| Video data-based intelligent tracking and target behavior analysis technologies have become emerging research directions in intelligent transportation systems.With the help of computers to learn and combine video sequences on the surveillance network,it can effectively determine the driving trajectory and behavior of target vehicles Category or potential risk and effectively address it.However,the switching of cameras in the surveillance network will also bring about changes in perspective,lighting,or other factors that cause the target to track with low accuracy in multiple camera locations.And processing a huge amount of video data will also cause the tracking process to be lengthy and slow.So the research idea of this paper is to first use a professional detection model to detect the video image and screen the candidate vehicle set.Based on the idea of vehicle re-identification and retrieval,the vehicle descriptor is formed by multi-feature fusion.Through the joint distance measurement and ranking of vehicle descriptors,the optimal solution in the selected candidate vehicle set is determined to be the target vehicle for cross-camera vehicle tracking.In order to verify the effectiveness of the research route selected by the subject,and then to construct a single vehicle cross-camera tracking system with high accuracy and robustness,the paper has carried out research in many aspects and angles.(1)Select the SSD model as the basic structure to construct the vehicle detection module.After an in-depth understanding and analysis of the shortcomings of the original network,the Inception-dense-SSD network structure was established in conjunction with the latest technical routes(such as dense modules,fusion of multi-scale features,normalized loss functions,etc.),and Effective training with professional data sets.The effectiveness and superiority of the improved algorithm in complex backgrounds,occlusions,and scale changes are verified through detection experiments under different data sets and in actual scenes.(2)Study the application of the multi-task learning framework in the field of vehicle re-identification,and improve the Rep Net framework according to the actual situation.The Dense Net-201 model structure with better performance and higher speed was used to replace its original feature extraction structure,and the L2 norm normalization process for the fully connected layer was added before the Triplet constraint in the metric branch.The realization of more distinguished fine-grained features of the vehicle(such as the details of the paste of the annual inspection logo,the placement of interior decorations,body scratches,hood bumps and other detailed information)extraction.(3)Study the vehicle cross-camera tracking strategy based on multimodal information for comprehensive vehicle characterization.After each input image frame passes the vehicle detection module,the vehicle pictures are extracted and composed to detect the vehicle picture set,and then the color similarity,SURF similarity,and multi-task learning between each image to be retrieved and the target vehicle picture are used The fine-grained feature similarity extracted by the frame is used to determine the joint distance.The optimal search result of the vehicle is selected without the joint distance exceeding the established threshold parameter,thereby achieving continuous tracking of the target between different camera positions.In summary,based on the algorithm design ideas of detection-multi-dimensional feature extraction-re-identification and retrieval,the paper studies and discusses the relevant theories and solutions of each module.Based on the actual needs,the implementation scheme of each module was optimized and optimized.The effective tracking test of the tracking algorithm model finally established showed that the false detection rate was reduced to a certain extent,and the tracking accuracy was more than 70%.For multi-camera vehicles The study of tracking has some significance. |
PDF Full Text Request