Research On Vehicle Detection And Depth Estimation In Traffic Scene Image

Intelligent driving in the transportation field is becoming more and more important.It is of great significance to reduce traffic accidents and prevent loss of life and property.However,the conventional intelligent vehicle driving system has the problems of slow efficiency and low accuracy.With the implementation of deep learning,Forward Collision Warning System(FCWS)can promptly issue a warning when the vehicle in front is too close,thereby avoiding accidents.The main purpose of FCWS is to detect the distance of the vehicle in front in real time,which is supported by two important technologies: vehicle detection and depth estimation.Vehicle detection is used to identify and monitor the vehicle in front,and depth estimation is responsible for calculating the distance to the vehicle in front.These two technologies jointly ensure driving safety.Therefore,this thesis focuses on the two key technologies of vehicle detection and depth estimation in the Forward Collision Warning System.The main contents are as follows:(1)The vehicle detection method is studied,in order to solve the problem of time-consuming and low accuracy of traditional methods in feature extraction,an improved YOLOv3 vehicle detection model is proposed.In order to solve the problem of simultaneously requiring high real-time performance and accuracy in traffic scenarios,the shortcomings of the original model's architecture are analyzed.In order to further improve the real-time nature of the network,the architecture of the network model has been simplified.Experiments show that the improved YOLOv3 model improves the real-time performance while maintaining a high accuracy rate.Secondly,there are often problems in the quality of image acquisition in the actual traffic,which may be caused by the dim light,leading to blurred vehicles in front,reducing the accuracy of the model.In order to solve this problem,the original preprocessing algorithm is improved by the threshold segmentation method.The experimental results show that the improved preprocessing algorithm improves the accuracy of the model.(2)In order to further obtain the distance of the detected vehicle,the depth estimation method is researched,in order to solve the problem that it is difficult to effectively use the underlying features in the deep convolution network,a CNN depth estimation method is proposed which fuses multi-level features.First,the problems of CNN in processing depth estimation tasks are analyzed.Because in CNN networks,the final effect of depth estimation often only uses the high-level feature information of the network,it is difficult to use the information of the underlying features,which results in the effect of the depth estimation is relatively fuzzy.In order to solve this problem,a method of fusing multi-level features is adopted,so that the network can use the bottom and high-level information for depth estimation at the same time.Experiments on the dataset prove that this method effectively improves the accuracy of depth estimation.(3)In order to further improve the accuracy of depth estimation,on the basis of point(2),combined with the game theory in GAN,the method of stacking CycleGAN is proposed.Through two-stage training,from rough and fine scale training and stacking model,the end-to-end RGB image to depth image conversion is realized.In order to make full use of the data set,L1 loss function is introduced to make the network learn the supervised information.Experiments show that the improved network further improves the accuracy of depth estimation.