Research On Feature Detection Based Vehicle Tracking Technology

As core technologies of Intelligent Transportation System(ITS), vehicle detection and tracking are pushed forward continually along with its good prospects for development and upgrading application demands. Traditional detection methods make video image processing based vehicle detection popularized, but their inherent limitations become bottleneck of system performance improvement. In recent years, more researchers focus on detecting vehicles based on some distinct visual features on their bodywork. This paper takes design and realization for feature extraction based vehicle tracking method as subject to do further research. In good lighting condition during daytime, the shape information of vehicle is outstanding and sufficient in traffic images, so that efficient object detection can be realized by employing corners as feature. For tracking target, the best estimation to large magnitude motion of objects at present can be achieved by feature point-based multi-resolution optical flow algorithm. To make performance of detection system perfect, vehicle detection in the evening is carried out by means of edge extraction, as edge information is legible and abundant relatively. At night, its realization relies on headlights recognition, because they are the only distinct and stable feature in that scene. Experiments prove that a real-time processing speed and high detection precision are available at toll-gate in different lighting conditions, with methods mentioned above. Moreover, the system is robust to image rotation, shadow and noise.Local Binary Pattern (LBP) is famous as an effective texture descriptor, but no previous work has focused on utilizing it for feature description to vehicles in ITS domain. The main contribution of this paper is that LBP texture descriptor is brought forward to extract feature for vehicles in traffic video images for the first time, and a novel Niblack-LBP (N-LBP) algorithm based vehicle tracking method is developed. The new algorithm is adaptive to various illumination changes, thus, it can make up fatal limitations of previous one, namely that it is sensitive to sudden illumination changes in local areas and its detection methods in day time and at night are repugnant. The approach put forward in this paper includes three steps:firstly the general critical ingredients (GCI for short) that can indicate vehicles effectively are selected from N-LBP features, through analyzing changing rules of distribution histograms of N-LBP features acquired from region of interest(ROI) in traffic video images. Then GCI are extracted from ROI in the new image and a typical part of them, which satisfy some quantitative standard, are selected for object detection and identification. Linear Kalman filter is employed for feature based movement estimation and affirmance to labeled objects. Finally, capture is carried out according to statistical standard to improve correct rate of the system. Experimental results demonstrate the superiority of N-LBP feature over basic LBP feature. Correct rate is increased further, while, false alarm rate is decreased effectively. Also, performance of the new system is more stable and reliable.