汽车工程 ›› 2023, Vol. 45 ›› Issue (6): 1022-1030.doi: 10.19562/j.chinasae.qcgc.2023.06.012

所属专题: 智能网联汽车技术专题-感知&HMI&测评2023年

• • 上一篇    下一篇

典型汽车碰撞事故场景中行人运动轨迹预测方法

韩勇1,2(),林旭洁1,黄红武1,2,蔡鸿瑜1,罗金镕1,李燕婷1   

  1. 1.厦门理工学院机械与汽车工程学院,厦门  361024
    2.福建省新能源汽车与安全技术研究院,厦门  361024
  • 收稿日期:2022-11-24 修回日期:2022-12-21 出版日期:2023-06-25 发布日期:2023-06-16
  • 通讯作者: 韩勇 E-mail:Yonghanxmut@gmail.com
  • 基金资助:
    国家自然科学基金(51775466);福建省财政厅专项(闽财教指)[2021]103号和福建省工信厅项目(2022G43)

An Approach for Predicting Pedestrian Trajectories in Typical Car Crash Scenarios

Yong Han1,2(),Xujie Lin1,Hongwu Huang1,2,Hongyu Cai1,Jinrong Luo1,Yangting Li1   

  1. 1.School of Mechanical and Automotive Engineering,Xiamen University of Technology,Xiamen  361024
    2.Fujian Institute of New Energy Vehicle and Safety Technology,Xiamen  361024
  • Received:2022-11-24 Revised:2022-12-21 Online:2023-06-25 Published:2023-06-16
  • Contact: Yong Han E-mail:Yonghanxmut@gmail.com

摘要:

为提高未来自动驾驶车辆对弱势道路使用群体的感知和决策融合的可靠性,本文提出一种基于目标检测算法(YOLOv5)、多目标跟踪算法(Deep-Sort)和社交长短时记忆神经网络(social-long short-term memory, Social-LSTM)的行人未来运动轨迹预测方法。结合YOLOv5检测和Deep-Sort跟踪算法,有效解决行人检测跟踪过程中目标丢失问题。提取特定行人目标历史轨迹作为预测框架的输入边界条件,并采用Social-LSTM预测行人未来运动轨迹。并对未来运动轨迹进行透视变换和直接线性变换,转换为世界坐标系中的位置信息,预测车辆与行人的可能未来碰撞位置。结果显示目标检测精度达到93.889%,平均精度均值达96.753%,基于高精度的检测模型最终轨迹预测算法结果显示,预测损失随着训练步长的增加呈递减趋势,最终损失值均小于1%,其中平均位移误差降低了18.30%,最终位移误差降低了51.90%,本研究可为智能车辆避撞策略开发提供理论依据和参考。

关键词: 汽车碰撞行人事故, 行人轨迹预测, 目标检测, 多目标跟踪, 社交长短时记忆神经网络

Abstract:

In order to improve the reliability of perception and decision fusion of vulnerable road users in future autonomous vehicles, in this paper, a pedestrian future trajectory prediction method based on target detection algorithm (YOLOv5), multi-target tracking algorithm (Deep-Sort) and Social-Long Short-Term Memory (Social-LSTM) neural network is proposed. Combining the YOLOv5 detection and Deep-Sort tracking algorithms, the problem of target loss during pedestrian detection and tracking can be effectively solved. A specific pedestrian target history trajectory is extracted as the input boundary conditions of the prediction framework, and Social-LSTM is used to predict the pedestrian future motion trajectory. The future motion trajectory is also subjected to perspective transformation and direct linear transformation, which is then converted into position information in the world coordinate system to predict the possible future collision locations of vehicles and pedestrians. The results show that the target detection accuracy reaches 93.889% and the average accuracy reaches 96.753%. The results of the final trajectory prediction algorithm based on the high accuracy detection model show that the prediction loss is decreasing with the increase of the training step, and the final loss values are less than 1%, among which the average displacement error is reduced by 18.30% and the final displacement error is reduced by 51.90%. This study can provide a theoretical basis and reference for the development of intelligent vehicle collision avoidance strategies.

Key words: car to pedestrian crash, pedestrian trajectory prediction, target detection, multi-target tracking, Social-LSTM