Video Object Detection And Segmentation Based On Visual Cognition Theory

Moving object detection and object segmentation in complex scene are difficult issues and have important application value in intelligent surveillance, intelligent robot, object based image and video analysis fields. Automatic object detection and segmentation are challenging problem for computer, but it is easy for human brain since human vision system has excellent vision cognition ability, which can accomplish some difficult vision tasks such as detecting, tracking and recognizing objects. Therefore, studying and making use of biological cognition mechanism of human vision system for video analysis is widely concerned. In this thesis, the research is focus on spatio-temporal saliency calculating model conforming to human being vision attention in complex dynamic scene, robust moving object detection and video object segmentation. The innovation and improvement work include:Firstly, a biologically-inspired approach for detecting moving objects in complex outdoor scenes is proposed. In neuroscience, scientists find that visual motions are perceived in the medial superior temporal (MST) area of the human brain, and motion field from visual perception is separated into independent components at the MST area. Inspired by this research, a novel approach combining independent component analysis (ICA) with principal component analysis (PCA) is proposed in this thesis for moving objects detection in complex scenes. In the proposed approach, taking advantage of the capability of separating the statistically-independent sources from signals, the ICA algorithm is employed to analyze the optical flows of consecutive visual image frames. As a result, the optical flows of background and foreground can be approximately separated. Since there are still many disturbances in the foreground optical flows caused by the complex scene, PCA is then applied to the optical flows of foreground components so that major optical flows corresponding to multiple-moving objects can be enhanced effectively and the motions resulted from the changing background are relatively suppressed at the same time. Moving objects can be detected finally. Comparative experimental results with existing popular motion detection methods demonstrate that the proposed biologically inspired approach can remove redundant noises in the background and detect moving objects effectively in a complex scene.Secondly, a new cognitively-inspired visual attention computational model is proposed for spatio-temporal saliency maps detection in complex dynamic scenes. Firstly, simulating the motion cognition mechanism in human brain, we employ robust ICA to detect salient foreground optical flows; second, inspired by the research that human eyes always focus on foreground objects, after clustering the foreground optical flows with similar direction and intensity by Meanshift algorithm, the foreground objects are then generated by a novel strategy of relative salient motion regions selection, thus we obtain temporal saliency map by calculating optical flow intensity of the foreground object areas. Meanwhile, we adopt the classical graph-based visual saliency(GBVS) model to get spatial saliency. At last, the temporal saliency map is further normalized and fused with the spatial saliency map to generate the final attention map. Extensive experiments and evaluation results demonstrate that our cognitively-inspired attention modelling can obtain better salient region locations at human eye gaze positions in complex scenes.Finally, a spatio-temporal saliency based approach for fast automatic video object segmentation is proposed. We perform fast optical flow to get motion information, and then calculate accurate motion saliency based on this temporal information, detecting the presence of global motion and adjusting the initial optical flow results accordingly. This is then fused with a region contrast-based (RC) image saliency method. Our joint weighted saliency map is then used as part of a foreground-background labelling by Markov Random field algorithm to produce the final segmented moving and static video objects. Experimental results in a wide range of environments are presented, showing that our method is faster, more robust and effective compared to the state-of-the-art approaches, especially for object segmentation in dynamic background.