Cross-camera Multi-target Tracking Algorithm Based On Human Keypoint Detection Siamese Location Network

Cross-camera multi-target tracking refers to the continuous tracking of multiple targets within the field of view of multiple cameras,generating their complete motion trajectories.It is widely used in video surveillance,intelligent transportation,crowd management and other fields.In recent years,with the development of deep learning,cross-camera multi-target tracking has been viewed as a trajectory matching task,including single-camera and cross-camera trajectory matching.The most common method is to use the appearance features of pedestrians extracted by deep neural network models as the matching basis.Existing work often focuses on how to extract discriminative features from the whole pedestrian image,ignoring background interference and significant changes in pedestrian posture in cross-camera situations.Once there is occlusion or pedestrian overlap during tracking,the discriminative power of the pedestrian appearance features extracted from the whole image will decrease,resulting in trajectory misalignment.Therefore,this thesis conducts in-depth research on cross-camera multi-target tracking to address the above problems,introducing human keypoint information to reduce background interference,and designing a Transformer based Siamese location network to extract more accurate pedestrian overall and keypoint features.The main contributions are as follows:(1)A cross-camera multi-target tracking algorithm based on human keypoint detection is proposed.To address the problem that pedestrian images generated by object detection algorithms contain a large amount of background information,this thesis designs a human keypoint detection method to generate images of each keypoint of pedestrians.In the process of trajectory matching,the global features of the complete pedestrian image and the local features of the keypoints detected in both trajectories are combined for similarity calculation,in order to reduce the impact of background and posture changes on the appearance description of pedestrians.In the process of cross-camera trajectory matching,based on the similarity of global and local features,a hierarchical clustering algorithm with added constraints is used to integrate trajectories in single cameras into complete cross-camera trajectories.Finally,the effectiveness of the proposed algorithm is demonstrated on public datasets.(2)A cross-camera multi-target tracking algorithm based on keypoint Siamese location network is proposed.The feature extraction model based on convolutional neural network is limited by the size of the receptive field and often only focuses on the most prominent parts of the image.At the same time,when extracting features of irregular pedestrian keypoints in the image using square convolution kernels,it will also cover some background areas,resulting in keypoint features that include some background features.Therefore,to address the above issues,this thesis proposes a Siamese Transformer Localization Network.Compared with the feature extraction model based on convolutional neural network in existing work,the self-attention mechanism in the Transformer network can obtain global perception of the entire image,enabling the model to focus on more positions in the image.The Siamese network with shared weights can obtain differential features between the original pedestrian image and the keypoint image with the background erased.The location module can locate the background part in the differential features,and remove this part from the original features,which can retain the keypoint features as much as possible and improve the accuracy of the model in describing the overall appearance and keypoint positions of pedestrians.Finally,the effectiveness of the proposed algorithm is verified on publicly available cross-camera multi-target tracking and pedestrian re-identification datasets.There are 29 figures,17 tables and 98 references in this thesis.