Research On Object Detection And Tracking Method In Active Vision System

With the evolution of human vision research, the field of visual system is developing from static to active vision. Active vision system simulates the initiative of human vision, and makes decision according to mission requirements, i.e., it is task-driven. Due to its high flexibility and a greater field of view compared with the stationary camera, active vision has actual using value in many areas, such as video surveillance, human-computer interaction and so on.1. On moving object detection, a Mixture-of-Gaussian based background modeling for target detection method is proposed. Mixture of Gaussian model used to establish the background which can adaptive update. It is a good solution for the issue of light changing and background interference in a dynamic environment and improves the detection accuracy. In the process of image subtraction and binarization, the OSTU algorithm used for adaptive binary-threshold selection, and able to adapt to various light intensity for moving object segmentation. Then, morphological operations used to remove the bright spots in binary image. Finally, the geometric characteristics of target used to remove the interference of the larger background objects. Experiments demonstrate that the method is effective for moving object segmentation in a low computational complexity.2. On moving object tracking, a new Mean-shift tracking algorithm of adaptive kernel bandwidth is developed. To meet the requirements of real-time in active vision system, Mean-shift tracking algorithm is adopted. For solving the issue of locating failure caused by target size changing, SURF feature points algorithm is introduced for detecting the feature points in the two adjacent frames. A scale factor is gained according to these points pairs in the two frames, which is used for adjusting the kernel bandwidth of Mean-shift. At the same time, in order to enhance the stability of tracking, linear prediction used to predict the possible location of track window in current frame, according to the locations in the two previous frames. The results show that the algorithm is stable and keeps a low computational complexity for real-time tracking.3. Finally, on the basis of analyzing object detection and tracking algorithms, a prototype of active vision based system is implemented. Technologies such as component, multi-threading programming are used for modification and extends easily. Experiments show that the system can achieve detection and tracking tasks, and satisfies real-time application requirement.