Real-time Vehicle Detection And Tracking Based On Road Area Analysis

Real-time Object detection and tracking from video captured by in-vehicle camera is an important research area in intelligent transportation systems and an important method to reduce the traffic accidents. It has promising prospect in driving assistance and accident warning field.In this thesis, upon the background of driving assistance on highway, we present a new framework for object detection and tracking on the basis of road area analysis. Based on this framework, we proposed new object detection and tracking algorithms. The main content of this thesis can be listed as follows:1,Based on the character of the road image and the result of road boundary detection, we proposed a new framework for object detection and tracking. In the process of lane boundary detection based on Hough transform, we improve the Sobel operator to do edge detection and combine the result with the output of the pixel classification, which is done by multi-Gaussian method. This combination enhances the robustness of the algorithm to noises.2,We proposed a new object diction algorithm via fusion of global classifier and part-based classifier. The global classifier is constructed as an Adaboost cascade; and in building the part-based classifier, we firstly invite SIFT-like feature to describe the vehicle parts, and then combine Adaboost and BP neural network to select the most effective parts and fuse the part-based classifiers. The method by fusion of global classifier and part-based classifier could not only detect object fast from the image, by also extract more effective local features from the object, so it could enhance the discriminate ability of the classifier.3,We implement a fast vehicle tracking algorithm based on Kalman filter. In the measuring process of Kalman filter, we proposed a new measuring method integrating the Adaboost confidence map and local shadows of the vehicle. This integration ensures the adaptability of the tracking algorithm. In doing shadow detection, we improve the Maximally Stable Extremal Regions method and get promising result.4,Grounded on these algorithms, we design and implement a real-time side/rear view vehicle detection and tracking system. It has been successfully tested under several environments.