A Study On Omnidirectional Vision Based Target Detection And Tracking

With the rapid development of computer vision, omnidirectional sensor is widely utilized in some areas, such as video surveillance, virtual reality, video conference, robot navigation, etc, because it can thoroughly grasp the environment information via its wide field of vision. Based on ominidirectional vision technology, this thesis investigates the unwrapping of omnidirectional image and target detection and tracking technology.As known to all, it is difficult to obtain the center of image in the process of omnidirectional image unwrapping. In order to overcome this problem, a quick omnidirectional image unwrapping algorithm based on Hough transformation is proposed. In this algorithm, the center and the radius of the scene are firstly detected through Hough transformation, and then the image size can be determined by the radius of the scene circle, finaly the image is unwrapped based on the center of the circle. Experiment results show us that the proposed algorithm can unwrap omnidirectional images quickly and effectively, and reduce distortion of the unwrapped images. During the process of unwrapping of omnidirectional video sequence, frame dropping mechanism is proposed in this thesis. By comparing the time sequence of omnidirectional images and the interval of inter-frame instantly, one can appropriately deal with the case of frame dropping. Moreover, experiment results are also listed to show the high real-time property of the proposed mechanism.In order to meet the needs of high real-time property of omnidirectional image and target detection,a target detection algorithm, which is based on an adaptive mixture Gaussian model and uses the omnidirectional camera for detection, is also presented. Firstly, the color space of the unwrapped omnidirectional video sequence can be transformed from RGB to YUV. Secondly, the luminance component can be extracted, and according to such luminance component, one can adjust the sampling frequency, i.e., frame filtering. Finally, the video sequence after the frame filtering can be modeled through mixture Gaussian model. Furthermore, experiment results are given to show the effectiveness of detecting moving targets and the high real-time property of the proposed method.On the basis of targets detecting, a background subtraction and Camshift based algorithm is stated. Firstly, the feature information, the color and location information of moving objects, is gained by background subtraction, and Camshif algorithm is used to detect targets via color information. Subsequently, by background subtraction, the position fusion between the targets'position detected and the targets'position tracked by Camshift algorithm, which can be regarded as the real position of moving targets, can be realized. Therefore, the detecting and tracking of moving targets can be achieved. Finally, experiment results demonstrate the effectiveness of detecting and tracking of moving targets, and illustrate that it can also be quite efficient for circumstances of multiple moving targets.We sum up this thesis with an established online target detecting and tracking system based on the omnidirectional vision. Using this system, one can let online video acquisition, image unwrapping, targets detecting and tracking come true, which can show that this system has a certain practicability.