Research On Cross-camera Pedestrian Tracking Algorithm Based On Walking Characteristics

With the rapid development of economy and science and technology,monitoring equipment in our country is gradually found in every corner of public places,forming a complex monitoring network.Target tracking technology is widely used in daily video surveillance,but most of the existing tracking technologies are tracking within the field of view of a single camera,which can only achieve short-time tracking in a limited area.In order to solve the limitation of single-camera tracking,a cross-camera pedestrian tracking algorithm based on walking characteristics is proposed in this thesis.Aiming at the problem of low recognition rate of pedestrian detection,a pedestrian detection method with HOG-SILTP feature is proposed.This method extracts the HOG feature and SILTP feature of the image and performs dimensionality reduction and fusion,and uses SVM classifier to classify and detect pedestrians.Tested on INRIA dataset,the results show that,compared with HOG detection algorithm and HOG-LBP detection algorithm,the pedestrian recognition rate of this algorithm is improved by 10.5% and 8.22% respectively.Aiming at the problem of tracking frame size and target size cannot be adapted by the Kernel Correlation Filter(KCF)tracking algorithm in single camera pedestrian tracking,a scale filter is added.And a tracking abnormality judgment mechanism is introduced,and PSR and APCE are used to judge the tracking situation,which solves the problem that the KCF algorithm cannot judge the tracking abnormality in time and lead to the tracking failure.Tested on the OTB-100 dataset,the results show that the improved KCF tracking algorithm has good scale adaptability and robustness.Aiming at the problem of low recognition accuracy due to changes in carrying objects and clothing changes during cross-camera re-recognition,a pedestrian re-recognition method based on walking characteristics is proposed.In this method,the joint length and joint movement sequence are calculated by extracting multi-frame 2D joint points of pedestrians,and the walking sequence is obtained by fusion.Then the spatial and temporal features are extracted by CNN network and BLSTM network respectively,and then fused.Finally,the similarity between pedestrians is calculated for re-recognition.Tested on the Dataset B dataset,the results show that the average re-recognition accuracy of the proposed re-recognition algorithm is 85.44% in a single perspective.