| With the continuous development of deep learning,computer vision has been widely used in practical life,which has laid the foundation for the construction of intelligent monitoring systems.Pedestrian tracking is a key component of intelligent monitoring systems and has high research and application value.Single-camera tracking has been widely used,but it cannot meet the long-term tracking needs of pedestrians in practical applications,and its monitoring range is also greatly limited.Cross-camera tracking can track pedestrians for a long time and over a large range through multiple cameras,thus overcoming the inherent defects of singlecamera tracking.Against this background,this paper carried out a series of studies on cross-camera pedestrian tracking and re-identification under non-overlapping views.This paper first researched the feature extraction network,pedestrian re-identification,and target tracking algorithm respectively,and then based on the proposed research techniques,carried out research on cross-camera pedestrian tracking and re-identification under non-overlapping views.The specific research contents are as follows:1.To better fuse the channel and spatial information of image features,and improve the ability of the feature extraction network to recognize targets in images,this paper designed a new attention mechanism that fills the feature maps obtained by large kernel convolution with the image’s features.This allows feature maps to be equally decomposed in the spatial dimension,and the small feature blocks obtained are finally concatenated in the channel dimension and fused with the original feature map information.This method was experimented on multiple datasets and compared with three existing classic attention mechanisms,and achieved the highest experimental performance.The robustness of this method can also be intuitively seen from the heat map visualizations.2.To achieve more accurate pedestrian matching in cross-camera pedestrian tracking under non-overlapping views,this paper designed a self-supervised learning network with a symmetric convolution structure.(1)To solve the problem that existing feature extraction networks are pretrained on the Image Net dataset and cannot extract granular information from pedestrian images and existing contrastive self-supervised learning may destroy the original properties of pedestrian images,this paper designed a mask reconstruction pre-training task to obtain a pre-trained model with strong robustness and used it for pedestrian re-identification tasks.(2)Improved the centroid-based triplet loss to train and optimize the pre-trained network model,incorporating masked images as additional samples into the loss calculation,so that the trained network can be better applied to pedestrian matching tasks.This method achieved an m AP that was about 5% higher than existing self-supervised learning pedestrian reidentification methods and a Rank1 that was about 1% higher on the Marker1501 and CUHK03 datasets.3.This paper presents methods for single and multiple pedestrian tracking and re-identification in cross-camera under non-overlapping views.(1)To address the problem of tracking failure caused by selfocclusion,mutual occlusion,or changes in camera perspective,this paper improved the Siam Mask single-object tracking algorithm and combined the above two works to complete the task of cross-camera single pedestrian tracking and re-identification,achieving good results on tracking after occlusion and tracking pedestrians from different perspectives.(2)For the problem of multi-camera multi-pedestrian matching,this paper first used the above self-supervised learning network to extract pedestrian features and calculated centroids for each pedestrian using their features across all historical frames to generalize their scale feature,resulting in better matching performance.Finally,the matched results were passed into BotSort multi-object tracking algorithm to complete the task of cross-camera multi-pedestrian tracking and re-identification. |
PDF Full Text Request