Automatic moving object detection using motion and color features and bi-modal Gaussian approximation
Victor Mejia, Eun-Young Kang · 2011
Automatic moving object detection is essential for various computer vision applications like video surveillance systems. Many previous detection methods work for usually low-res video sequences under certain constraints and are based on background learning and/or pixel-level motion analysis or they focus on detecting particular objects. We introduce a hybrid moving object detection scheme with motion-color features, followed by a statistical optimization step to increase the accuracy of the boundaries of the detected objects in hi-resolution video sequences taken in the presence of camera motions such as camera vibrations. Motion analysis involves both the motion vector information extracted from a reference H.264 decoder and a moving-edge map in order to produce an overestimate of the moving object blobs. Pyramid color segmentation connecting multiple components that might be under different motions is performed to extract the solid bodies of moving object blobs with accurate boundaries. Results from motion and color analysis are fused and a region-growing technique based on the blobs' Gaussian distribution of its RGB information is performed to further refine moving blobs. Results are shown to demonstrate the accuracy of our method.