Vehicle Retrieval Using Deep Learning Under Cross-region

With the improvement of the security system,a huge monitoring network has gradually formed in the urban roads.The network plays a vital role in monitoring illegal traffic behavior,thus ensures the safety of people's production and life.Today,we mainly rely on human to track illegal vehicles,which greatly reduces retrieval efficiency.Therefore,the development of automated cross-regional vehicle retrieval system has important application values.But vehicle retrieval in complex environments faces a great challenge.because of the two reasons: First,the illumination,scale,and viewing angle of the same vehicle under different cameras are different.Second,different vehicles of the same model are very similar,which increases the difficulty and influences the performance of vehicle retrieval.To tackle the problems above,we propose a deep feature extraction model under multilocal regions.the model first uses Adaboost algorithm under a large number of training samples to locate the regions(car face and car windshield).Compared with the global appearance of vehicles,local regions can better identify the vehicle,thus improve the performance of vehicle retrieval.Based on local regions,we propose an improved two-branch deep feature extraction network on VGG model.Compared with the traditional methods,our model can extract and fuse the features of local regions effectively to ensure the accuracy of cross-region vehicle retrieval.Based on effective vehicle feature,we propose cross-region vehicle retrieval algorithm based on spatio-temporal constraint model uses Baidu map API.the vehicle retrieval algorithm considers the path to estimate the time period under other cameras of query vehicle in the future,and then calculation the visual and temporal similarity of the query vehicle and all candidate vehicles within the time period to determine where the query vehicle appears.In order to verify our method,we conduct different experiments on three data sets.The experimental results show that our model under multi-local regions can improve 16.87%and 11.14% respectively than global regions or car face of vehicle,which prove our model is more effective in identifying vehicles.The cross-region retrieval algorithm based on spatio-temporal constraint model speeds up retrieval(time is 260 millisecond)and ensurs the performance(mAP is 0.521)of vehicle retrieval.In addition,compared with the current methods,our method has better accuracy and can guarantee real-time vehicle retrieval.