基于LSTM概率多模态预期轨迹预测方法  

The Method of Probabilistic Multi-modal Expected Trajectory Prediction Based on LSTM

作　　者：高镇海[1] 鲍明喜 高菲[1] 唐明弘 Gao Zhenhai;Bao Mingxi;Gao Fei;Tang Minghong(Jilin University,State Key Laboratory of Automotive Simulation and Control,Changchun130022)

机构地区：[1]吉林大学,汽车仿真与控制国家重点实验室,长春130022

出　　处：《汽车工程》2023年第7期1145-1152,1162,共9页Automotive Engineering

基　　金：国家自然科学基金(52202495、U1564214)资助。

摘　　要：针对单模态轨迹预测无法充分表示未来预测空间以及解决轨迹预测固有的不确定性问题,本文构建了驾驶行为意图识别及交通车辆预期轨迹预测模型。驾驶行为意图识别模块识别被预测车辆车道保持、左换道、右换道、左加速换道和右加速换道的概率;交通车辆预期轨迹预测模块采用编码器-解码器架构,输出被预测车辆未来6 s内可能发生的多种行为和轨迹。通过HighD数据集对模型进行训练、验证与测试。试验结果表明:基于意图识别的预期轨迹预测模型生成的多模态概率分布可提高本车行驶安全性,与其他方法相比显著提高轨迹预测精度,在预测长时域轨迹上具有明显的优势。To address the problem that unimodal trajectory prediction cannot adequately represent the future prediction space and solve the inherent uncertainty of trajectory prediction,this paper constructs a driving behavior intention recognition and traffic vehicle expectation trajectory prediction model.The driving behavior intention recognition module identifies the probability of lane keeping,left lane change,right lane change,left acceleration lane change and right acceleration lane change of the predicted vehicle;the traffic vehicle expected trajectory prediction module uses an encoder-decoder architecture to output multiple behaviors and trajectories of the predicted vehicle that may occur within the next 6 seconds.The model is trained,validated and tested with the HighD dataset.The test results show that the multi-modal probability distribution generated by the intention recognition-based expected trajectory prediction model can improve the driving safety of the vehicle,significantly improve the trajectory prediction accuracy compared with other methods,and have obvious advantages in predicting long time domain trajectories.

关 键 词：轨迹预测 行为意图识别 LSTM 交互式行为 

分 类 号：U463.6[机械工程—车辆工程]