Vision-based multi-scaled vehicle detection and distance relevant mix tracking for driver assistance system

This paper aims to improve the robustness of vision-based multi-scaled vehicle detection and tracking for an actual driver assistance system. Considering the problem of discontinuity of detection and tracking for multi-scaled vehicles especially in an ultra-close area, we propose a novel detection framework which concludes short-range local feature (license plate) detection and long-range skeleton detection. Specially, the rear license plate can be located accurately by introducing a multi-scaled morphological operator and analyzing the color information. Then, vehicles in a long supervising range can be detected with a Look-up Table-based AdaBoost classifier synchronically. Finally, an inverse perspective mapping-based tracking strategy is proposed to unite the location results in the framework. It is proved to make up the leak vehicle detection in the near supervising area and improve the robustness of tracking. The accuracy of license-based detection and the robust mix tracking have both been testified in several groups of experiments.