车辆视角下的行人穿行意图识别与行为预测
Pedestrian intention recognition and action prediction from the perspective of vehicles
何友国 1孙义芝 1蔡英凤 1袁朝春 1田力威2
作者信息
- 1. 江苏大学 汽车工程研究院,江苏 镇江 212001
- 2. 沈阳大学 信息工程学院,沈阳 110044
- 折叠
摘要
预测行人穿行马路的行为一直是智能驾驶汽车研究领域的重点方向,对于行人安全保护至关重要.现有研究通常使用轨迹或姿势对行人的行为进行建模,但是行人行为复杂多变,需对其进行更加深层的语义解读.将行人的穿行意图与行为相结合,设计多任务网络来识别行人意图并预测行人行为.行人的穿行意图会影响其行为,故使用行人未来的行为作为先验来检测行人目前的意图和行为,同时考虑到行人周围的交通目标与自车运动对于行人的影响,设计特征融合模块来融合行人特征与交通目标特征.在自动驾驶数据集(PIE和JAAD)上验证实验模型,结果表明,该模型展现了其在行人意图与行为建模方面的优势.
Abstract
Predicting pedestrian crossing action is key for intelligent driving vehicles and is crucial to ensure the safety of pedestrians.Existing methods typically model pedestrian action using trajectories or postures,but pedestrian action is complex and variable.Without a deeper semantic interpretation of pedestrian behaviors,a better understanding of pedestrian behaviors can hardly be achieved.This paper investigates the integration of pedestrian crossing intentions and actions,and designs a multitask network to identify pedestrian intentions and predict pedestrian actions.It suggests pedestrian crossing intentions influence their actions and utilizes future pedestrian actions as prior information to detect current intentions and actions.Additionally,it considers the impact of surrounding traffic targets and vehicle movements on pedestrians,and designs a feature fusion module to integrate pedestrian features with traffic target features.Finally,our model is validated on two autonomous driving datasets (PIE and JAAD),showing superior results and demonstrating advantages in modeling pedestrian intentions and actions.
关键词
行人意图/行人行为/智能驾驶/行人安全Key words
pedestrian intention/pedestrian action/intelligent driving/pedestrian safety引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据