Research On Visual Vehicle Detection Algorithm Based On Improved Faster R-CNN

Vehicle detection algorithm is an important link of vehicle aided driving and automatic driving,which plays an important role in the related technology of intelligent network connection.Detection algorithms with pyramid structure often have good detection accuracy and detection efficiency.This kind of algorithm has been widely applied to various kinds of automatic driving systems.At present,the problems of vehicle detection algorithms mainly come from the following three aspects: 1.The algorithm has poor detection effect for distant vehicles;2.The detection effect of the blocked vehicle target is not good;3.The detected target positioning information is not accurate enough,which affects the conversion accuracy of image coordinates to world coordinates.It can be seen the current vehicle detection algorithm still has a large space for progress.In this paper,vehicle detection algorithm based on pyramid structure in deep learning image detection is deeply studied.The specific research work is as follows:(1)On the basis of deep analysis of the pyramid structure principle,this paper proposed the vehicle detection algorithm based on the double-channel residual pyramid.The basic idea is,first of all,from the characteristics of the extracted from the images of different resolution figure to save,respectively,as an original features,and then the characteristics of the resolution of these different figure first do a top-down feature fusion,the high-level features with rich semantic information gives the underlying characteristics,then put the original characteristics of preserved before finish with each layer of different resolution characteristic figure to do together,realize the thought of the residual,further enrich the layer characteristic information,finally will sum up the layers of different resolution fusion features figure from bottom to do it again,can carry on the position information of the supplement to the senior figure,No longer confined to one-way traffic.The vehicle detection algorithm based on the residual double-channel pyramid structure can effectively reduce the detection probability of the detection algorithm for the smaller vehicle targets with a longer distance.The experimental results show that the proposed algorithm is better than the existing algorithm and can meet the practical requirements of the application of safe driving assistance technology.(2)Aiming at the problem of inaccurate positioning information obtained after vehicle detection and small improvement of detection accuracy of larger vehicle targets.On the basis of a double-channel residual pyramid structure,the paper adds a module of enhanced global roi module,further to extract the global features to add with original local features,at the same time,we fuse the original local features with the expanded areas of roi makes network can locate the target more accurately.Experimental results show that this method can effectively improve the accuracy of big targets of the algorithm in different scene data sets.(3)Aiming at the problem that the detection model is not good for the detection of occluded targets.On the basis of the double-channel residual pyramid structure,this paper improved the loss function of the original vehicle detection algorithm,joined the attractive loss and balance loss module,when the occluded vehicle targets exist in the training sample,this loss will pull away occluded target as much as possible from the surrounding test box and take the method on the occluded data set for testing.Experimental results show that the method can effectively improve the accuracy of the algorithm for the occluded objects in the data set.This paper trains a vehicle detection model based on double-channel residual pyramid structure and global context roi module,designed the visual interface and validates the improved algorithm on the unmanned vehicle platform.The source code is based on Python3.6 edition deep learning framework Pytorch1.0,the operating system is Ubuntu18.04 64 bit operating system,and the development tools are Pycharm and Vim,in order to improve the running speed of the model during detection and inference,this study further converts Pytorch model into TensorRT model for deployment,which can effectively reduce the model's inference time and meet the requirements of real-time vehicle detection.