Research of A General Surveillance Algorithm by Images

提出了一种利用图像信息的行人检测、识别和跟踪方法。然而，由于该算法不假设目标的特定形状，因此该方法适用于除人之外的其他目标。该方法包括三个阶段。第一个是传入对象检测，它通过逐像素操作检测进入对象。第二阶段是行人识别，识别单个行人，第三阶段是行人跟踪，跟踪已识别行人的位置。对于目标检测，我们采用了双摄像机的方法，该方法对光照变化和阴影不敏感。本文的一个新颖点是开发一种方法来确定准确的几何变换和校准相机的灵敏度，这是需要一个正确的操作的传入的物体检测过程。另一个创新点是使用一个类似于维纳滤波器的过程，我们称之为正则化逆滤波器，用于行人识别。这改善了附近行人的分离。室外场景实验表明，该方法对光照变化和阴影问题具有稳定的检测和跟踪效果。成功追踪行人的比率为93%。

We propose a pedestrian detection, identification, and tracking method using image information. However, since, the algorithm does not assume a particular shape of the target, this method is applicable to other targets than persons. The method consists of three stages. The first one is incoming object detection, which detects entering objects by pixel-wise operations. The second stage is pedestrian identification, which identifies individual pedestrians, and the third stage is pedestrian tracking, which tracks the position of the identified pedestrians. For the incoming object detection, we adopt a method using two cameras, which is insensitive to illumination change and shadows. One novel point of this paper is to develop a method for determining accurate geometrical transformation and calibrating camera sensitivities, which are needed for a proper operation of the incoming object detection process. The other novel point is to use a process similar to the Wiener filter, which we named regularized inverse filter, for the pedestrian identification. This improves separation of close-by pedestrians. Experiments using outdoor scenes proved that the method provided a stable detection and tracing result against the illumination change and the shadow problems. The ratio of successfully traced pedestrians was 93%.