Research Of Real-Time Detection And Attributes For Fast Vehicles

As one of the key components of the intelligent traffic system,video surveillance collects real-time road traffic pictures,providing real and reliable information for the road traffic managers to assist the research and determination,effectively improving the efficiency of road management and reducing the working pressure of the managers.At the same time,the amount of video surveillance and the data it generates are also proliferating,which not only makes it difficult for managers to use efficiently,but also puts enormous pressure on long-term storage.Due to various factors,such traditional video surveillance can no longer meet the daily needs of managers.Therefore,some tools are needed to extract useful information from these videos and filter out most redundant and low-value information.First,we collected real-world road data recorded by surveillance cameras in different road conditions and different weather conditions,including urban road,expressway,sunny,rainy,cloudy,night,congestion.Sencondly,we conducted the research of real-time detection for fast vehicles,and on the basis of the real-time and precise location,we carried out the research of joint vehicle detection and attributes for fast vehicles.Research of Fast Vehicle Detection Network(FVDN).Firstly,k-means ++ clustering algorithm is adopted to optimize the preset value of anchor box,so that the model can better adapt to a large variance of scales during training.Secondly,we present Rapidly Extract Module(REM),which is to quickly shrink the feature map size and enhance the run-time efficiency.Then,we present the Two Stream & Two Dual-Cross Connection(TS-TDC)Module to increase the network's width for enhancing features utilization rate and improving the performance.Further,we present the Cross-Stage and Multi-Scale(CS-MS)Module to optimize scale-invariant design because that vehicles on video surveillance are always with a large variance of scales.Finally,we used the visual loss function curve and mAP curve to adjust the network hyper-parameters.Research of Real-Time Detection and Attributes for Fast Vehicles(DAFV).Only with high precision location,high accuary attributes recognition is possible.In view of the common shortcomings of the existing algorithms,we proposed a joint algorithm.Firstly,according to the rich correlation information of those attributes,which plays an auxiliary role in detection,we achieved the combination of detection and recognition through multi-task learning and multi-label learning.Secondly,by referring to the design of ResNeXt block,we optimized the feature extractor of FVDN to enrich the feature.At the last,we used a similar configuration for training,and we performed functional tests,performance tests and ablation experments.