Traffic Information Acquisition based on Robust Tracking Techniques Using Images and Sound

基于从交通视频中提取各种有助于实现智能交通系统（ITS）的信息的思想，人们开展了大量的研究工作来开发利用交通监控序列的车辆跟踪技术。然而，由于鲁棒性问题，跟踪技术仍然难以实际应用。本研究的目的是通过大大提高车辆跟踪技术的鲁棒性，扩大其应用范围。在科研资助项目研究期间，我们首先建立了一种低阶跟踪过程与高阶跟踪过程相结合的跟踪方法，在贝叶斯框架下实现了目标物体在高维状态空间的随时间变换跟踪，以适应光照条件快速变化和物体运动的影响。采用该方法，实现了跟踪的鲁棒性和实时性的比较好结合。其次，在上述框架下，重点研究和实现了具有低层次跟踪功能的识别机制。这种低级跟踪过程既可以作为一个跟踪器单独工作，也可以与高级跟踪过程协同工作，使整个系统更加可靠。此外，(1)为了在恶劣天气、夜间或遮挡（由于车辆在图像上重叠）的情况下保持系统的鲁棒性，我们开发了一种基于识别立体声麦克风拍摄的道路交通声音的车辆通道检测方法，并证明了该方法特别适用于夜间图像序列；(2)为了在没有道路基础设施的情况下获得必要的信息，我们研究开发了一种利用车载摄像头检测和跟踪相邻车辆的方法；(3)为了支持安全驾驶，我们开发了在十字路口识别和区分车辆和行人的技术。与本研究项目有关的结果已在若干国际会议和期刊上发表。少

Based on the idea that a variety of information useful for realization of Intelligent Transportation Systems (ITS) can be extracted from traffic movies, a lot of research efforts have been deveted to develop vehicle tracking techniques using traffic surveillance sequences. However, because of issues in robustness, the tracking techniques are still difficult to be put to practical use. The purpose of this research is to expand the application scope of the vehicle tracking techniques by greatly enhancing the robustness of them.In the period of the research supported by Grant-In-Aid for Scientific Research, first we have established a tracking method that integrates low-level tracking process with the high-level tracking process, which enables transformation of target objects to be tracked over time in high-dimensional state space, in Bayes framework to adapt to the influence from rapid changes of lighting conditions and object's motions. By using this method, we have achieved the compati … More bleness between robustness and real-time execution of tracing.Secondly, within above-mentioned framework, recognition mechanism that functions as low-level tracking has been mainly investigated and implemented. This low-level tracking process can not only work individually as a tracker but can also work jointly with high-level tracking process to make the whole system more reliable. Additionally, (1) to keep the robustness of the system even in case of bad weather, night or occurrence of occlusion (because of overlap of vehicle on images), we have developed a method to detect the passages of vehicles based on recognizing road traffic sound taken by a stereo microphones and proved its effectiveness especially for night-image-sequences ; (2) to obtain necessary information in the situation where road-side-infrastructures are not available, we have investigated and developed a method to detect and track neighboring vehicles by using in-vehicle cameras ; (3) to support safe driving, we have developed the techniques to identify and distinguish vehicles and pedestrian at intersections.The results related to this research project have been published in several proceedings of international canferences and journals. Less