改进YOLOv5+DeepSort的行人跟踪算法

Improved YOLOv5+DeepSort algorithm for pedestrian tracking

针对复杂环境道路行人跟踪易发生身份丢失、切换的问题,提出一种改进的YOLOv5检测并结合DeepSort跟踪算法。检测阶段,融合注意力模块CBAM与YOLOv5颈部网络增强对行人特征的提取;用SIoU边界框损失函数代替CIoU边界框损失函数,加速边界框回归的同时提高准确定位度。跟踪阶段,改进DeepSort利用拓展卡尔曼滤波器对非线性环境行人位置进行预测,通过匈牙利算法匹配预测和检测轨迹,优化复杂环境下行人身份切换频繁的问题。最后连接改进后的YOLOv5与DeepSort算法,对MOT⁃16数据集进行检测跟踪。实验结果表明:改进YOLOv5算法较原算法平均精准度提高4%,结合DeepSort跟踪,平均跟踪精确度为63.5%,比原始算法提升了3.4%;行人身份切换次数减少52次,比原始算法减少了6.5%。

An improved YOLOv5 detection combined with DeepSort tracking algorithm is proposed to solve the problem that the pedestrian identity is prone to loss and switch when tracking the pedestrian on roads in complex environments.In the detection stage,CBAM(convolutional block attention module)and the YOLOv5 neck network are integrated to enhance the extraction capacity of pedestrian features,and the CIoU bounding box loss function is replaced by the SIoU bounding box loss function,which accelerates the bounding box regression and improves the positioning accuracy.In the tracking stage,the extended Kalman filter(EKF)is used in the improved DeepSort to predict the position of pedestrians in a nonlinear environment,and the Hungarian algorithm is adopted to match the prediction and detection trajectories to optimize the frequent pedestrian identity switching in complex environments.Finally,the improved YOLOv5 and DeepSort algorithms are connected to detect and track the MOT⁃16 dataset.The experimental results show that the average accuracy of the improved YOLOv5 algorithm is increased by 4%compared with the original algorithm;in combination with Deepsort tracking algorithm,the average tracking accuracy is 63.5%,which is 3.4%higher than the original algorithm,and the number of pedestrian identity switching is reduced by 52 times,which is 6.5%less than the original algorithm.