基于深度学习和特征信息关联的多行人目标跟踪算法  

Multi-pedestrian target tracking algorithm based on deep learning and feature information association

作　　者：潘继财[1] PAN Jicai(Institute of Scientific and Technical Information of China,Beijing 100038,China)

机构地区：[1]中国科学技术信息研究所,北京100038

出　　处：《电子设计工程》2022年第9期31-36,共6页Electronic Design Engineering

摘　　要：多行人目标跟踪是智能安防监控系统的关键技术之一,其跟踪准确度的高低直接关系到监控系统的效果。针对复杂监控场景下多行人目标跟踪困难的问题,提出了一种YOLOv3网络模型与SORT跟踪算法相结合的鲁棒跟踪方法。通过简化网络模型输出以提高模型效率,对YOLOv3模型针对行人检测数据集进行重新训练。为了避免因长时间遮挡导致的目标跟踪失败,设计行人重识别网络(Re-ID)来提取目标表征特征,并通过计算特征向量的余弦距离来判别帧间行人目标的关联程度。实验结果表明,文中设计的改进YOLOv3检测器使行人检测率有明显的提高,提出的行人目标跟踪算法有效提高了跟踪效果,在MOT16数据集上的跟踪准确率和跟踪精准率相比于SORT算法分别提高了15.72%和3.14%。Multi-pedestrian target tracking is one of the key technologies of intelligent security monitoring system,and its tracking accuracy is directly related to the effect of the monitoring system. Aiming at the difficulty of multi-pedestrian target tracking in complex monitoring scenes,a robust tracking method combining YOLOv3 network model with SORT tracking algorithm is proposed. Simplifying the output of the network model to improve the model efficiency,the YOLOv3 model is retrained for the pedestrian detection data set. In order to avoid target tracking failure caused by long-time occlusion,a pedestrian recognition network(Re-ID) is designed to extract the characteristics of the target representation,and the cosine distance of the feature vector is calculated to determine the degree of association between pedestrian targets inter frame. The experimental results show that the improved YOLOv3 detector designed in this paper has significantly improved the pedestrian detection rate,the proposed pedestrian target tracking algorithm effectively improves the tracking effect and the MOTA and MOTP are improved by 15.72% and 3.14% respectively on MOT16 data set compared with SORT algorithm.

关 键 词：多行人跟踪 YOLOv3模型 目标检测 Re-ID网络 特征信息关联 

分 类 号：TN919.8[电子电信—通信与信息系统]