Abstract: To handle the difficulties in the dense pedestrian objects detection and trajectory tracking, as well as a lack of motion semantic information analysis, we propose a method based on FairMOT network and K-means cluster for pedestrian spatial-temporal trajectory characteristic extraction in the dense crowd. First, we obtain the pedestrians' crossing street motion feature vectors from the surveillance video clips by tracking each object. Then we leverage a covariance filtering method STCCF to exclude the abnormal trajectory data and generate a trajectory set. We further investigate the semantic information in the trajectory set by the K- means algorithm utilizing the S coefficient (Silhouette Coefficient) and DB index (Davies Bouldin Index) as indicators. Experimental results show our algorithm successfully identify 179 abnormal trajectories in 2689 extracted trajectories, with the detected rate at 81.73% . The semantic information that represents where the pedestrians come and where they leave is perceived. The results verify the effectiveness of our proposed method in both trajectory extraction and motion semantic analysis.

Key words: intelligent transportation, trajectory extraction and motion semantic perception, FairMOT, dense crowd, trajectory point clustering, K-means

摘要： 针对基于视频监控的密集行人群识别难度大，运动轨迹提取困难，运动语义信息挖掘不足 等问题，本文提出基于多目标跟踪FairMOT框架及K-means聚类的行人轨迹捕获和运动语义信 息感知方法。首先，利用多目标跟踪算法提取视频中行人群目标过街时的运动轨迹特征向量；然 后，通过分析轨迹坐标的数值分布特点，设计了一种协方差滤波算法STCCF，以检测和剔除“准静 态轨迹”，得到行人群目标运动轨迹簇；最后，针对已提取的轨迹簇，应用K-means聚类方法，选取 S系数(Silhouette Coefficient)和DB指数(Davies Bouldin Index)两个指标，感知行人聚集和消散点 等场景语义特征。实验分析表明，算法从提取到的2689条轨迹中辨识出179条“准静态轨迹”，检 出率为81.73%；视频场景中的行人目标源点和消失点的解析结果与人工辨识结果吻合，验证了密 集行人群轨迹提取和运动语义信息感知方法的有效性。本文研究可为数据驱动的行为预测和轨 迹建模提供基础。

关键词: 智能交通, 轨迹提取和运动语义感知, FairMOT, 密集行人群, 轨迹点聚类, K-means