Research On Vehicle Target Recognition In ADAS Based On Monocular Vision

With the rapid development of China's economy and the continuous improvement of people's living standards,the number of cars is increasing day by day.With the frequent occurrence of traffic accidents,it is urgent to improve the active safety and intelligent performance of automobiles.Therefore,intelligent vehicle technology and intelligent safe driving assistance system(ADAS)have attracted much attention and become a hot topic for scholars and researchers.Vehicle environment perception system is the key of vehicle intelligent driving system or driverless technology.Vehicle recognition is the most important part of the environmental perception system.Accurate and rapid identification of the target vehicle is of great research significance for the intelligent control of the vehicle.Therefore,based on visual information and image processing technology,this paper studies the vehicle target recognition in front of the intelligent vehicle assisted driving system,which is mainly divided into three parts:(1)Selection of vehicle candidate areas based on taillight characteristics.By adopting various image processing methods,such as median filtering,sharpening,HSV color model transformation and morphological processing,the possible regions of taillights can be screened.Based on the inherent position distribution and area size of the taillights,the taillights are matched,so as to realize the final detection of the vehicle candidate area.(2)The discrimination of vehicle candidate regions by the SVM classifier based on HOG feature.By constructing the training set of positive and negative samples of the vehicle,and HOG feature extraction,the feature is sent to the SVM for training.The SVM detector is obtained by optimizing the training parameters,and the vehicle candidate regions detected based on the taillight feature are distinguished.Finally,the effectiveness of the proposed method is further verified based on the experimental results.(3)Vehicle identification method based on CNN deep learning.It mainly focuses on the basic principle analysis based on CNN deep learning,the network architecture design of fever-rcnn,as well as the training and testing of neural network.First,the CNN network module was built.After adjusting the convolutional layer,pooling layer and other structures and parameters,the deep convolutional features of the vehicle target were obtained.Then,the RPN network module was constructed,and multiple candidate regions were obtained through regression and classification loss calculation and post-processing.Finally,a fast-rcnn network was constructed,and multiple candidate regions were refined with the full connection layer,and finally the candidate box of detection target was obtained.Finally,the effectiveness of the proposed method is further verified based on the experimental results.