基于深度学习的行人目标检测与跟踪

Pedestrian detection and tracking based on deep learning

传统的目标检测跟踪方法存在实时性低、准确性差等问题。采用TBD(Tracking by Detection)方式,基于目标检测算法YOLOv5和多目标跟踪算法DeepSort,提出一种深度学习的多目标检测与跟踪方法。通过K-means聚类算法得出最佳聚类anchor大小对输入的视频帧中不同大小的目标进行检测和跟踪。基于残差结构的卷积神经网络构建行人重识别网络,在Market-1501数据集上离线训练提取行人特征。实验结果表明,算法在Crowdhuman数据集的测试集上的检测精度可达83.6,验证了检测算法的准确性;将算法应用在连续的视频帧上,可为同一个目标分配并保持相同的身份ID,在保证跟踪实时性的同时,验证了算法的可行性。

Traditional target detection and tracking methods have problems such as low real-time performance and poor accuracy.Using TBD(Tracking by Detection,TBD)method,based on the target detection algorithm YOLOv5 and the multi-target tracking algorithm DeepSort,a multi-target detection and tracking method under deep learning is proposed.The K-means clustering algorithm is used to obtain the optimal cluster anchor size to detect and track targets of different sizes in the input video frame.A pedestrian re-identification network is built based on the convolutional neural network of the residual structure,and the pedestrian features are extracted by offline training on the Market-1501 dataset.The experimental results show that the detection accuracy of the algorithm on the test set of the Crowdhuman dataset can reach 83.3%,which verifies the accuracy of the detection algorithm;applying the algorithm to continuous video frames can assign and maintain the same ID for the same target,while ensuring the real-time tracking,the feasibility of the algorithm is verified.