Design Of Vision-based Frontal Vehicle Detection And Tracking System

Recent years, with the development of intelligent transport systems(ITS) and intelligent vehicles(IV), vision-based preceding vehicle detection and tracking technology as the fundamental module for both ITS and IV becomes an attractive field. Preceding vehicle detection and tracking system has many practical applications as diverse as monitoring, intelligent scheduling and autonomous navigation.This thesis presents a designing of vision-based preceding vehicle detection and tracking system. The system contains three main components: lane detection, vehicle detection, and vehicle tracking. The main work of this thesis includes points shown as follows.Firstly, a frame of lane detection system is proposed. In this module, we calculate the mean value of the image, and take projection in vertical direction, combine its first order differential information to capture the air-line. The OTSU algorithm with bilateral threshold is then proposed to take binaryzation for the image to determine the accurate area of lane and its line; After those steps finished, we use improved Furiman Chain Yards combined with Hough transformation to detect the lane line.Secondly, we design a Hog features based on feature template, which can well describes the shape of vehicles. On this basis,we present a double cascade structure of vehicle detection algorithm, this algorithm follow two basic steps: Hypothesis Generation(HG); and Hypothesis Verification(HV). First of all using cascade Adaboost algorithm hypothesizes the locations of the vehicles in images; then using cascade deformable part models algorithm with higher accuracy to verify the locations, builds a real-time vehicle detection framework with high accuracy.Thirdly, we design a tracking algorithm based on the theory of compressed sensing. The algorithm construct image multi-scale spatial of the target, and then get the feature set by using Difference of Gaussian; Then design a feature extraction method based on compressed sensing, modeling for each feature set, using the naive Bayesian classifier to detect target in the next frame.Finally, integrate the algorithms and implemented the frontal vehicle detection and tracking system. Experimental results demonstrate that the proposed algorithm is effective for vehicle detection and satisfies the real-time requirement.