Driverless Road Based On Deep Learning Research On Vehicle Detection And Classification

The emergence of driverless vehicles,which can play a positive role in reducing traffic congestion,reducing traffic accidents and solving exhaust emissions,is one of the hot spots in the current automotive industry and research workers.Driverless vehicle requires a strong ability of environmental awareness,and the results of environmental awareness will directly affect the driverless vehicle's decision-making system to make what kind of driving behavior response.Therefore,the research on the detection and classification technology of driverless vehicle on the road is of great practical significance.The results have been studied as follows:(1)Research on vehicle image preprocessing algorithm and data-set establishment.Aiming at the characteristics,illumination and occlusion of the image of driverless road vehicles,this paper preprocesses the data set image,proposes an improved image preprocessing algorithm,and evaluates the image quality before and after processing.Firstly,the improved image preprocessing algorithm is used to denoise and enhance the image;secondly,the mask algorithm is used to block some images in the training data set,so that the sensitivity of the network model to the occluded image is improved in the training process,so that the detection ability of the network model to the occluded vehicle in the image is improved.Building image data sets,including 10 different vehicle types,different types of vehicle body size and speed will be different,network model detects the vehicle types in the image,combined with its size in the image,you can judge the distance between vehicles,speed and other information,so that the information passed to the decision-making system is more accurate,the decision system makes different driving behavior decisions according to different types of vehicles around.At the same time,in the training data set,the images are added with noise,rotation,scaling and the images under different road conditions and weather conditions.(2)Road vehicle detection and classification based on improved Faster R-CNN.In view of the high requirements of driverless on road vehicle detection accuracy,Faster R-CNN is selected as the detection model and improved.First of all,the training parameters of the model are optimized to make the training process more stable and faster convergence and avoid over fitting.Then,optimize the network model structure:the original Faster R-CNN model has poor detection effect for small objects in the image.Therefore,the pyramid feature network FPN is used to fuse different levels of image features.The experimental results show that the improved network can improve the detection accuracy of small objects in the image to 86.1%;the original network model has many parameters and poor speed performance,so the combined pruning algorithm is used to compress the network model.After pruning,the accuracy of the model test is 85%,the training time is reduced to half of the original,and the model parameters are reduced by 45%(3)Network model detection and classification performance evaluation and experimental results analysis.The improved Faster R-CNN model is trained and the performance of network detection and classification is tested by experiments.The experimental results show that the improved Faster R-CNN keeps the high accuracy of the original network model in detection and classification tasks,and also has a significant improvement in speed.Each detection and classification evaluation index is increased by 3%-4%,which verifies the effectiveness of the improved network model.