Fast 3D symmetric pattern indexing for modeling, tracking and detection of semantic objects

Most existing approaches to detecting semantic objects in a video use an object detector to locate the target object in the first frame, which is followed by an object tracking to locate objects in successive frames. However, automatic modeling, tracking and detecting semantic-rigid objects remain a challenge since the object shapes are deformable across video frames. This paper presents an analytical method to detect the 3D symmetric pattern in a small video cube. A symmetric pattern indexing and retrieval scheme is also proposed to speed up the model-based tracking in order to detect semantic objects across frames. The Generalized Hough Transform (GHT) is used to add spatial constraints on these symmetric patterns to define the average shape of the target object in a small video slice. In the stage of model-based tracking, the rigid motion of the target object is estimated by rigid registration in both the color space and the Hough voting space. Experimental results demonstrate the good performance of the proposed method in terms of segmentation accuracy.