Research And Implementation Of Vehicle Region Of Interest Detection Algorithm In Traffic Image

With the rapid growth of the motor vehicles,there are a series of problems about serious safety and traffic in China.At the same time,the amount of video image files grows explosively,which brings troubles for the public security monitoring,criminal investigation and the detection of the case.Especially with the increase in vehicles,cases which involve cars are more and more.The detection and recognition of cars and its character features draw a lot of attention.However,the performance and accuracy of the detection and recognition system shoud be improved,so it is of great significance to propose a highly efficient and accurate vehicle detection algorithm.In this paper,the algorithm of extracting the vehicle and its region of interest(ROI)by using convolution neural network is studied,including accuracy and processing performance.The primary research contents are as follows:(1)The vehicle and its ROI image data set were built up with the help of vehicle training data set of the city's Public Security Bureau,which includes four types of vehicles,for instance,passenger cars,minibuses,buses,trucks and 4 Classes(regions of interest),such as windshields,sun visors,license plates,and year-of-the-art,the 8samples of the categories are used as training and testing samples to verify the effectiveness of the algorithm.(2)This paper presents a coarse-grained detection method for vehicles and their ROI based on improved convolution neural network.The network structure uses the ELU activation function and can realize the multi-scale training of the image.The structure quickly detects and identifies the category of the 8 kinds of data,and decides which kind it belongs to.Compared with the target detection methods proposed in recent years,the results show that the processing rate and accuracy are improved obviously,and the average accuracy rate is 91.2% in the case of 20000 iterations.(3)The advantages of CUDA parallel architecture and stand-alone multi-GPU card are used to further improve the detection speed of vehicle and ROI detectionalgorithm.The same traffic data has achieved nearly four times of the acceleration effect and speed up the efficiency of target retrieval.In this paper,based on the study of the parallel computing model of traffic image models for the heterogeneous platform of the National Natural Science Foundation of China,and the parallel cooperative retrieval system of the vehicle image based on the depth of the international cooperation project of Hebei Academy of Sciences.The research of vehicle target and its ROI detection uses convolution neural network algorithm,which achieves parallel detection of one machine with multi-GPU and improves the accuracy and processing performance.Through these studies,we can provide technical support for the rapid excavation of suspected vehicles,the service of police operations and the improvement of the level of public safety services.