系统工程与电子技术 ›› 2025, Vol. 47 ›› Issue (7): 2371-2382.doi: 10.12305/j.issn.1001-506X.2025.07.29

• 制导、导航与控制 • 上一篇    

基于行人轨迹预测的无人车动态避障方法

张岳, 王晶, 王鼎衡, 代昌华, 段援朝, 刘保荣, 张霄嵩   

  1. 西北机电工程研究所, 陕西 咸阳 712000
  • 收稿日期:2024-09-18 出版日期:2025-07-16 发布日期:2025-07-22
  • 通讯作者: 张岳
  • 作者简介:张岳(1999—), 男, 研究实习员, 硕士, 主要研究方向为无人车自主导航、路径规划
    王晶(1986—), 女, 副研究员, 博士, 主要研究方向为目标轨迹预测
    王鼎衡(1988—), 男, 助理研究员, 博士, 主要研究方向为神经网络轻量化
    代昌华(1995—), 男, 助理研究员, 博士, 主要研究方向为车辆底盘控制、人机共驾
    段援朝(1994—), 男, 研究实习员, 硕士, 主要研究方向为即时定位与建图
    刘保荣(1998—), 男, 研究实习员, 硕士, 主要研究方向为多传感器融合目标检测
    张霄嵩(1998—), 男, 研究实习员, 硕士, 主要研究方向为目标检测、语义分割
  • 基金资助:
    咸阳市重大科技创新项目(L2023-ZDKJ-JSGG-GY-018)

Dynamic obstacle avoidance method for unmanned vehicle based on pedestrian trajectory prediction

Yue ZHANG, Jing WANG, Dingheng WANG, Changhua DAI, Yuanchao DUAN, Baorong LIU, Xiaosong ZHANG   

  1. Northwest Institute of Mechanical and Electrical Engineering, Xianyang 712000, China
  • Received:2024-09-18 Online:2025-07-16 Published:2025-07-22
  • Contact: Yue ZHANG

摘要:

针对无人车自主导航过程中对行人避障时未考虑目标运动趋势所导致的避障效果欠佳的问题, 提出一种基于行人轨迹预测的无人车动态避障方法。首先, 利用深度相机和目标检测算法完成行人的识别及位置解算, 并通过跟踪获得行人位置的时序信息。其次, 将一维卷积神经网络(convolutional neural network, CNN)和双向门控循环单元(bidirectional gated recurrent unit, BiGRU)相结合, 并融入注意力(attention, ATT)机制, 构建CNN-BiGRU-ATT行人轨迹预测模型, 由行人当前位置序列预测未来轨迹。最后, 通过Bresenham算法将行人当前位置及轨迹预测结果叠加在代价地图上, 使得全局和局部路径规划算法提前规避碰撞风险。实验表明, 基于深度相机的行人检测、定位方法能够准确计算行人历史轨迹, 为轨迹预测模型提供可靠的输入; 基于CNN-BiGRU-ATT的行人轨迹预测模型在多种运动状态下预测误差更小。融入轨迹预测结果的代价地图可提高路径规划的前瞻性, 避免无人车频繁更改全局和局部路径, 降低导航任务用时和碰撞风险。

关键词: 自主导航, 轨迹预测, 代价地图, 动态避障

Abstract:

A dynamic obstacle avoidance method for unmanned vehicle based on pedestrian trajectory prediction is proposed to address the problem of poor obstacle avoidance performance caused by the lack of consideration of target motion trends during autonomous navigation of unmanned vehicle. Firstly, use depth camera and object detection algorithm to complete pedestrian recognition and position calculation, and obtain temporal information of pedestrian position through tracking. Secondly, by combining one-dimensional convolutional neural network (CNN) with bidirectional gated recurrent unit (BiGRU) and incorporating attention (ATT) mechanism, a CNN-BiGRU-ATT pedestrian trajectory prediction model is constructed, which predicts future trajectories based on the current position sequence of pedestrians. Finally, the Bresenham algorithm is used to overlay the current position and trajectory prediction results of pedestrians on the cost map, enabling global and local path planning algorithms to avoid collision risks in advance. Experiments show that the pedestrian detection and localization method based on depth camera can accurately calculate the historical trajectory of pedestrian, providing reliable input for trajectory prediction model. The pedestrian trajectory prediction model based on CNN-BiGRU-ATT has smaller prediction errors in various motion states. The cost map integrated with trajectory prediction results improves the foresight of path planning, avoids frequent changes of global and local paths by unmanned vehicle, and improves the efficiency of task execution.

Key words: autonomous navigation, trajectory prediction, cost map, dynamic obstacle avoidance

中图分类号: