Research on Simultaneous Tracking of Moving Targets by Multiple Cameras 0277045CA 00262113

本研究的目的是建立一种平行跟踪视觉技术，能够同时进行详细的、高分辨率的跟踪。通过使用两个或多个分辨率高但视野窄的电视摄像机，可以同时跟踪两个或多个独立运动的目标。为此，笔者从以下几个方面进行了具体的研究，取得了如下成果：同时检测和跟踪两个或多个运动目标首先，作者研究了一种检测和跟踪两个或多个运动目标的技术的发展。具体地，使用帧之间的差异、掩蔽处理、光流提取和提取的流的统计分类来提取视场中的运动元素并对其进行分类。通过对运动目标的提取和分类，实现了对运动目标的检测和跟踪。实现光流提取…的健壮性通过不正确的对应控制获得更多n上述光流提取使用基于相关性的模板匹配。为了实现光流提取的稳健性，作者解决了匹配过程中错误对应的控制问题。具体地说，介绍了匹配的成败检查。它考虑了局部区域的相关值分布情况，特别是峰的存在、峰的尖锐程度等。因此，可以控制和避免匹配误差和伪流提取。此外，通过对时间轴顺序的两个方向进行相互匹配，成功地控制了除对象边界的伪对应之外的伪流的提取。它在本质上是不可能与之匹敌的。两个或多个摄像机协作进行大范围跟踪将开发的单摄像机视场中运动元素的提取和分类技术应用于两个或多个摄像机，以实现大范围跟踪。这是通过在时间和空间上给出模板匹配来实现的。每个摄像头的视觉处理是如此独立。每个摄像头的并行度都很高，这是可以预期的。因此，技术的并行化对其性能的改善较小

The purpose of this research is to establish a paraliel tracking vision technique which can continue carrying out tracking in detail and in high resolution simultaneously. By using two or more television cameras with high resolution but narrow view, two or more independently moving targets can be tracked simultaneously. For this purpose, the author have investigated the following topics as concrete research items and obtained several results described below1. Simultaneous detection and tracking of two or more moving targetsFirst, the author have tackled development of a techique to detect and track two or more moving targets. Specifically, difference between frames, masking processing, optical flow extraction, and statistical classification of the extracted flow are used to extract and classify motional elements in a view field. As a result of the extraction and classification, detection and tracking of moving targets have been succeeded2. Realizing robustness of optical flow extractio … More n by incorrect correspondence controlThe optical flow extraction mentioned above uses correlation based template matching. In order to realize robustness of the optical flow extraction, the author have tackled control of incorrect correspondence in the matching process. Concretely, success-or failure check of the matching is introduced. It considers situation of correlation value distribution in partial area, especially, existence of a peak, sharpness of the peak, and so forth. As a result, matching error and fake flow extraction can be controlled and avoided. Moreover, by matching both direction of time-axis order mutually, it succeeds in controlling extraction of fake flow except for fake correspondence of the object boundary. Its matching is impossible in essence3. Wide range tracking by cooperation of two or more camerasThe developed technique to extract and classify motional elements in the single camera view is applied to two or more cameras to realize wide range tracking. This is realized by giving template matching both temporally and spatially. Vision processing with each camera is so independent. And high parallelism for every camera can be expected. Thus parallelizing of the technique improve its performance Less