Explainable lane change intention prediction based on LSTM-multi-head mixed attention

doi:10.3969/j.issn.1674-8484.2024.05.014

Abstract

Abstract:

An interpretable lane change intention prediction model was proposed to enable the autonomous vehicle to accurately predict the lane change intention of the vehicles around them. This model based on the Long Short-Term Memory (LSTM) and the multi-head mixed attention, which can fully extract the spatiotemporal interaction between the target vehicle and its surrounding vehicles. A Shapley additive interpretation method (SHAP) based on maximum entropy was proposed to explain the degree of influence of each feature on the model output at a specific time step, and experiments on the HighD dataset were carried out. The results show that the comprehensive accuracy of the proposed model is 4.03%, 9.51%, and 5.16% higher than that of the LSTM, the Convolutional Neural Network (CNN), and the multi-head attention, respectively, before lane change, which fully proves the validity of the model in the long time horizon. And the wrong prediction samples can be attributed to model defects or sparse samples on the other hand, guiding users to optimize the model.

Key words: autonomous driving vehicles, lane change intention prediction, attention mechanism, long short-term memory (LSTM), Shapley additive explanations (SHAP)

CLC Number:

U461

GAO Kai, LIU Jian, LIU Linhong, LIU Xinyu, ZHANG Jinlai, DU Ronghua. Explainable lane change intention prediction based on LSTM-multi-head mixed attention[J]. Journal of Automotive Safety and Energy, 2024, 15(5): 763-773.

Figures/Tables 9

References 16

[1]	YANG Xing, CHEN Lu, WANG Huaji, et al. Driver lane change intention inference for intelligent vehicles: Framework, survey, and challenges[J]. IEEE Trans Vehi Tech, 2019, 68(5): 4377-4390.
[2]	LUO Wenjie, YANG Bin, Urtasun R, et al. Fast and furious: Real time end-to-end 3D detection, tracking and motion forecasting with a single convolutional net[C]// Proc IEEE Conf Computer Vision Pattern Recog, 2018: 3569-3577.
[3]	徐杰, 裴晓飞, 杨波, 等. 融合车辆轨迹预测的学习型自动驾驶决策[J]. 汽车安全与节能学报, 2022, 13(2): 317-324.
	XU Jie, PEI Xiaofei, YANG Bo, et al. Learning autonomous driving decision making with vehicle trajectory prediction[J]. J Auto Safe Energ, 2022, 13(2): 317-324. (in Chinese)
[4]	蔡英凤, 邰康盛, 王海, 等. 无人驾驶汽车周边车辆行为识别算法研究[J]. 汽车工程, 2020, 42(11): 1464-1472+1505. doi: 10.19562/j.chinasae.qcgc.2020.11.003
	CAI Yingfeng, TAI Kangsheng, WANG Hai, et al. Research on Vehicle Behavior Recognition Algorithm around driverless vehicles[J]. Autom Engineering, 2020, 42(11): 1464-1472+1505. (in Chinese)
[5]	XU Yiran, YANG Xiaoyin, GONG Lihang, et al. Explainable object-induced action decision for autonomous vehicles[C]// Proc IEEE/CVF Conf Computer Visi Pattern Recog, 2020: 9523-9532.
[6]	SI Nianwen, ZHANG Wenlin, QU Dan, et al. Fine-grained visual explanations for the convolutional neural network via class discriminative deconvolution[J]. Multimedia Tool Appl, 2022, 81(2): 2733-2756.
[7]	Liessner R, Dohmen J, Wiering M A. Explainable reinforcement learning for longitudinal control[C]// ICAART (2) (2), 2021: 874-881.
[8]	CUI Zhihao, LI Meng, HUANG Yanjun, et al. An interpretation framework for autonomous vehicles decision-making via SHAP and RF[C]// 2022 6th CAA Int'l Con Vehi Contr Intell (CVCI), 2022: 1-7.
[9]	LI Meng, WANG Yulei, SUN Hengyang, et al. Explaining a machine-learning lane change model with maximum entropy Shapley values[J]. IEEE Trans Intel Vehi, 2023, 8(6): 3620-3628.
[10]	Krajewski R, Bock J, Kloeker L, et al. The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems[C]// 2018 21st Int'l Conf Intell Transp Syst (ITSC), 2018: 2118-2125.
[11]	GAO Kai, LI Xunhao, CHEN Bin, et al. Dual transformer based prediction for lane change intentions and trajectories in mixed traffic environment[J]. IEEE Trans Intel Transport Syst, 2023, 24(6): 6203-6216
[12]	Doroudgar S, Chuang H M, Perry P J, et al. Driving performance comparing older versus younger drivers[J]. Traff Injur Prev, 2017, 18(1): 41-46.
[13]	QI Xuanhao, ZHI Min. A review of attention mechanisms in computer vision[C]// 2023 8th Int'l Conf Image, Visi Computing (ICIVC), 2023: 577-583.
[14]	NIU Zhaoyang, ZHONG Guoqiang, YU Hui, et al. A review on the attention mechanism of deep learning[J]. Neurocomputing, 2021, 452: 48-62.
[15]	CUI Yutao, JIANG Cheng, WANG Limin, et al. Mixformer: End-to-end tracking with iterative mixed attention[C]// Proc IEEE/CVF Conf Compu Visi Pattern Recog, 2022: 13608-13618.
[16]	Lundberg S M, Lee S-I. A unifed approach to interpreting model predictions[C]// Proc 31st Int’l Conf Neur Info Process Syst, Long Beach, California, USA, 2017: 4768-4777.

隐藏单元数	平均准确率/ %	平均损失	训练时间/ s
8	95.50	0.126	271.42
16	97.50	0.078	283.87
32	97.50	0.072	307.08
64	95.50	0.122	327.17
128	95.75	0.112	412.26

隐藏单元数	平均准确率/ %	平均损失	训练时间/ s
8	95.50	0.126	271.42
16	97.50	0.078	283.87
32	97.50	0.072	307.08
64	95.50	0.122	327.17
128	95.75	0.112	412.26

模型	意图	R = 3 s			R = 2 s			R = 1 s
模型	意图	准确率/ %	召回率/ %	F1分数%	准确率/ %	召回率/ %	F1分数%	准确率/ %	召回率/ %	F1分数%
本文模型	LLC	94.03	91.30	92.65	98.54	97.58	98.06	98.54	97.58	98.06
	RLC	88.27	83.57	85.86	95.15	94.69	94.92	95.15	94.69	94.92
	LK	87.50	94.36	90.95	93.81	95.17	94.48	93.81	95.17	94.48
	平均值	89.86	89.86	89.82	95.81	95.81	95.82	95.81	95.81	95.82
LSTM	LLC	89.57	91.30	90.43	95.71	97.10	96.40	95.77	98.55	97.14
	RLC	91.25	70.53	79.56	93.94	86.86	91.85	95.77	98.55	97.14
	LK	79.20	95.65	86.65	92.02	94.69	93.33	98.46	92.75	95.52
	平均值	85.83	85.83	85.55	93.88	93.88	93.86	96.62	96.62	96.60
CNN	LLC	81.68	79.71	80.68	89.16	87.44	88.30	97.51	94.69	96.08
	RLC	80.00	67.63	73.30	81.04	82.61	81.82	88.16	97.10	92.41
	LK	79.50	93.72	86.03	88.89	88.89	88.89	96.35	89.37	92.73
	平均值	80.35	80.35	80.00	86.31	86.31	86.33	93.72	93.72	93.74
多头注意力	LLC	88.67	86.96	87.80	95.53	90.82	92.16	93.69	93.24	93.69
	RLC	83.89	72.95	78.04	84.16	89.86	86.92	88.94	95.17	88.94
	LK	81.93	94.20	87.64	92.96	89.37	91.13	90.96	85.02	94.48
	平均值	84.70	84.70	84.50	90.02	90.02	90.07	91.14	91.14	91.19

模型	意图	R = 3 s			R = 2 s			R = 1 s
模型	意图	准确率/ %	召回率/ %	F1分数%	准确率/ %	召回率/ %	F1分数%	准确率/ %	召回率/ %	F1分数%
本文模型	LLC	94.03	91.30	92.65	98.54	97.58	98.06	98.54	97.58	98.06
	RLC	88.27	83.57	85.86	95.15	94.69	94.92	95.15	94.69	94.92
	LK	87.50	94.36	90.95	93.81	95.17	94.48	93.81	95.17	94.48
	平均值	89.86	89.86	89.82	95.81	95.81	95.82	95.81	95.81	95.82
LSTM	LLC	89.57	91.30	90.43	95.71	97.10	96.40	95.77	98.55	97.14
	RLC	91.25	70.53	79.56	93.94	86.86	91.85	95.77	98.55	97.14
	LK	79.20	95.65	86.65	92.02	94.69	93.33	98.46	92.75	95.52
	平均值	85.83	85.83	85.55	93.88	93.88	93.86	96.62	96.62	96.60
CNN	LLC	81.68	79.71	80.68	89.16	87.44	88.30	97.51	94.69	96.08
	RLC	80.00	67.63	73.30	81.04	82.61	81.82	88.16	97.10	92.41
	LK	79.50	93.72	86.03	88.89	88.89	88.89	96.35	89.37	92.73
	平均值	80.35	80.35	80.00	86.31	86.31	86.33	93.72	93.72	93.74
多头注意力	LLC	88.67	86.96	87.80	95.53	90.82	92.16	93.69	93.24	93.69
	RLC	83.89	72.95	78.04	84.16	89.86	86.92	88.94	95.17	88.94
	LK	81.93	94.20	87.64	92.96	89.37	91.13	90.96	85.02	94.48
	平均值	84.70	84.70	84.50	90.02	90.02	90.07	91.14	91.14	91.19

样本	SHAP值 / 10^-3
样本	左换道	右换道
ΔX	-3	12
D_f	-15	192
D_lf	104	216
D_rf	288	-916
D_lb	71	-65
D_rb	66	400
v_f	-5	59
v_lf	7	121
v_rf	26	166
v_lb	-7	170
v_rb	-5	27
模型检查	模型稀疏	模型缺陷