Research On Pedestrian Tracking Across Multi-camera

Pedestrian tracking is widely applied in the field of public security.However,pedestrians always travel across multiple cameras,enjoying a high degree of freedom and being vulnerable to environmental interference.The existing multi-camera pedestrian tracking algorithm cannot meet the application requirements.Therefore,the research on pedestrian tracking across multicamera has attracted a lot of attention.This thesis focuses on some key technologies about pedestrian tracking across multi-camera,including pedestrian detection,visual tracking and pedestrian re-identification.Moreover,the software is developed for pedestrian tracking across multi-camera.The main contents of this thesis are as follows:With respect to pedestrian detection,a pedestrian detection algorithm based on HOG feature and a pedestrian detection algorithm based on Faster R-CNN architecture are implemented.The Caltech Pedestrian Detection Benchmark is used to analyze the efficiency and accuracy of the two methods in feature extraction and pedestrian detection.The results show that the former is superior in computational efficiency,while the latter is better in accuracy.With respect to visual tracking,an improved ECO tracking algorithm is proposed.Aiming at the problem that ECO tracker is short for rotation,occlusion and other complex scenarios.Our algorithm optimizes the update strategy of the correlation filter,and proposes a tracking performance evaluation criterion.Firstly,the HOG feature and Color Names feature are fused by efficient convolution operators,and the dimension of the feature matrix is reduced.Secondly,the correlation between the feature matrix and the sample model is computed.The target is located by searching the region with the greatest correlation response.Finally,the tracking performance is evaluated based on the correlation response distribution,and the strategy of correlation filter updating is optimized according to the evaluation.The experiments on OTB-50 and OTB-100 show that our algorithm achieves the results with Precision Plot of 88.4% and 84.7%,and with Success Plot of 84.1% and 77.8%,respectively.The performance is improved in illumination variation,occlusion,rotation and other complex scenarios compared with the baseline algorithm.With respect to pedestrian re-identification,a pedestrian re-identification algorithm is proposed based on pedestrian orientation and body region guided feature decomposition and fusion.Aiming at the difficulty of angle adaptability of the traditional body region based pedestrian recognition algorithm,the strategies of feature fusion and body region segmentation are optimized.Firstly,the convolutional neural network is used to pose estimation,body segmentation and orientation analysis.Secondly,the feature fusion networks are trained under different matching relationships based on orientation analysis.Finally,the Triplet Loss is employed for metric learning to obtain the pedestrian re-identification model.The experiments on Market-1501 show that our algorithm can achieve a Rank-1 accuracy of 87.5% and a mAP of 70.62%.With respect to software development,the software for pedestrian tracking across multicamera is developed based on Matlab2016 b and the deep learning framework named Caffe.The functions of pedestrian detection,visual tracking,pedestrian re-identification and pedestrian tracking across multi-camera are implemented.