Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (15): 206-213.DOI: 10.3778/j.issn.1002-8331.2204-0105

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Pedestrian Multi-Object Tracking Algorithm for Intelligent Driving

YAN Chenyang, LIU Hongzhe, XU Cheng, LI Xuewei   

  1. 1.Beijing Key Laboratory of Information Service Engineering, Beijing Union University, Beijing 100101, China
    2.Institute for Brain and Cognitive Sciences, College of Robotics, Beijing Union University, Beijing 100101, China
  • Online:2023-08-01 Published:2023-08-01



  1. 1.北京联合大学 北京市信息服务工程重点实验室,北京 100101
    2.北京联合大学 机器人学院 脑与认知智能北京实验室,北京 100101

Abstract: Multi-object tracking(MOT) is a research hotspot in intelligent driving scenarios. Most modern MOT networks follow the “tracking-by-detection” paradigm, and the trajectory association of tracked objects is one of the hotspots that needs to be solved urgently. Aiming at the problem of missing detection caused by scene confusion and accidental occlusion caused by overlapping objects, which increases the difficulty of data association, a multi-object tracking algorithm PDTNet for intelligent driving that integrates attention mechanism and anchor-free detection is proposed. Firstly, the pyramid segmentation attention module is integrated into the deep aggregation network to improve the representation ability of multi-scale features. Secondly, a simple re-identification module is designed. The target detection obtained by the anchor-free frame detector is combined with the existing tracking trajectories to perform multi-step matching to achieve robust multi-object tracking. The experimental results show that the effectiveness of the proposed algorithm is verified on the MOT16, MOT17 datasets and BUIISE datasets, and the proposed algorithm improves the detection accuracy, association accuracy and total tracking accuracy of multi-object tracking. It has great application in target tracking scenarios.

Key words: multi-object tracking, intelligent driving, attention mechanism, deep aggregation network

摘要: 多目标跟踪(multi-object tracking,MOT)是智能驾驶场景中的一个研究热点,大多数现代MOT网络遵循“逐检测跟踪”范式,跟踪目标的轨迹关联是其中一个急需解决的热点问题。针对场景混乱以及意外的遮挡造成的对象重叠往往会导致遗漏检测,进而增加了数据关联的难度等问题,提出融合注意力机制和无锚框检测的智能驾驶多目标跟踪算法PDTNet。将金字塔分割注意力模块融入深层聚合网络,提高多尺度特征的表示能力;设计一个简单Re-identification模块,将由无锚框检测器获得的目标检测与已有的跟踪轨迹相结合进行多步匹配,实现强鲁棒性的多目标跟踪。实验结果表明,在MOT16、MOT17数据集和BUUISE数据集上验证了算法的有效性,提高了多目标跟踪的检测准确率、关联准确率以及跟踪总精度等,在智能驾驶多目标跟踪场景中有很大应用。

关键词: 多目标跟踪, 智能驾驶, 注意力机制, 深层聚合网络