Detection, Tracking And Localization Of Multiple Moving Objects Based On Omnidirectional Vision

With the rapid development of Internet, image processing and communication technology, intelligent video surveillance has become an emerging research focus in the computer vision fields, and it has been widely applied in all walks of life. This paper uses omnidirectional camera which can obtain 360-degree environmental information as vision equipment, detecting, tracking and localization of moving objects in indoor environment.Aiming at the problem that how to use the omnidirectional vision equipment to get surrounding information more accurate, this paper has studied the calibration technology of omnidirectional vision equipment. Firstly Analyzed the optical properties of hyperbolic mirror and the imaging principle of omnidirectional vision equipment, then introduced the perspective projection model of hyperbolical omnidirectional camera, lastly demonstrated calibration method of omnidirectional camera based on Taylor series. The calibration results are satisfactory, high precise, and it provide the basis for human localization based on the omnidirectional vision.To the static scene in the surveillance system, the background subtraction technique is used for detection. Firstly background model is established with statistics and updated in real-time mode in order to adapt the slow changes of the environment, and then obtain the foreground area by the threshold segmentation technique, finally, the morphology processing technique is used to filter the noise in the binary image. This method can extract the foreground area accurately from the omnidirectional image.A preliminary study has also been done in tracking moving objects. Kalman filter is mainly adopted to achieve the goal of tracking human body in real-time mode, combined with the information, such as center coordinate of moving object, external rectangle and so on, to improve the robustness of tracking objects, the experimental results show that it can accurately estimate parameters of the object in the next frame image, such as the position and velocity etc, which will offer much more facilities to the followed works.The binary image is processed by the labeling algorithm of connecting area, with this algorithm it can record the information of center point,enclosing rectangle and the feet point in each connecting area. According to the imaging characteristics of human body in omnidirectional image, the statistical method is adopted to extract the feet point of human. According to the imaging principle of omnidirectional vision equipment, the mapping relation between the image plane and the ground plane is obtained based on the extracted feet point, the human position on the ground is achieved. Combined with the detection and racking of moving objects, the real-time localization of human body in video surveillance system is achieved. This method is simple and has good precision it can distinguish the different human bodies in the scene.