The Research Of Vehicle Target Recognition Algorithm

With the continuous development of the social economy,a lot of families have their vehicles as an transportation.And other people will purchase different types of vehicles for different purposes in life.It has caused the vehicles to increase exponentially over several years.It also makes traffic congestion,traffic accidents and motor vehicle crime rates increasing.The difficulty of managing all types of parking lots has also grown.In order to solve these thorny problems,a method of detecting different types of vehicles in an original image using a feature-based vehicle target type detection method is used to effectively solve the problem of vehicle identification and vehicle type detection of large-scale images.In current feature-based target recognition algorithms,one of the methods to recognize with the fast speed and the better recognition effect is a convolutional neural network(CNN).One of the methods(regional convolutional neural network,R-CNN) is widely used in target classification and recognition.Firstly,it generates the most likely candidates for the target object in the graph,then uses the convolutional network to extract features,and finally trains the classifier to classify candidate regions.This thesis is based on Faster R-CNN,an algorithm with superior recognition rate and recognition accuracy in the R-CNN series algorithm.After analyzing the characteristics of convolutional neural network output characteristics of different convolution layers,an improved method for network structure is proposed by combining different convolutional layer output features.Then,the recognition accuracy of the Faster R-CNN algorithm is improved by using the context features of the original graph and the optimization strategy of the target bounding box.Finally,the improved network model is trained by the data set which consisting of vehicle data sets of MIT and Caltech plus vehicle images of three types(cars,SUVs and minibuses) on the network.Then the comparative experiment is conducted on the improved network model.The experiment verifies that the improved vehicle identification network can effectively identify cars,SUVs and minibuses in the image.And based on the Faster R-CNN algorithm,the recognition accuracy is improved.The mean Average Precision(AP,Average Precision) of the three types of vehicle was 83.2%,79.2%,and 78.4%.It is 1.7% higher on the AP than the traditional Faster R-CNN network.