Monocular Vehicle Detection For Forward Collision Warning System

With the development of road traffic and the increase of vehicle population, road traffic safety has become an acute problem, in particular the high incidence of rear-end collision rate. Forward collision warning systems based on machine vision, radar and other technologys to detect forward vehicles and warning when there will be an accident. Detection of forward vehicles is an important part of these systems. This article summarizes and analyzes the existing vehicle detection technology, focuses on vehicle detection during daytime and nighttime, vehicle tracking, and proses the improved algorithm. The main research work as follows:(1) The forward vehicles detection in the daytimeThe forwrd vehicles detection method is divided into two steps: hypothesis generation and hypothesis verification. In the hypothesis generation, we have explored various methods, eventually we chose shadow underneath to generate hypothesis. Histogram valley analysis method(HVAM) is proposed to get adaptive shadow threshold.This method can significantly solve the influence of illumination and the gray value mutation of road surface to shadow segmentation.In hypothesis verification, we propose V-HOG features instead of traditional HOG features. Taking one-dimensional gradient information mapping to two-dimensional, to enhance the distinction between background and foreground.(2) The forward vehicles detection at nightVehicles’ taillights information is effectively used to detecte vehicles at night. For existing vehicles’ taillights extraction, threshold was based on experience. In order to address this deficiency, we collect a lot of taillights images. Statisticing taillights image’s each color components in a variety of color spaces. Obtain the threshold by experiment. Then the taillights region is extracted. On this basis, taking into account the vehicle taillights region intermediate brightness is higher, and there was a red halo around it, combining high brightness region detection in the image, vehicle taillights extraction is further improved. Finally, according to the geometry of the vehicle taillight, establishing a priori constraints on the segmented region. Taillights are filtered and associated to locate the the vehicles’ position in the image.(3) Vehicle trackingDesign of a vehicle tracking method based on Kalman filter. The position of the vehicle and the external rectangular box are the tracking components; the motion state of the vehicle is predicted. For vehicles are susceptible to the effect of illumination, occlusion, deformation in the tracking process, we propose a vehicle observation method based on online learning. Establishing training dataset online, and training a classifier, the target vehicle’s motion state is measured. Most important, we present a strategy for multiple vehicles tracking by observing each target by turns and updating the tracking queue, which improves the real-time performance and ensures that the new vehicle can be detected.(4) Using the vehicle detection and tracking results, we extract the forward vehicles’ motion parameters, and their behavior is further analysised to provide effective theoretical basis for forward collision warning system.