| Real-time people counting are the hottest research in the field of computer vision. And it has important practical application value. Now many methods of people counting are by tracking people or establishing human models which base on the color image captured by the monocular camera. These algorithms are mostly effective in a particular kind of environment.For the scenes of pedestrians goal blocked or congested they can’t count accurately. Consider the characteristics of the depth image and the features of head target in the three-dimensional space. This paper uses SVM classifier to sort targets.First of all, Basing on the imaging principle of binocular stereo vision, get the depth image of the scene by binocular camera calibration and stereo disparity matching. Then use the perspective projection geometry of the camera to restore a top plan projection of the depth image. In this way, the head features including geometric shape, size, height information, and the structure of the upper body are all restored. Next, extract the features of the locked foreground objects which include head objects and non-head objects, then analysis the distribution of the different types of objects features to determine the threshold of the head objects, so we can remove the non-head objects preliminary. At last, we match the targets to get the tracking trajectories which include the right trajectories formed by the head objects and the wrong trajectories by non-head objects. Then we get the optimal SVM classification model which is trained by the above marked object features to detection the objects on the trajectories. When a trajectory has more objects detected as head objects than the threshold,sort this trajectory as the right one, and judge it as a person.In this paper, the method has been tested by the flow of people statistics in two different scenes. The result shows that the method of detection persons basing SVM classifier trained by object features in 3D space is real-time. The accuracy is over 94%, it is adaptive to different scenes. |
