Machine Vision Based Front Vehicle Detection And Tracking

In the wake of the rapid growth of social economy,the intelligent transportation system has become a hot issue of great concern in today's society.As an important part of the system,intelligent vehicle,which has integrated functions such as environmental perception,planning and decision-making,and driving assistance,is of great theoretical research value,and has broad prospects for application.This technology has been applied to real life more and more.In this paper,a vehicle recognition and tracking technology is studied based on monocular vision sensor.Taking the driver-less vehicle project of SWJTU as the background,this research is aiming at the environmental perception problem in intelligent vehicle technology.The main contents of the following research include: vehicle feature selection,machine learning algorithm research,front vehicle identification and tracking.In this paper,the characteristics commonly used in vehicle detection are studied,and the features of Haar-like,HOG,shadow and symmetry are introduced in detail.To compensate for the limitations of single feature,this paper proposes a method combining the symmetry feature and rectangle feature to detect the front vehicle.First,the edge detection Sobel operator is used to marginalization of the image,and the region of interest is extracted with the feature of vehicle symmetry.The extraction principle of the vehicle rectangle feature is introduced.The extended rectangle feature and the integral image method are used to calculate the features quickly.A preliminary study is carried out for the machine learning algorithms commonly used in vehicle detection.The adaptive enhancement algorithm and support vector machine are introduced.The AdaBoost algorithm and the extended Haar-like rectangle feature are used to verify the region of interest.Image preprocessing is studied.Average value method and median filtering are used to process the original image.Using the characteristics of vehicle symmetry to extract the candidate area,the front vehicle recognition algorithm is designed with the rectangular feature and the trainer generated by the AdaBoost algorithm,combined with the multi scale transformation method and the reward and punishment window fusion scheme.The experimental results are analyzed,and the recognition results under different samples and different environments are compared.The results prove that the algorithm is more reliable and accurate than the common AdaBoost algorithm.After studying the commonly used vehicle tracking algorithms,Calman filter is applied to track vehicle targets,and a moving target tracking experiment based on Calman filtering is carried out,and some achievements have been obtained.Finally,the paper summarizes the work of this research,puts forward the shortcomings,and looks forward to the future of the research content.