Vehicle Identification And Detection Based On Improved Mask R-CNN

With the rapid development of the automobile industry,traffic accident management is becoming more and more challenging for traffic management departments.In order to further improve the efficiency of dealing with traffic accidents,it is the main content of text research to identify and detect the accident vehicles in static images.The traditional vehicle image recognition algorithm divides the feature extraction and classification into two stages.The feature selection,mainly depending on human experience,needs to be constructed manually,and the whole process is inefficient.Especially in real complex traffic scenes,due to the variations of weather,illumination,environment and optical jitter,the target cannot be effectively detected.In view of the shortcomings of traditional vehicle detection algorithms,convolution neural network will become the mainstream of target detection algorithm.its main advantage is that the four steps of candidate region generation,feature extraction,classification,location refinement in target detection are unified into a depth network framework,which effectively improves the detection efficiency,and has certain invariance for target rotation and displacement,which can be applied to complex scenes.This paper studies the advanced convolution neural network knowledge,and applies it to the accident vehicle detection in static images.The main work is presented as follows:1)The structure and working principle of convolution neural network in target detection are studied,and the representative neural networks R-CNN,SSPNet,Fast R-CNN,Faster R-CNN,Mask R-CNN algorithm in the field of target detection are analysed respectively.Through the research and comparison of each network structure,the construction principle of each network can be intuitively understood,and then the advantages of mask R-CNN network structure in target detection can be more clearly understood.2)Combining with the vehicle detection problem in the accident scene,this paper improves the network structure and some details based on the original mask R-CNN network model,so as to better adapt to the practical application problems and realize the application value of the network.Because vehicle detection can be regarded as a two-class problem approximately,the redundant feature information generated in feature extraction is eliminated by reducing the number of network layers in the feature extraction network structure,and the operation speed of the algorithm is improved.The experimental results show that the algorithm has a better performance when the number of network layers is reduced by 8 layers.In order to prevent over-fitting,the threshold of dropout layer is set to 0.5 in the design of candidate window classifier,and the pool layer is added in the candidate window position.In the process of candidate frame generation,the non-maximum suppression algorithm is improved to learning network,which makes the model more flexible to select the maximum,get rid of the influence of human experience on the model,and improve the generalization ability of it.3)Combined with the research background,the model is trained by ImageNet and Pascal VOC vehicle images.In order to make the model more suitable for vehicle identification and detection problems in accident images,two kinds of data sets are established in this paper,one is accident vehicle data set,the other is aerial vehicle data set.Since that data set of an accident vehicle in the disclose data set is relatively small,in order to expand the data set,the method of data enhancement is adopted to further improve the generalization ability of the model.