Improve YOLOv3's Traffic Vehicle Detection

With the development of economy and technology,the number of urban vehicles keeps rising,and the intelligent transportation system is gradually applied to People's Daily life.In modern urban traffic,video surveillance covers most driving areas,and the vehicle detection method based on computer vision and deep learning can be widely used in the scene of traffic video image.However,the accuracy of the existing object detection methods in the application of real scenarios still cannot meet the actual needs.Therefore,how to achieve accurate vehicle detection in complex environments is of great research significance for intelligent transportation.In this paper,the vehicle detection problem in urban traffic scenarios is studied,an improved vehicle detection method based on improved YOLOv3 was proposed,and the network model was optimized with the data augmentation method for training.In this paper,on the research of data augmentation method,a method of applying mix-up data augmentation and CutMix data augmentation to object detection dataset was proposed and applied to KITTI vehicle dataset.Through experiments,it was concluded that the optimal method of data augmentation for KITTI dataset was CutMix,which can be used as the training data set constructed in this paper.In this paper,a vehicle detection method based on YOLOv3 is proposed.The improvement was divided into two points.First,the original darknet-53 backbone network structure of YOLOv3 is optimized by drawing on the idea of Deformable Convolutional Networks(DCN),and the Deformable convolution networks was added into the residual module.Secondly,the multi-scale fusion part was improved,the feature fusion part was promoted to 4 scales,and the up-sampling was carried out by transposed convolution,which increases the detection ability of the model to small targets.Through comparing follow-up study found that after using the CutMix KITTI datasets in this paper,the improved YOLOv3 trained for vehicle detection,got 92.37% of the mAP,proved that in vehicle detection,the improvement of the model and the data augmentation method of the training dataset have improved the performance of vehicle detection,completed a precise vehicle detection.