在线多行人图像目标跟踪算法设计与仿真

Design and Simulation of Online Multi Pedestrian Image Target Tracking Algorithm

由于已有算法在进行多行人跟踪时,存在跟踪结果误差较大的问题,提出一种基于YOLOv4和DeepSORT的在线多行人目标跟踪算法,采用YOLOv4对目标进行检测。将检测结果作为输入,根据检测结果初始化卡尔曼滤波器,建立目标运动模型;提取检测结果的外观特征,建立目标外观模型。DeepSORT算法利用运动模型和外观模型所得到的数据将新的检测结果与现有轨道联系起来,通过匈牙利算法进行关联度量,完成轨迹和检测的匹配,实现在线多行人目标跟踪。测试结果表明,所提算法能够有效提升在线多行人目标跟踪结果的准确性。

In order to improve the accuracy of multi pedestrian tracking results,an online multi pedestrian target tracking algorithm based on YOLOv4 and DeepSORT is proposed in this paper.Firstly,YOLOv4 was used to detect the target,and the detection result was regarded as the input content.Based on the detection content,the Kalman filter was initialized to establish the target motion model.The appearance features of the detection results were extracted and the target appearance model was established.The new detection results obtained from the dynamic model and appearance model of DeepSORT algorithm were combined with the existing track.Based on the Hungarian algorithm,the above results were correlated and measured to complete the matching of track and detection.Finally,online multi-pedestrian target tracking was achieved.The results show that the algorithm can effectively improve the accuracy of online multi pedestrian target tracking.