用于自动驾驶车辆的融合注意力机制多目标跟踪算法

doi:10.3969/j.issn.1674-8484.2021.04.010

摘要/Abstract

摘要：

为提高自动驾驶车辆的跟踪准确性,建立了一种融合注意力机制的多目标跟踪算法。基于YOLOv3神经网络并融合注意力机制,增强了目标外观特征提取网络的性能。用该多目标检测算法,提取目标或背景的具有辨别性的特征。用长短时记忆（LSTM）网络,提取物体的运动特征。用该跟踪算法对目标轨迹进行动态建模。借助追踪目标的相似度数值和数据匹配关联,完成了多目标的跟踪任务。在多目标跟踪数据集MOT16上进行了实验。结果表明：与YOLOv3相比,考虑注意力机制的多目标检测算法的成功率提高了1.9%;该算法的准确度53.9%,精确度79.0%;因而,本算法实现了对目标的稳定跟踪。

关键词: 自动驾驶车辆, 多目标跟踪算法, 深度神经网络, 注意力机制

Abstract:

This paper established a multiple-object tracking algorithm for autonomous-vehicles integrated with an attention mechanism to improve the tracking accuracy. The multiple-object tracking algorithm extracted object appearance features by using YOLOv3 neural network with attention mechanism that enhanced the performance of feature extraction network. The algorithm extracted discriminative features of the target or the background. A Long Short-Term Memory (LSTM) network extracted the target motion features, while the object tracking algorithm modeled the target trajectory dynamically. The algorithm accomplished a multiple object tracking through data matching and association based on tracked target similarity degree. Experiments were done on a multiple object tracking dataset MOT16. The results show that the object detection success rate increases 1.9% compared with YOLOv3 network, using the object detection algorithm with attention mechanism with a tracking accuracy of 53.9% and a tracking precision of 79.0%. Therefore, the algorithm achieves a stable target tracking.

Key words: autonomous-vehicles, multiple-object tracking algorithm, deep neural networks, attention mechanism

中图分类号:

U461.99

张平, 迟志诚, 陈一凡, 惠飞. 用于自动驾驶车辆的融合注意力机制多目标跟踪算法[J]. 汽车安全与节能学报, 2021, 12(4): 516-521.

ZHANG Ping, CHI Zhicheng, CHEN Yifan, HUI Fei. Multiple object tracking algorithm integrated with attention mechanism for autonomous vehicles[J]. Journal of Automotive Safety and Energy, 2021, 12(4): 516-521.

图/表 8

参考文献 19

[1]	卢湖川, 李佩霞, 王栋. 目标跟踪算法综述[J]. 模式识别与人工智能, 2018, 31(1):61-76.
	LU Huchuan, LI Peixia, WANG Dong. Visual object tracking: a survey[J]. Patt Recogn Artif Intell, 2018, 31(1):61-76. (in Chinese)
[2]	崔洲涓, 安军社, 崔天舒. 基于多层深度卷积特征的抗遮挡实时跟踪算法[J]. 光学学报, 2019, 39(7):(0715002)1-14.
	CUI Zhoujuan, AN Junshe, CUI Tianshu. Real-time and anti-occlusion visual tracking algorithm based on multi-layer deep convolutional features[J]. Acta Optica Sinica, 2019, 39(7):(0715002)1-14. (in Chinese)
[3]	Reid D B. An algorithm for tracking multiple targets[J]. *IEEE Trans Automat Contr*, 1979, 24(6):843-854. doi: 10.1109/TAC.1979.1102177 URL
[4]	Fortmann T E, Bar-shalom Y, Scheffe M. Sonar tracking of multiple targets using joint probabilistic data association[J]. *IEEE J Oceanic Eng*, 1983, 8(3):173-184. doi: 10.1109/JOE.1983.1145560 URL
[5]	种衍文, 匡湖林, 李清泉. 一种基于多特征和机器学习的分级行人检测方法[J]. 自动化学报, 2012, 38(3):375-381.
	CHONG Yanwen, KUANG Hulin, LI Qingquan. Two-stage pedestrian detection based on multiple features and machine learning[J]. Acta Automatica Sinica, 2012, 38(3):375-381. (in Chinese) doi: 10.3724/SP.J.1004.2012.00375 URL
[6]	徐渊, 许晓亮, 李才年, 等. 结合SVM分类器与HOG特征提取的行人检测[J]. 计算机工程, 2016, 42(1):56-60, 65.
	XU Yuan, XU Xiaoliang, LI Cainian, et al. Pedestrian detection combining with SVM classifier and HOG feature extraction[J]. Comput Eng, 2016, 42(1):56-60, 65. (in Chinese)
[7]	Bewley A, Ge Z Y, Ott L, et al. Simple online and realtime tracking[C]// *IEEE Int'l Conf Image Process*, 2016: 3464-3468.
[8]	Ren S, He K, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. *IEEE Trans Patt Anal Mach Intell*, 2017, 39(6):1137-1149. doi: 10.1109/TPAMI.2016.2577031 URL
[9]	Wojke N, Bewley A, Paulus D. Simple online and realtime tracking with a deep association metric[C]// *IEEE Int'l Conf Image Process*, 2017: 3645-3649.
[10]	李勇, 杨德东, 韩亚君, 等. 融合扰动感知模型的孪生神经网络目标跟踪[J]. 光学学报, 2020, 40(4)(0415002)1-12.
	LI Yong, YANG Dedong, HAN Yajun, et al. Siamese neural network object tracking with distractor-aware model[J]. Acta Optica Sinica, 2020, 40(4): 1-12. (in Chinese)
[11]	Redmon J, Farhadi A. YOLOv3: An incremental improvement[J/OL]. arXiv preprint (2018-04-08). https://arxiv.org/abs/1804.02767v1, 2018.
[12]	Woo S, Park J, Lee J Y, et al. CBAM: convolutional block attention module[C]// *Europ Conf Comput Visi (ECCV)*, 2018: 3-19.
[13]	Hochreiter S, Schmidhuber J. Long short-term memory[J]. *Neural Comput*, 1997, 9(8):1735-1780. pmid: 9377276
[14]	Rezatofighi S H, Milan A, Zhang Z, et al. Joint probabilistic data association revisited[C]// *IEEE Int'l Conf Comput Visi (ICCV)*, 2015: 3047-3055.
[15]	Tang S, Andriluka M, Andres B, et al. Multiple people tracking by lifted multicut and person re-identification[C]// *IEEE Conf Comput Visi Patt Recogn (CVPR)*, 2017: 3701-3710.
[16]	Sadeghian A, Alahi A, Savarese S. Tracking the untrackable: learning to track multiple cues with long-term dependencies[C]// *IEEE Int'l Conf Comput Visi (ICCV)*, 2017: 300-311.
[17]	Kieritz H, Becker S, Hübner W, et al. Online multi-person tracking using integral channel features[C]// *IEEE Int'l Conf Adv Video Sign Based Surveill (AVSS)*, 2016: 122-130.
[18]	CHU Qi, OUYANG Wanli, LI Hongsheng, et al. Online multi-object tracking using CNN-based single object tracker with spatial-temporal attention mechanism[C]// *IEEE Int'l Conf Comput Visi (ICCV)*, 2017: 4846-4855.
[19]	Sanchez-matilla R, Poiesi F, Cavallaro A. Online multi-target tracking with strong and weak detections[C]// *Europ Conf Comput Visi (ECCV)*, 2016: 84-99.

模块	输入	输出
YOLO-v3网络	原始图像	目标位置、类别
注意力机制模块	目标特征图	加权目标特征图
LSTM网络	目标位置、类别	跟踪结果

模块	输入	输出
YOLO-v3网络	原始图像	目标位置、类别
注意力机制模块	目标特征图	加权目标特征图
LSTM网络	目标位置、类别	跟踪结果

方法	平均精度 / %	检测速度 /（帧·s^-1）
YOLO-v3	81.5	27
本文	83.4	25

方法	平均精度 / %	检测速度 /（帧·s^-1）
YOLO-v3	81.5	27
本文	83.4	25

算法	引文	MOTA↑/ %	MOTP↑/ %	MT↑/ %	ML↓/ %	FP↓	FN↓	IDS↓
JPDA_m	[14]	26.3	76.3	4.1	67.5	3 689	130 549	365
LMP	[15]	48.8	79.0	18.2	40.1	6 654	86 245	481
AMIR	[16]	47.2	75.8	14.0	41.6	2 681	92 856	774
oICF	[17]	43.2	74.3	11.3	48.5	6 651	96 515	381
STAM	[18]	46.0	74.9	14.6	43.6	6 895	91 117	473
EAMTT	[19]	38.8	75.1	7.9	49.1	8 114	102 452	965
本文		53.9	76.2	21.6	42.4	5 352	81 444	423