交通运输系统工程与信息

交通运输系统工程与信息 ›› 2024, Vol. 24 ›› Issue (5): 79-90.DOI: 10.16097/j.cnki.1009-6744.2024.05.008

• 智能交通系统与信息技术 • 上一篇    下一篇

基于改进视觉算法的自动驾驶风险预判模型

赵红专*1, 2，张继康1，潘佳雯1, 3，袁泉4，许恩永2，魏金占5，周旦1，刘承堃1   

  1. 1. 桂林电子科技大学，广西智慧交通重点实验室，广西 桂林 541004；2. 东风柳州汽车有限公司，商用车技术中心，广西 柳州 545000；3. 桂林电子科技大学，南宁研究院，南宁 530000；4. 清华大学，车辆与运载学院，北京 100084；5. 桂林航天工业学院，电子信息与自动化学院，广西 桂林 541004
  • 收稿日期:2024-05-17 修回日期:2024-07-26 接受日期:2024-07-31 出版日期:2024-10-25 发布日期:2024-10-22
  • 作者简介:赵红专(1985- )，男，广西桂林人，副教授，博士。
  • 基金资助:
    国家自然科学基金(52362045, 52072214)；广西科技重大专项(桂科AA22068101)。

Automatic Driving Risk Prediction Model Based on Improved Vision Algorithm

ZHAO Hongzhuan*1, 2, ZHANG Jikang1, PAN Jiawen1, 3, YUAN Quan4, XU Enyong2,WEI Jinzhan5, ZHOU Dan1, LIU Chengkun1   

  1. 1. Guangxi Key Laboratory of Intelligent Transportation, Guilin University of Electronic Science and Technology, Guilin 541004, Guangxi, China; 2. Commercial Vehicle Technology Center, Dongfeng Liuzhou Automobile Company Limited, Liuzhou 545000, Guangxi, China; 3. Nanning Research Institute, Guilin University of Electronic Science and Technology, Nanning 530000, China; 4. School of Vehicles and Transportation, Tsinghua University, Beijing 100084, China; 5. Faculty of Electronic Information and Automation, Guilin University of Aerospace Technology, Guilin 541004, Guangxi, China
  • Received:2024-05-17 Revised:2024-07-26 Accepted:2024-07-31 Online:2024-10-25 Published:2024-10-22
  • Supported by:
    National Natural Science Foundation of China (52362045, 52072214)；Guangxi Science and Technology Major Special Project (桂科AA22068101)。

PDF

摘要： 针对传统车辆切入过近导致自动驾驶产生脱离的问题，本文提出一种YOLOV7-Tiny (YouOnly Look Once Version 7 Tiny)和 SS-LSTM(Strong Sort Long Short Term Memory)的自动驾驶风险预判模型。首先，模型改进了视觉目标检测模型YOLOV7-Tiny，增加小目标检测层；其次，引入SimAM (A Simple, Parameter-Free Attention Module for Convolutional Neural Networks)无参注意力机制模块，优化训练损失函数，并对其目标车辆进行轨迹跟踪及预测，通过改进的多目标跟踪算法 StrongSORT(Strong Simple Online and Realtime Tracking)的短期预测不断矫正 LSTM(Long Short Term Memory)的长期预测，即建立SS-LSTM模型，并将预测的超车轨迹与智能网联车自身轨迹在同一时间纬度下进行拟合，得到传统车辆切入时的风险预判模型。实验结果表明，本文的自动驾驶风险预判方法有效预判了传统车辆切入时的风险。仿真实验表明，改进YOLOV7-Tiny相比于原有算法 mAP(mean Average Precision)提高了 2.3 个百分点，FPS(Frames Per Second)为61.35 Hz，模型大小为12.6 MB，模型满足车载端轻量化的需求。实车实验表明，根据SS-LSTM模型所得到的风险预判准确率为90.3%。

关键词: 交通工程, 风险预判, YOLOV7-Tiny, 自动驾驶, 长短期记忆网络, 轨迹预测

Abstract: In order to deal with the problem of traditional vehicles cutting too close to each other resulting in disengagement of automatic driving, this paper proposes an automatic driving risk prediction model with improved YOLOV7-Tiny and SS-LSTM. The model improves the visual target detection model YOLOV7-Tiny(You Only Look Once Version 7 Tiny), adds a small target detection layer, introduces the SimAM (A Simple, Parameter-Free Attention Module for Convolutional Neural Networks) attention mechanism module, optimizes the training loss function, and performs trajectory tracking and prediction of its target vehicle. The short-term prediction of Strong SORT (Strong Simple Online and Realtime Tracking) is utilized to continuously correct the long-term prediction of LSTM (Long Short Term Memory) to establish the SS-LSTM model. And the predicted overtaking trajectory is fitted with the trajectory of the intelligent networked vehicle itself at the same time latitude, so as to obtain the risk prediction model when the traditional vehicle cuts in. The experimental results show that the automatic driving risk prediction method in this paper effectively predicts the risk of traditional vehicles when cutting in, and the simulation experiments show that the improved YOLOV7-Tiny improves the prediction accuracy by 2.3% compared with the original algorithm mAP (mean Average Precision). The FPS (Frames Per Second) is 61.35 Hz. The model size is 12.6 MB, and the model meets the lightweight demand of the vehicle end. The real-vehicle experiments show that the accuracy of risk prediction based on the SS-LSTM model is 90.3%.

Key words: traffic engineering, risk prediction, YOLOV7-Tiny, autonomous driving, long short term memory networks, trajectory prediction

中图分类号: