Detection of deformable objects in a non-stationary scene
Physical copy available from RIT's Wallace Library at TA1634 .A92 2005
Image registration is the process of determining a mapping between points of interest on separate images to achieve a correspondence. This is a fundamental area of many problems in computer vision including object recognition and motion tracking. This research focuses on applying image registration to identify differences between frames in non-stationary video scenes for the purpose of motion tracking. The major stages for the image registration process include point detection, image correspondence, and an affine transformation. After applying image registration to spatially align the image frames and detect areas of motion segmentation is applied to segment the moving deformable objects in the non-stationary scenes. In this paper, specific techniques are reviewed to implement image registration. First, I will present other work related to image registration for feature point extraction, image correspondence, and spatial transformations. Then I will discuss deformable object recognition. This will be followed by a detailed description on the methods developed for this research and implementation. Included is a discussion on the Harris Corner Detection operator that allows the identification of key points on separate frames based on detecting areas in frames with strong contrasts in intensity values that can be identified as corners. These corners are the feature points that are comparable between frames. Then there will be an explanation on finding point correspondences between two separate video frames using ordinal and orientation measures. When a correspondence is made, the data acquired from the image correspondence calculations will be used to apply translation to align the video frames. With these methods, two frames of video can be properly aligned and then subtracted to detect deformable objects. Finally, areas of motions are segmented using histograms in the HSV color space. The algorithms are implemented using INTEL'S open computer vision library called OpenCV. The results demonstrate that this approach is successful at detecting deformable objects in non-stationary scenes.