Electric bus operations are vulnerable to disruptions, creating complex recovery challenges that necessitate balancing economic costs, battery constraints, and the consistency of driver schedules. To address this, this study proposes a spatiotemporal optimization model for electric bus disruption recovery. Unlike traditional rescheduling frameworks for fuel-based buses, this model accounts for dynamic battery consumption and charging decisions under disrupted conditions. A key innovation of the study is introducing the "spatiotemporal consistency" as a constraint, which is designed to maintain the stability of driver schedules in both time and space. To solve the associated Mixed-Integer Non-Linear Programming (MINLP) problem efficiently, the study proposes a customized gradient heuristic search algorithm. Simulation results based on the real-world bus network of Hengshui City show that the model produces feasible recovery plans rapidly across three typical scenarios: severe weather, sudden accidents, and cumulative disrupted events. The recovery duration was successfully controlled within a range of 95 to 153 minutes. Ultimately, the results demonstrate that the proposed model reduces operating costs and quickly generates recovery plans while safeguarding the spatiotemporal consistency of schedules, proving its adaptability in complex operational environments.

In recent years, lane detection has become a key technology in autonomous vehicles. However, lane detection is difficult to achieve accurate recognition because it greatly affected by complex environmental factors. In order to detect the lane markings with high accuracy and effectiveness in various complex environments, this paper innovatively proposes a lane markings detection model DG-UFLD based on anchor point classification. Firstly, a novel Global Spatial Channel Attention Module (GSCA) is designed to improve the overall performance of the model, which can uniquely capture multidimensional global information and enhance feature representation. Secondly, dynamic snake convolution (DSConv) and triplet loss function are introduced into the model to further enhance its detection capability for slender targets. Finally, the integrated model adopts coarse-grained lane grid classification instead of fine-grained segmentation, which achieves extremely high detection speed with high accuracy. To verify the effectiveness of the proposed DG-UFLD algorithm, case studies are conducted on two publicly available lane detection datasets (TuSimple and CULane) in this paper. The results show that compared to the original algorithm, the proposed DG-UFLD algorithm improves the detection accuracy on the TuSimple dataset from 96.05% to 96.45% at extremely high detection speed; the detection accuracy on the CULane dataset increases from 68.4% to 75.3%. Meanwhile, compared with other mainstream lane detection networks, the improved network achieves better results in lane detection under extreme conditions. The verification results demonstrate that this method can quickly and accurately detect the lane markings.

Focusing on the emerging mixed traffic environment where connected and automated vehicles (CAVs) and human driven vehicles (HVs) coexist, this paper studies the cooperative evacuation problem under flexible lane allocation strategies. First, considering the platooning behavior of CAVs, the cell transmission model is adopted to simulate the dynamics of the mixed traffic flows. Then, to minimize total evacuation time, the cooperative evacuation problem of mixed traffic is formulated as a mixed integer nonlinear programming model, and an improved Benders decomposition algorithm is designed to solve it. Numerical experiments are conducted on the road network of the core urban area in Xi'an to analyze the evacuation efficiency, lane allocation schemes, and optimal platoon sizes under different evacuation demands and CAV penetration rates. The results indicate that: (1) compared with the fixed lane allocation, the flexible lane allocation can improve the evacuation efficiency by more than 10% by reusing released road resources; (2) the total evacuation time decreases with the increase of CAV penetration rate. When the CAV penetration rate exceeds a certain threshold, flexible lane allocation is equivalent to the CAV priority strategy; (3) when the evacuation demand or CAV penetration rate is high, allocating more lane resources to CAVs can reduce the total evacuation time; (4) there is an optimal CAV platoon size, which is dynamically adjusted with the changes in the evacuation demand and CAV penetration rate. In this paper, the optimal value is mostly 5 or 6.

To reduce the high infrastructure cost of the full-line catenary installation mode in electric road systems, this paper develops a life cycle cost optimization model for the "segmented catenary installation + on-board energy storage" technology. The model comprehensively considers the costs of infrastructure construction, vehicle purchase, energy consumption, system maintenance and battery replacement. The model uses battery capacity, charge-discharge rate and catenary proportion as the key decision variables, and adopts the particle swarm optimization algorithm for multi-period dynamic optimization. Taking a typical coal freight scenario as an example, the study analyzes the optimal parameter configuration for different operational lifespans under the conditions of battery capacity ranging from 50 Ah to 150 Ah and charge-discharge rate from 1 C to 3 C. The results show that the optimal system configuration presents a significant dynamic variation with the operational lifespan: in the 25~30 years operational cycle, the combination of high battery capacity (150 Ah) and high charge-discharge rate (3 C) is more economical, corresponding to a catenary proportion of about 43%. When the operational cycle is extended to 40~45 years, the optimal strategy shifts to low battery capacity (50 Ah) and low charge-discharge rate (1 C), with the catenary proportion increased to about 87%, so as to reduce battery replacement costs and exert the long-term amortization effect of infrastructure. Sensitivity analysis indicates that battery unit price and line construction cost have a significant impact on system configuration: when the battery price rises by 20% or the line construction cost drops by 20%, the optimal catenary proportion both increases to about 87%. The research results can provide a reference for the planning and investment decision-making of electric road systems under different operational cycles.

Uncertainty in passenger load often leads to continuous fluctuations in instantaneous train power, which are further amplified when multiple trains operate simultaneously within the same power-supply section, increasing the likelihood of instantaneous power violations. To jointly address energy consumption minimization and peak power shaving, this study develops an energy-efficient optimization model with peak-power constraints within a speed profile selection framework. A distributionally robust optimization method based on the Wasserstein ambiguity set is used to characterize the uncertainty of train passenger-load probability distributions. To mitigate this challenge where evaluating peak-power constraints requires fine-grained temporal resolution and causes the model size explosion, this study proposes an exact dynamic programming algorithm enhanced with a time-domain reduction strategy. By exploiting the structural and periodic characteristics of traction power within each power supply section, the method identifies and retains only the time instants that are critical for peak-power evaluation. It further establishes that the reduced time set yields the same feasible solution space as the full temporal domain with respect to peak-power constraints, thereby reducing computational burden while preserving model fidelity. Numerical experiments based on the Beijing Subway Yizhuang Line demonstrate the effectiveness of the proposed approach. Relative to a baseline dynamic programming method, the proposed algorithm improves computational efficiency by approximately 47.4% in large-scale instances. Compared with stochastic and robust optimization models, the distributionally robust optimization model consistently achieves lower energy consumption while raising the minimum instantaneous power stability ratio to 98.99%, delivering a more balanced performance between energy saving and peak-power regulation. Sensitivity analyses further indicate that appropriate tuning of peak-power limits and confidence levels enables reliable trade-offs between energy efficiency and power-supply stability, confirming the practical applicability of the proposed framework for energy-efficient train operation.

This study aims to reveal the spatial agglomeration patterns and regional functional differences of urban freight activities. It also intends to conduct an in-depth analysis of the core influencing factors and differentiated action mechanisms of meso-scale freight functional positioning. To achieve these goals, this study proposes "freight communities" as the basic meso scale unit for characterizing urban freight spatial organization, and carries out an empirical study in Shanghai. Based on the heavy duty truck trajectory data in Shanghai, a freight mobility network is constructed. The Louvain algorithm is adopted to identify 79 freight communities with clear structures and close internal connections, achieving a modularity of 0.89. A multi-dimensional indicator system is established, covering freight scale, flow structure, operation characteristics, and specialization level. The k means++ clustering algorithm is used to classify the freight communities into five types: core gateway hubs, large-scale manufacturing bases, peripheral multi-service areas, central urban consumer areas, and specialized enclosed industrial zones. Furthermore, the Extreme Gradient Boosting (XGBoost) model combined with SHapley Additive exPlanations (SHAP) values is used to analyze the influence mechanism of community functional positioning. Five core driving variables are identified: distance to ports, distance to railway freight stations, distance to expressways, distance to urban expressways, and population density, with a cumulative contribution rate of 70%~80%. It is found that the impacts of these variables on different functional communities exhibit significant non-linear characteristics and threshold effects. This study clarifies the functional positioning and spatial organization logic of freight communities. It can provide a theoretical basis and practical guidance for the precise governance of urban freight space, the optimal allocation of infrastructure, and the formulation of differentiated management and control strategies.

Meteorological conditions like wind, waves, and currents often reduce vessel speed in navigational channels, impacting transit times and carbon emissions. To optimize vessel sequencing and scheduling, this study investigates a multi-meteorological factor channel scheduling problem. A bi-objective mixed-integer programming model is developed to minimize carbon emissions and scheduling time deviations, incorporating constraints such as vessel speed loss, tidal time windows, channel capacity, and several transit rules. An improved Non-dominated Sorting Genetic Algorithm II (NSGA-II), enhanced with deep learning, is proposed to obtain Pareto-optimal solutions. The ML component combines k-means clustering and recurrent neural networks to estimate actual vessel speeds under varying weather conditions, while an adaptive parameter adjustment mechanism maintains population diversity in NSGA-II. Experimental results show that the proposed approach reduces speed estimation error from 11.2% to 1.4%. The Pareto front achieves two extremes: zero scheduling time deviation and a 60.8% reduction in carbon emissions. In terms of meteorological effects, the model effectively mitigates significant speed loss, improving the two objectives respectively by 11.8% and 67.0%. The sensitivity analysis indicates that vessel speed loss increases from 0.8 knots under mild conditions to 2.9 knots under moderate conditions, accompanied by rises of 90.4% in carbon emissions and 36.9% in time deviation. No feasible solution exists under severe conditions. These findings demonstrate the model's dual role: it provides trade-off solutions under most weather scenarios and works as an early warning system for extreme conditions.

Traffic accident liability determination is a key aspect in improving the governance capacity of traffic safety. In view of the high manual dependence, strong subjectivity, and low efficiency in current traffic accident liability determination in China, this paper proposes a Multimodal Large Language Model that incorporates Retrieval-Augmented Generation (RAG) and Chain-of Thought (CoT). First, accident keyframe sequences are sampled and extracted from surveillance videos. Then, a traffic law knowledge base is constructed, and relevant legal clauses are retrieved using a hybrid strategy that combines BM25 and vector retrieval with Reciprocal Rank Fusion (RRF), to suppress "hallucinations" in legal citations. A CoT prompting strategy is then designed to drive the model to perform multi-step liability reasoning, automatically generating a structured Road Traffic Accident Determination Report, including accident facts, legal bases, and the final liability conclusion. Experimental results on a real-world accident dataset show that the proposed method attains an accident liability determination accuracy of 80.00%, improving by 6.67% over the baseline without RAG and CoT. Meanwhile, the legal provision citation accuracy increases from 6.67% to 56.67%, significantly alleviating erroneous citations and fabricated legal content. The generated reports demonstrate practical usability in terms of key information coverage, semantic consistency, and format standardization.

Multi-entry underpass tunnel has prominent safety concerns due to the distance and grade of the entrance, the tunnel's enclosed space, limited sight distance and visual field, particularly in the underground merging area which has high risk of crashes. This paper analyzed drivers' safety from the main entrance of the underpass to the merging area, and conducted a real vehicle test using an eye tracker to collect drivers' pupil data, saccade duration, and saccade angle. The tunnel section from the main entrance to the underground merging area was divided into five sections to analyze drivers' visual characteristics and driving safety across different sections. The results show that the relative change rate of pupil area increased significantly upon entering the tunnel, rising from 41.56% in the shading section to 286.84% in the entrance internal section and further to 407.40% in the underground merging section, where it then stabilized. The cluster analysis indicated that the saccade angles in the underground merging section and its adjacent sections were grouped into the same category, suggesting consistent saccade behavior patterns before and after the underground merging section. Based on the optimal thresholds of saccade duration (36.00 milliseconds) and saccade angle (11.60 degree), saccades were classified into four categories. Moreover, the underground merging section exhibited the highest peak-to-mean ratio of saccade speed in the tunnel, reaching 3.62 decibel, indicating elevated visual instability and increased driving safety risk. The findings provide an understanding of drivers' visual behavior patterns in the underground merging area and their implications for driving safety, providing references for road design and traffic management.

To reveal the differential characteristics and intrinsic driving mechanisms of the spatial evolution of airport clusters, promote their coordinated development, and optimize air network layout, this study takes the Beijing-Tianjin-Hebei (BTH) and Chengdu-Chongqing (CCQ) airport clusters as research objects. Focusing on the distribution characteristics of aviation resources on both sides of the Hu Huanyong Line, it systematically analyzes the spatial heterogeneity from a multi-scale perspective. By constructing the quantitative indicators including coverage capacity index, coverage heterogeneity index, spatial influence model, spatial influence heterogeneity index, route overlap coefficient, and route heterogeneity index, the study comprehensively compares the connectivity structure, service coverage, and network differences of the two airport clusters under different radiation radii. The results show that: (1) the 1 500 kilometers radiation radius is a turning threshold for changes in coverage capacity and spatial heterogeneity of airport clusters; (2) both of two airport clusters exhibit a common feature of "stronger in the east and weaker in the west", and the east-west gap is more significant in the CCQ. (3) the comprehensive influence proportions of the eastern sides of the BTH and CCQ are 77.73% and 93.55% respectively, which highlights differences in the agglomeration degree of regional aviation resources; (4) overlapping routes are concentrated along the Hu Huanyong Line, with high structural heterogeneity but low density on the northwest side, and "high density and low difference" on the southeast side; (5) the sensitivity analysis of weight coefficients indicates that the model results have good stability and reliability. This study reveals the coupling mechanism between the spatial pattern of Chinese airport clusters and regional economic development, providing the theoretical basis and decision-making reference for optimizing aviation resource allocation, promoting airport group coordination, and balancing the development between the east and west.

The port-hinterland intermodal transport network consists of multiple types of coupled nodes, including ports and rail/ road container terminals. Major disruptions can reduce local connectivity and, through the reallocation of transport operations, trigger cascading functional degradation. To quantify the resilience of multimodal container networks under major disruptions, this paper proposes a "resilience quadrilateral" model that decomposes the resilience process into four sequential phases, including resist, absorb, adapt, and recover. Using the Space-P approach, this study built an undirected weighted hinterland network for the Port of Shanghai which encompasses 253 nodes (port terminals, rail container yards, and road container depots) and 1 067 links. On this basis, four node attack-recovery strategies are designed and simulated to evaluate differences in network resilience under different disruption and recovery rules. The results reveal that (1) the overall network attains its highest resilience under the "random attack + prioritized recovery" regime ( NT = 0.16, SNT = 172.50) and has its lowest resilience under "targeted attack + random recovery" ( NT = 0.27, SNT = 257.50); (2) at the sub-network level, the road network shows the smallest performance decline rate and the fastest recovery, the resilience of road is better than inland-waterway, and inland-waterway has better resilience than rail; (3) after removing all waterway nodes, all rail nodes, and all road nodes, the remaining network performance values are respectively 0.57, 0.91, and 0.37, indicating that the sub-network importance ranks as road, then waterway, then rail.

As a type of Flexible Transit (FT), Deviated Fixed Route (DFR) bus services allow vehicles to moderately deviate from a predetermined baseline route to pick up and drop off passengers. The current researches in this field primarily focus on the vehicle deviation rules, service constraints, and scheduling strategies based on the baseline route. Typically, the baseline route is either assumed to be an existing bus route or constructed virtually within a simulated network structure. It lacks data-driven planning methods that reflect the actual characteristics of passenger demand. Therefore, this paper proposes a planning method for baseline route driven by taxi trajectory data. The method comprises three stages: bus-stop generation, route generation, and route evaluation &selection. During the bus-stop generation, a multi-period origin-destination (OD) matrix is first constructed with taxi trajectory data to identify the high-frequency OD pairs. Candidate stops are then generated through grid clustering and merge-and-split strategies. During the route generation, a weighted point graph is constructed by integrating the candidate bus stops with the rail transit stations. Initial candidate routes are generated with the demand-intensity-guided probabilistic path search. These routes are then filtered through operational constraints. A Pareto analysis is subsequently applied to derive a feasible baseline route set under dual objectives. Subsequently, a multidimensional evaluation system is employed to quantitatively assess the candidate routes and select the optimal one as the benchmark route plan. The Shenzhen case study validated the feasibility of this method. The following results are demonstrated: the benchmark routes distributed along primary travel corridors and effectively connected subway stations, with over 14% of their stops located at the subway stations and over 89% non-overlapping with the existing bus routes; the benchmark routes for opposing directions during the same time period exhibit the differences in metrics such as stop locations, route alignment, passenger flow intensity, average travel time, and metro connection ratios. Such variations also occur among the benchmark routes generated for the same origin-destination pair but different time periods, demonstrating that this method can generate benchmark routes matching with spatio-temporal heterogeneity of passenger flow characteristics.

Under the background of the integration of metro and active mobility, this study uses the data of 2025 Beijing metro smart-card transactions and dockless bike-sharing trip in one week to identify bike-metro feeder trips. Then it develops a method to delineate the bike-sharing catchment areas around metro stations. This paper quantifies the bike-sharing feeder ratio across time periods and feeder modes. And then it integrates the Fuzzy C-Means (FCM) clustering with Multiscale Geographically Weighted Regression (MGWR) to investigate the spatiotemporal heterogeneous effects of built environment on transfer shares for different station types. The case study of Beijing shows that bike sharing expands the traditional walking catchment radius of metro stations, reaching an average distance of approximately 1.6 km. During peak hours, the average bike-sharing feeder ratio is about 16%, with low-ridership stations showing a slightly higher proportion than other station types. The MGWR results indicate that high job housing density suppresses the bike-sharing feeder ratio, whereas a high level of land-use mix promotes it. Sufficient bike-sharing supply is a key prerequisite for maintaining a high feeder ratio. The results for bus stop density reveal spatial differentiation in the relationship between bus and bike sharing, showing both competition and complementarity: the two modes tend to be complementary in central urban areas but competitive in peripheral areas. In addition, the bike-sharing feeder ratio is not only affected by the built environment but also moderated by station functional attributes. Specifically, residential-oriented stations are more strongly affected by competition from bus, whereas employment-oriented stations are more sensitive to transport location factors.

This paper addresses the problem of temperature control coordination failure at origin-distribution nodes in cold chain road-rail intermodal transport by constructing a three-party differential game model for road carriers, railway carriers, and the government. A synergy gain coefficient is introduced to characterize the cross-modal spillover effects of temperature control investments, and its theoretical expression is derived from three channels: temperature handover, information synergy, and learning spillover. This paper establishes the equivalence between the positive synergy gain coefficient and strategic complementarity, thereby reveals the game-theoretic root of coordination failure. Analytical equilibrium solutions are obtained for decentralized, leader-follower, and cooperative scenarios. The equilibrium existence boundary is derived, and the marginal efficiencies of investment and operational subsidies are compared. Results show that: (1) the three channels contribute 33.4%, 43.3%, and 23.3% of the total synergy gain, respectively, with a greatest improvement potential of over 40% in the information synergy exhibiting; (2) under the decentralized decision-making, the investments of road and railway temperature control reach only 30.8% and 37.3% of the cooperative optimum, respectively, resulting in an efficiency loss of 72.9%; (3) under the small-synergy ( β→0 ) and symmetric-parameter assumptions, the unconstrained optimal subsidy rate is approximately 90%, yet equilibrium existence and system stability constraints compress the feasible range to 25%~30%; within this range, the marginal efficiency of investment subsidies is approximately 1.73 times that of operational subsidies (theoretical upper bound under symmetry is 2); and (4) an "incentive-stability" trade-off exists: increasing the subsidy rate from 0 to 30% raises total investment by 65.3%, yet the equilibrium existence threshold decreases from 0.62 to 0.43. These conclusions are robust to parameter variations, providing a theoretical basis for coordination mechanism design and subsidy policy optimization.

To address the "last-mile" challenge in rural logistics and expand the service scope of instant retail, this paper proposes a scheduling method for urban-rural bus of passenger-freight co-transport based on a zonal alternating response mechanism, considering the instant retail orders. A delivery model is developed to leverage the existing fixed-route resources of urban-rural buses to transport both standard express packages and instant retail orders. The rural operating segment was partitioned into several zones with an alternating response mechanism designed to aggregate passenger and freight demand and enhance transport coverage. To account for variations in delivery timelines, this study uses an urgency coefficient to characterize the degree of priority, enabling the dynamic adjustment of delivery sequences. A mixed-integer programming model is developed to minimize the total mileage, and a hybrid algorithm combining Adaptive Large Neighborhood Search and Variable Neighborhood Search is designed for the solution. Taking Route H22 in Huaiyin District, Huai'an City, Jiangsu Province as a case study, the study performed simulations under varying transportation intensities. Two passenger-freight co-transport scenarios were examined: moderate passenger demand with moderate freight flow, and low passenger demand with high freight flow. The results indicate that under two passenger-freight co-transport scenarios, the additional freight revenue accounted for 39.5% and 72.1% of the total income respectively, while ensuring passenger service quality. This demonstrates that the integration can effectively utilize spare bus capacity to boost total revenue. Furthermore, the expedited delivery strategy for instant retail orders increased the total revenue by 6.35% and 6.14%, respectively. The zonal alternating response strategy effectively reduced the mileage by 16.9% and 18.6%, and shortened average passenger delays, with its advantages suitable for adoption during periods of low passenger flow.

To address uncertainties in material demand and drone coverage radii during disaster relief, this paper proposes a fuzzy optimization method for locating drone emergency distribution centers. Using k-means clustering on geographical data to identify candidate sites, the study uses triangular and trapezoidal fuzzy numbers to characterize parameter uncertainty. A Fuzzy Maximal Covering Location Problem (FMCLP) is defined with credibility constraints, aiming to maximize covered demand, minimize total costs, and optimize spatial equity. A Hybrid Simulated Annealing (HSA) algorithm, embedded with fuzzy simulation and Monte Carlo sampling, is designed to achieve global optimization via local search and adaptive cooling. A case study in Jiangning District of Nanjing city demonstrates that compared to deterministic models, the proposed method increases covered demand by 82 units, reduces total costs by 61 400 yuan, and shortens the maximum uncovered distance by 7.63 km, while exhibiting superior robustness against facility failure. The optimal weight combination for coverage, cost, and equity is determined as (0.7, 0.2, 0.1). The findings validate the model's applicability in complex, uncertain environments, providing a scientific basis for urban drone logistics planning.

This paper reviews 200 studies that employ virtual reality (VR) technology in pedestrian traffic research, and systematically summarizes the existing work from the perspectives of experimental tasks, experimental scenarios, research subjects, and research themes. A comprehensive overview of the current research progress is provided, and future development directions are proposed. Results show that evacuation and street-crossing tasks take account for the highest proportion of VR-based pedestrian experiments, which reflects the advantages and applicability of VR in simulating high-risk and complex traffic environments. Different tasks emphasize different scenario components: evacuation and wayfinding tasks focus on spatial structure and signage, and street-crossing tasks emphasize roadway and vehicle characteristics, while walking tasks highlight the interpersonal influences. By enabling the flexible configuration and fine-grained control of dynamic and static environmental variables, VR provides a precise and controllable technical support for exploring the underlying mechanisms of pedestrian traffic. With the respect to research subjects, path choice and street-crossing decision-making constitute the core areas of investigation, and physiological and psychological attributes appear with comparable frequency across different tasks. Research has gradually shifted from simple behavioral observation to the analysis of multidimensional coupled mechanisms. In terms of research themes, studies which examine pedestrian interactions with the environment, vehicles, and other individuals dominate across all types of experimental tasks. As new technologies, facilities, and services emerge in the era of digital intelligence, interaction-oriented research is expected to further expand. Research on heterogeneous and special populations has become a hotspot, while hazardous behavior identification remains concentrated in street-crossing tasks. VR provides dynamic tools for observing and analyzing the individual cognition, responses, and behavior, which offers a substantial potential for further exploration of individual differences and behavioral diversity. Studies on VR validity, system performance, and intervention effectiveness are found across various tasks, and the scientific evaluation and calibration of VR systems will help unlock their full potential. In view of the current research landscape, future work may focus on improving the accuracy and applicability of VR experiments, enhancing multimodal data fusion and behavioral-mechanism modeling, advancing pedestrian street-crossing studies under automated-vehicle scenarios, and further leveraging VR to identify and mitigate hazardous pedestrian behaviors across diverse environments.

There is an optimization objective conflict issue in the joint optimization of traffic signals and trajectories of connected and automated vehicles (CAV) based on deep reinforcement learning. To address this conflict, a dual-agent system consisting of traffic signals and CAVs is designed, and a dynamic weight joint optimization framework (Dynamic-Weighted Optimization into Positioning with Double Deep Q-Network, DOP-DDQN) is proposed. By constructing a joint reward function that integrates traffic signals and CAVs, the balance between the efficiency of traffic signal and CAV on safety, energy consumption, and efficiency is achieved. The weights of traffic signals and CAVs in the joint reward are dynamically adjusted by using fuzzy logic, which convert the objective conflict into a dynamic priority allocation problem, and thus enable the adaptive adjustment of the optimization focus based on traffic conditions. Simulation results show that compared to MaxPressure, DOP-DDQN reduces average queue length, travel time, and fuel consumption by 15.97%~23.74%, 8.69%~9.89%, and 4.19%~9.53%, respectively. Compared to other similar methods, these three indicators decrease by 4.06%~15.19%, 2.30%~6.62%, and 1.42%~5.11%, respectively. Further research indicates that the control effectiveness of DOP-DDQN significantly increases with the penetration rate of CAVs, but the improvement tends to level off when the penetration rate exceeds 0.6.

To address the multiple objectives challenge of urban traffic signal control that is difficult to simultaneously balance the traffic efficiency, public transport priority, and environmental impact, this paper proposes a modeling method for traffic signal agent based on a graph attention mechanism. First, in terms of road network spatial modeling, an innovative approach centered on "approach road segments" is constructed, and a graph attention mechanism is further introduced to characterize the spatial dependency relationships among road segments. Thereby it enhances the perception of agent in upstream traffic conditions. Second, in traffic flow temporal modeling, this paper compares two types of network architectures: a multilayer perceptron and a long short-term memory network. Third, for the multi-objective optimization task, a reward function with adaptive and dynamic weight adjustment based on real-time road conditions is designed, enabling the coordinated optimization of traffic efficiency, public transport priority, and environmental impact. Finally, a square urban traffic signal grid consisting of three north-south and three east-west corridors with a side length of 1.2 km is selected as the experimental road network, on which the traffic signal agent is trained and tested under different traffic volumes and fluctuation scenarios. The results show that the proposed traffic signal agent significantly outperforms a simple adaptive control across all objectives: in terms of traffic efficiency, the achievement degree of ideal vehicle speed under high traffic demand is improved by 36.21%; in terms of environmental impact, fuel efficiency is improved by more than 11.0%; and in terms of public transport priority, the model increases the number of buses passing through by approximately 25.0%, with bus speed closeness improved by 14.5%. Therefore, the agent model of multi-objective traffic signal developed in this paper demonstrates a strong multi-objective coordination capability and provides a feasible pathway for building efficient, equitable, and environmentally friendly intelligent traffic signal control systems.

This paper conducts a study on cooperative lane-changing control for vehicular platoons based on an improved artificial potential field method for the lane-changing problem of platoon in mixed traffic flow scenarios. First, a two-layer decision-making model for platoon lane-changing is proposed. The first-layer decision evaluates the safety of surrounding environment, and the second-layer decision selects a lane-changing strategy which is suitable for the current scenario based on the surrounding vehicle types. Subsequently, the platoon lane-changing maneuver is realized with an artificial potential field (APF) method. To address the limitation of traditional methods which ignore the vehicle kinematic constraints, a Dubins curve-based path planning method is introduced to construct a multi-potential-field model that adapts to the networked environment. While the smooth lane-changing trajectories that meet vehicle kinematic constraints are generated to ensure the physical feasibility of the planned path. Finally, a lane-changing scenario is constructed and the feasibility and effectiveness of the collaborative lane-changing control method for the platoon are verified through simulation. The results show that the proposed method incorporating Dubins curves can plan smooth lane-changing paths; the proposed collaborative lane-changing method can achieve dynamic obstacle avoidance for the platoon; compared to the non-cooperative lane-changing, the cooperative lane-changing strategy can reduce the speed error between the following vehicle and the platoon to approximately 0 m·s-1, shorten the lane-changing time by 38.7%, and reduce the minimum following distance by 15.2%. The proposed cooperative lane-changing method not only satisfies the safety requirements of vehicular platoons during lane-changing, but also improves the operational efficiency of traffic flow.

To maximize the utilization of lane resources across the entire road network, a control strategy for Connected and Automated Vehicles (CAV) is proposed through borrowing dedicated bus lanes, under the premise of guaranteeing bus priority. First, by analyzing the characteristics of bay-style bus stops, a dynamic clearance distance model for buses is constructed, and the control rules for CAVs borrowing bus lanes are formulated for normal road sections and bus bay sections, respectively. Second, considering that the mandatory lane-returning of CAVs is susceptible to the traffic status of adjacent lanes, a cooperative control strategy for CAV lane-returning is proposed. Specifically, the critical density of adjacent lanes is introduced to regulate the mandatory lane-changing behavior of CAVs. The scenarios are classified based on the gap between the leading and following vehicles in an adjacent lane, and cooperative lane-changing rules for CAVs are formulated to ensure the priority buses. The simulation scenarios for various strategies are constructed and verified by using the SUMO simulation platform. The results indicate that: (1) Compared with the strategy prohibiting CAVs from using bus lanes, the average travel time of social vehicles is reduced by 5.4%, and the speed of joint dedicated lanes is increased by 21.21%. Moreover, under high traffic flow, the average speeds of CAVs and Human-driven Vehicles are increased by 13.14% and 9.63%, respectively, while the average speed of buses is reduced by only 0.99%. (2) Compared with the strategy that allows CAVs to borrow bus lanes without considering the bus priority, the proposed strategy increased the average travel time of social vehicles by 1.9%, and the average speed of buses by up to 5.91% respectively. The above results demonstrate that the proposed strategy can improve the traffic efficiency of road networks without compromising bus priority.

To address issues such as false detections, missed detections, and misjudgments of queue status in long-queue detection under surveillance views—caused by occlusions from dense traffic, aggregation of small distant targets, and inconsistent spatial scales between near and far regions in surveillance scenes—this paper proposes an edge-feature-enhanced long-queue detection method for complex traffic scenarios. In the object detection module, a novel Edge Feature Enhancement Pyramid (EFE-Pyramid) is designed, which incorporates an Edge Feature Convolution (EFConv) to explicitly extract shallow-layer edge response features, and employs a deep-shallow feature fusion mechanism to enhance edge features, thereby improving the model's discriminability and detection accuracy for occluded objects in dense traffic and small distant targets. On this basis, a task-aligned detection head is used to ensure real-time performance. For queue length estimation, an adaptive perspective-aware method is proposed to achieve unified alignment of target representations across near and far regions, enhancing the stability and generalization capability of queue length estimation. Experimental results demonstrate that, on both public datasets and a self-collected complex-traffic scenario dataset, the proposed algorithm achieves higher recognition performance in complex traffic environments with relatively low model complexity, improving detection accuracy by 2.4 and 4.8 percentage points over the baseline model, respectively, and outperforming current state-of-the-art methods. Furthermore, in queue length estimation experiments, it achieves an average absolute percentage error of 4.8%, fully validating the effectiveness and practicality of the proposed approach for long-queue detection in complex traffic scenarios.

Nighttime UAV(Unmanned Aerial Vehicle) aerial imagery suffers from inherent limitations such as low illumination, strong noise, and small targets, which leads to persistently high false detection and missed detection rates in vehicle monitoring. To address these issues, this paper proposes NS-YOLOv8 (NightSpin-You Only Look Once version 8), a vehicle detection algorithm designed to improve the detection accuracy of small target vehicles in low-light aerial scenarios. The method adopts a two-stage image enhancement strategy at the input end, which performs image restoration and enhancement through noise suppression and color restoration. Furthermore, a streamlined four-layer feature pyramid architecture is introduced into the backbone network to reduce redundancy and optimize feature representation. A deformable attention mechanism is embedded into the neck network to enhance the focus of model on target regions. Finally, cross-layer fusion between the backbone and neck networks achieves effective complementarity and integration of deep and shallow features. Experimental results demonstrate that NS-YOLOv8 outperforms methods such as Oriented R-CNN and YOLOv6-OBB in detection accuracy. It achieves Precision of 96.8%,mAP@0.5-0.95 of 80.3%, and an F1-score of 96.5%, with only 1.3 M parameters and 12.1 GFLOPs. Visualization analyses further confirm that the proposed algorithm effectively reduces false and missed detections of cars and trucks at night, while it improves detection confidence, and makes it suitable for precise small target detection in low-light environments.

To resolve the mismatch between transport supply and demand and improve transfer efficiency in multi-line bus systems, this paper proposes a scheduling optimization model for Modular Autonomous Vehicles (MAV). The model enables an "in-vehicle transfer" mode, where passengers complete cross-line transfers without alighting, facilitated by the dynamic coupling and decoupling of Modular Units (MU). A dual-stage optimization framework based on a Multi-Population Genetic Algorithm is designed. The first stage optimizes timetables and vehicle formations, while the second schedules specific MUs to minimize total generalized costs, comprising operator expenditures and passenger waiting times. Numerical experiments using real-world data from Beijing demonstrate that the proposed method resulted in a reduction of the total system cost by 7.0% compared to traditional fixed-capacity bus systems, a 3.1% reduction of cost compared to MAV systems relying solely on walking transfers, and a 1.5% reduction of total cost compared to scenarios adopting all-passenger in-vehicle transfers. The study result indicates that while in vehicle transfers enhance service quality, the marginal benefits diminish with an excessive transfer proportion, suggesting that the model effectively balances economic efficiency and passenger experience by identifying the optimal transfer ratio.

Large passenger hubs often face short-term surges of dispersive feeder demand, which can lead to accumulated detours and cross-area route interweaving, making customized bus routes difficult to organize and operate. This study proposes a static route planning method for hub-oriented customized buses by introducing a sector-based compactness constraint. Given candidate stops, stop-level demand, and road-network distance/time matrices, the study proposes a planning model that minimizes system wide total cost, including fixed route cost, distance-based operating cost, and in-vehicle time cost, while controlling detours through a detour-ratio constraint. To further suppress cross-sector interweaving and enhance route-shape controllability and operational stability, the study uses a sector-based compactness constraint to limit the azimuth coverage arc length of each route with respect to the hub. A region-aware Adaptive Large Neighborhood Search (ALNS) is conducted by sorting stops by bearing angles. Then, a "sector-boundary-prioritized" destroy operator and a region-aware greedy insertion repair operator are adopted, and local refinement is performed via 2-opt and inter-route relocation moves. A case study of Jinan Railway Station shows that, with a maximum angular-span threshold of 60°, the method yields 19 routes with a total cost of 175.3 thousand yuan, total mileage of 306.39 km, total operating time of 735.34 min, and a mean detour ratio of 1.08. Compared to the baseline without the compactness constraint, the maximum detour ratio decreases from 4.88 to 2.73, the mean angular span drops from 61.58° to 27.85°, and the average number of sector transitions reduces to 0.16 per route. The results demonstrate that sector-based compact organization can significantly improve route morphology and mitigate extreme detours without compromising service coverage.

Existing U-shaped dual-cycle scheduling models typically assume deterministic operation times. However, variations in vessel arrival times and equipment service durations may cause delays in export-container retrieval, yard transportation, and vessel or truck loading operations, which tend to accumulate along the operational chain and are difficult to capture during scheduling. To mitigate the risk of repeated postponements in container retrieval, yard transfer, and loading completion under time-varying operating conditions, this paper proposes a two-stage stochastic scheduling framework. In the first stage, task sequences and equipment resource allocations are jointly determined; in the second stage, operation completion times are evaluated under different arrival and service scenarios. Conditional Value-at-Risk (CVaR) is incorporated to characterize adverse operational outcomes with significant impacts on vessel-and truck-loading schedules. Computational experiments demonstrate that the proposed approach consistently produces stable solutions across different problem scales. Compared with the first-come-first served policy and a priori robust scheduling strategy, it achieves superior performance in terms of completion-time distributions, system operating conditions under severe disruptions, and delay propagation control. The average number of cascading propagation layers is reduced from 3.42 to 1.27, and the proportion of disruption-affected operations decreases from 38.6% to 11.8%. By accounting for completion-time risk during task sequencing and resource allocation, the proposed approach mitigates the accumulation of waiting times across interconnected equipment operations and limits the propagation of delays to downstream activities.

To enhance the accuracy of urban public transit travel demand clustering and overcome the limitations of traditional clustering methods in capturing non-convex patterns, handling noise, and reducing parameter dependency, this paper proposes a general framework for public transit travel demand clustering supporting cross-domain applications and develops a Geographic AutoEncoder-based Hierarchical Clustering (GeoAE-HC) algorithm. The sine-cosine positional encoding is introduced and a geographical self-attention mechanism is designed to capture both feature similarity and geographical proximity among transit travel demand data. A hierarchical clustering method integrating density-based clustering and mean-based clustering is developed and applied in the latent feature space extracted by the autoencoder, enabling accurate clustering of travel demand distribution. To improve the cross-domain generalization ability of the model across different cities, a transfer learning strategy combining domain adversarial training with a frozen encoder and fine-tuned decoder is designed to strip away city-specific noise. Simulation results show that in the Chengdu dataset with 375 cluster centers, all evaluation metrics of the GeoAE-HC algorithm are superior to those of the comparison algorithms. Specifically, the Silhouette Coefficient is improved by 11.82%, 23.14%, and 45.01% compared to Deep Embedded Clustering(DEC), DK-means, and k-means, respectively. The Calinski-Harabasz (CH) Index is improved by 6.14%, 5.62%, and 14.44%, respectively. After cross-domain transfer to the Beijing dataset, the performance of GeoAE-HC also outperforms those of comparative algorithms, validating its effectiveness and generalization capability.

The train operation scheme determined by the train formation plan serves as a primary basis for timetable optimization. The quality of freight train path connections also affects the optimization of train formation plan. However, this interaction is usually ignored in traditional studies on formation plan. To achieve the integrated optimization of formation plan and timetable, this study develops a bi-level programming model with a nested structure. The upper-level model optimizes the formation plan and incorporates the average travel time cost of freight trains as a disturbance function to reflect timetable quality. Based on the operation scheme provided by the upper-level, the lower-level model determines the departure and arrival times of freight trains. To adapt to the model structure, a nested optimization algorithm based on simulated annealing is designed, where iterative optimization is achieved through the nested solution of two levels and travel time feedback. Case studies with different numbers of fixed-path trains are conducted on a railway line with 20 stations to validate the proposed method. The results show that, compared with the independent method that first obtains the optimal train formation plan and then arranges the train paths, the proposed method achieves a lower total cost, with the reductions ranging from 4.11% to 5.59% across all cases, while maintaining a relatively high computational efficiency. In a representative case with 36 fixed-path trains, the integrated method yields a total cost of 89 197.85 car-hours, which is about 5.38% lower than that of the independent method, with the costs of train formation plan increasing by about 5.62% and the cost of train travel time decreasing by about 11.41%.

Demand-responsive transit (DRT) has emerged as a promising solution to address the dual challenges of urban congestion and diverse mobility needs. However, its widespread adoption is constrained by complex and uncertain public perceptions. This study aims to dissect the core themes of public discourse on DRT and attribute the underlying sentiment drivers. A dataset of 25 831 text entries was collected from two major social media platforms, Douyin and Bilibili, spanning from July 2022 to July 2025. The study uses a hierarchical topic modeling approach, integrating a priori framework with the Gibbs Sampling Dirichlet Multinomial Mixture (GSDMM) topic modeling, and leveraged a Large Language Model (LLM) for sentiment analysis. The analysis identified four primary themes: "travel experience," "operational efficiency," "economy," and "others," which were further categorized into eight distinct sub-topics. The results reveal an overall public sentiment score of 0.37, indicating a slightly negative tendency. Positive sentiments were predominantly linked to "travel experience," reflecting an appreciation for DRT's core value proposition. Conversely, a significant volume of negative feedback was concentrated on tangible operational issues, such as booking procedures and routing inefficiencies. Crucially, this study uncovered a strong correlation between regional sentiment heterogeneity and DRT's functional role within local transport ecosystems. In some developed metropolitan areas, DRT is perceived as a beneficial complement and receives higher ratings. However, in regions where DRT has been extensively deployed to compensate for service gaps in the existing public transit network, a significant disparity between high public expectations and actual operational performance emerges as the primary driver of concentrated negative sentiment. This study provides timely public opinion intelligence for DRT stakeholders. The study concludes that the widespread adoption of DRT must abandon the "one-size-fits-all" model and instead shift towards a refined strategy tailored to the differences in local transportation systems. Furthermore, addressing core operational bottlenecks must be its foremost task to win public trust.

This study addresses the dynamic and coordinated control of pedestrian flow within metro stations, specifically during peak hours and under the conditions of sudden passenger surges. A station-level pedestrian flow control as a Markov decision process is formulated, a state space is constructed centered on key-area crowd density and travel time, and a multi-objective design is rewarded which jointly account for safety, efficiency, and operating cost. Using the Proximal Policy Optimization (PPO) within a simulation environment, the agent performs online joint optimization over entrance inflow control, train headway adjustment, queue-guidance barrier deployment, and other strategies. The application of reinforcement learning enables coordinated pedestrian flow control through dynamic online decision-making, rather than traditional static offline optimization. Compared with a static optimization benchmark, the PPO policy reduces high-density risk and service delays: the share of high-density states in the security screening area decreases from 22% to 10%, and the proportion of passengers with security queue times more than 3 minutes drops from 23% to 13%. These results indicate that the proposed method improves the operational efficiency while it satisfies safety thresholds and balancing relative operating costs.

Existing path planning algorithms for electric vertical take-off and landing aircraft predominantly use global static planning, and have difficulties to respond in real-time to dynamic obstacles (such as other aircraft, birds, or airborne debris) within complex low-altitude environments. This limitation impedes their real-time obstacle avoidance capabilities. This paper integrates the velocity obstacle method with the artificial potential field approach to develop an obstacle avoidance path planning algorithm for electric vertical take-off and landing aircraft operating in complex three-dimensional low-altitude environments. This algorithm considers both the aircraft's dynamic constraints and the motion characteristics of obstacles during the planning process. It uses the velocity obstacle method to calculate parameters such as the relative velocity and angle between the aircraft and obstacles, thereby identifying conflict avoidance zones. Furthermore, it utilizes an artificial potential field model to dynamically adjust attraction and repulsion functions, guiding the aircraft to progressively return to its flight path. This approach achieves a unified solution for global path optimization and local obstacle avoidance. Using parts of Tianfu New Area in Chengdu city as the setting, the study conducted 1∶10 scale simulation experiments to analyze two scenarios—conventional open airspace and confined narrow spaces— under real-world geographical constraints. Additionally, a high-density obstacle scenario was established to test the algorithm's adaptability and robustness under highly complex conditions. The experimental results indicate that, in conventional open-space scenarios, the proposed algorithm reduces the number of iterations by 24.7% and shortens the path length by 19.8% compared to the velocity-obstacle method. Compared to the artificial potential field method, it reduces the number of iterations by 20.2% and shortens the path length by 30.4%. In confined spaces and high-density obstacle scenarios, the algorithm effectively overcomes the local oscillations and excessive detours commonly caused by traditional methods, and significantly reduces the cumulative yaw angle of the aircraft and improves the smoothness and physical flyability of the flight path.

In view of the significant spatiotemporal heterogeneity of passenger flow on the main and branch lines of Y-shaped rail transit lines and the transport capacity-demand mismatch arising from the traditional fixed marshalling scheme, this paper proposes an optimization method for multi-route train timetables integrating cross-line operation and flexible marshalling. The operation logic of the coupling-decoupling process in the flexible marshalling mode and the multi-route layout are clarified. Practical operational constraints, including coupling/decoupling operations, safety headways, rolling stock connections, and the number of available rolling stocks, are considered systematically. A bi-objective optimization model is developed to minimize the total passenger waiting time and the train-kilometers of main line operation. Furthermore, a multi-objective genetic algorithm is designed for the cross-line scenarios. Core decision variables such as route starting points, marshalling modes, and train timetables are represented by real-number coding. The Pareto optimal solution set is derived by combining non-dominated sorting and crowding distance. The effectiveness of the model and algorithm is verified through an empirical example in a cross-line Y-shaped rail transit line in a city. The results indicate that the proposed optimization scheme well considers the temporal-spatial distribution characteristics of passenger flows on main and branch lines through flexible marshalling and coordinated scheduling of multiple routes. Compared to the traditional independent branch line operation mode, the total passenger waiting time decreases by an average of 1.4%, and the vehicle operating kilometers are reduced by 10.8% on average. Whereas, in contrast to the through-merge operation mode, the total passenger waiting time decreases by 54.2%. This scheme achieves a Pareto-optimal balance between passenger experience and operational costs, and offers scientific and practical technical support for the interconnection operation of Y-shaped cross-line rail transit systems.

To improve the accuracy, quality and efficiency of automatically generating train working diagrams, a multi-objective collaborative optimization model is formulated for intercity railway timetable and station track utilization. A stepwise collaborative optimization strategy is proposed to address the complex variables and coupled constraints. By progressively adding constraints to control the single-step solution scale, an optimal feasible scheme can be generated efficiently within a limited time. Taking the Beijing-Tianjin Intercity Railway as a case study, the effectiveness of the proposed method is verified. Results indicate that the proposed stepwise collaborative solution process can output a second-level train working diagram and station track utilization scheme within 5 minutes, which outperforms the existing methods in efficiency and accuracy. Compared with the actual scheme, the optimized one increases the average train travel speed by 1.39 km·h-1, and reduces the average train connection time by 1.03 minutes. It achieves uniform utilization of station tracks, thus improves the quality of train working diagram. Moreover, the proposed method outputs the optimal solution under different parameter settings, demonstrating excellent applicability and stability

This study addresses the vehicle routing problem for hazardous materials transportation in a half-open multi-depot network. First, a risk assessment model integrating load factor, driving speed, vehicle type, and road type was developed to comprehensively quantify population exposure risk and environmental damage risk. Then, considering the impact of vehicle load and driving speed on fuel consumption and carbon emissions, the study proposed a multi-objective routing optimization model with time-dependent and inventory constraints, aiming to minimize total transportation cost, total transportation risk, and total travel time. The risk-prioritized dominance mechanism was introduced into NSGA-II (Non-dominated Sorting Genetic Algorithm II) to solve the model. A case study based on real-world chlorine transportation data from Beijing was conducted. The results show that compared with the conventional VRP mode, half-open mode reduces total transportation risk by 15.59%, total travel distance by 10.44%, and total travel time by 3.96%. Increasing driving speed shortens travel time but raises transportation cost and risk, whereas reducing speed yields the opposite effects. Moreover, the magnitude of the impact of speed changes on travel time is significantly smaller than that on transportation cost and risk. These findings can provide reliable decision support for achieving safe, economical, and efficient collaborative optimization of hazardous materials transportation in multi-depot scenarios.

To improve the existing detection methods for driver distraction in balancing accuracy with cost and identifying distraction levels, this paper proposes a hierarchical recognition method using XGBoost and Sparrow Search Algorithm. The study extracted 46 features from driving behavior, eye movement, and physiological data during visual and cognitive distraction. Highly sensitive indicators for distraction type and degree were selected through repeated-measures Analysis of Variance (ANOVA) and effect size comparisons to build a high-sensitivity feature set. A recursive feature elimination algorithm was then applied for secondary feature selection from multi-source indicators, resulting in two distinct sets: a single-source feature set containing only driving behavior features and a multi-source feature set integrating driving behavior, eye movement, and physiological features. A hierarchical recognition framework was established, employing the Sparrow Search Algorithm to optimize hyperparameters of the XGBoost classifiers at each level. The study also evaluated the model performance using the single-source and multi-source feature sets as input to develop a low-cost, high-accuracy model. Results demonstrate that for distraction type recognition, the single-source and multi-source models achieved accuracies of 92.4% and 95.9%, respectively. The Sparrow Search Algorithm optimized model for type recognition consistently outperformed Support Vector Machine, Random Forest, and standard XGBoost models, with accuracy improvements of 4%~6%, 11%, and 7%. For distraction degree recognition, the multi-source model achieved the accuracies of 81.0% for cognitive and 84.9% for visual distraction, while the single-source model showed relatively poor performance. Therefore, the single-source model is recommended for low-cost type recognition, and the multi-source model for high-precision degree assessment. The proposed method enables cost-effective and accurate identification of driver distraction, providing theoretical support for developing context-aware driver state monitoring systems.