Research On Application Of Region Convolutional Neural Network In Vehicle Detection

Vehicle detection is the basic component of the intelligent transportation system.It is beneficial to improve the robustness of the whole intelligent transportation system by solving the problem of vehicle detection in complex environment.Traditional vehicle detection methods extract features based on artificial rules,it is easy to be affected by factors such as illumination,occlusion and deformation in the complex environment,the generalization ability is poor.The region convolutional neural network and its extended model have good performance in target detection task.Therefore,it is important to use region convolutional neural network for vehicle detection.The main research contents are as follows:1.The widely used methods of target detection based on deep learning were analyzed.The Faster R-CNN model based on region convolutional neural network and the application of the model in vehicle detection were especially studied.2.For the two problems of the application of Faster R-CNN model in vehicle detection:(1)In the image,the size of the target is different,the small size vehicles are seriously missed;(2)The space of negative sample is large which made the model discriminant ability is low.Faster R-CNN model was improved in this paper.First,the size of anchor box was modified,multi-layer feature fusion strategy was introduced in the feature extraction stage;Next,the multi-scale image was used to train the Faster R-CNN model;Finally,the hard negative samples were excavated,and the hard negative samples were added to the training set to train the model again.3.KITTI dataset and data sets collected in real environment were selected to verify the effectiveness of the improved Faster R-CNN model.In addition,The influence of different size of anchor box,the number of layers of feature fusion,scales of training image and hard negative sample mining strategy on the improved Faster R-CNN model was analyzed.The experimental results indicate that:(1)The vehicle detection method based on the improved Faster R-CNN model has higher accuracy and faster detection speed;(2)The reset size of anchor box,multi-layer feature fusion strategy and multi-scale training increase the detection ability of the model for small size vehicles;(3)The hard negative sample mining strategy can eliminate the problem of low model discriminant ability caused by the large space of negative sample;...