Moving Object Detection Research In Dynamic Scenes

Moving objects detection has been one of the most basic issues in Computer Vision, and in various social areas, it shows great prospect for application, such as intelligent surveillance system, robot industries, medical diagnosis, self-guided missile and so on. Because of its profound civil and military values, researching moving object detection received high emphasis in the developing process of Computer Vision in recent years. And the fact that detecting in dynamic scenes includes both problems related to static scenes and problems introduced by camera ego-motion, increases research difficulty of moving object detection in dynamic scenes, whose development still has a long way to run. Therefore, in this thesis, we divide dynamic scenes into two major categories and research the moving object algorithm in these two cases respectively.Detection in 2D scenes. We introduce a widely applicable global motion model, and solve the motion compensation problem through feature points matching. The algorithm has been tested on several real shot sequences and standard sequences, and results shows that our algorithm is able to accurately achieve motion compensation and detect motion in real time performance. Our original work can be summarized as:(1) Analyze characteristics of global motion in various types of scenes, and introduce eight-parameter model which is capable to model multiple scenes.(2) As to the shortcoming of SIFT-low matching rate, we propose a feature searching and matching strategy based on position estimation. This can improve the efficiency while not reducing performance of original SIFT.(3) Feature points update strategy based on residual image. This update strategy makes sure that features can be updated frame by frame and thus the running speed increases.Detection in 3D scenes. This algorithm employs multiple views geometry to detect dynamic points in each frame. Two different geometric constraints are used in the process. The algorithm is tested on long-time indoor sequences and results show that it can robustly and correctly detect moving objects in 3D scenes. Original work of this part can be summarized as:(1) Introduce multiple views geometry into motion detection. Combining epipolar geometry and three-view geometry, the features are assigned with dynamic probabilities. (2) Propose 3D position consistency constraint: in 3 consecutive frames, reconstructed points with changing 3D position has a high dynamic probability.