Research And Implementation Of Monocular Vision-Based Vehicle Detection Algorithm

With the development of economy, modern management methods are needed to administer traffic on account of problems of traffic transportation being more and more severe, which brings about the research on Intelligent Transportation System-ITS. Driver Assistance System is an important component of ITS, which has bright prospects, especially in automobile safety improvement and accident avoidance measurement. Since vision-based vehicle detection technology is an important component in driver assistance systems, it is of great significance to study vehicle detection technology.This paper uses monocular fish eye camera to take the external environment information, improves some algorithms on the basis of analysis and summarization of existing vehicle detection algorithm, and proposes and implements vehicle detection algorithm on fish eye image. The method this paper proposed follows two steps:hypothesis generation (HG) and hypothesis verification (HV). HG generates ROI (Regions of Interest) which include candidate vehicle; HV verifies the existence of vehicle on ROI. Before HG, through studying the basic principle of fisheye image rectification, we propose to apply plane-based model, cylinder-based model and sphere-based model rectification methods in our algorithm. After experiment results contrast among these three methods, we select cylinder-based model as the final rectification methods. In the case that vehicles in the rectified images may not still maintain some features(horizontal edge may be curve for example), in HG stage, we propose new shadow region fusion methods to improve existing vehicle segmentation and location based on multi-features fusion, and in HV stage, we propose new feature computation methods and use better weak classifier generation algorithm to improve existing AdaBoost classifier based on haar-like feature.Experiment results show that the method proposed in our paper can detect existing vehicles in rectified image, and the method verifies good robustness and applicability.