Image parsing with automatic detection of symmetrical parts and its application on human activity recognition

In this paper we present a variant of Hough voting for detecting and parsing image objects into their constituent symmetrical parts. The parsing algorithm, given an input image, first uses the Hough voting scheme to detect the salient symmetrical parts by an integrating of color segmentation, depth coherence, and motion grouping. The output of the parsing algorithm is an object graph which is further denoised with the optimization of dominant sets. Subsequently, the detected objects and their parts are tracked across video frames to capture the object part movements which are used to learn and classify human activities. The proposed approach has a significant advantage: no models are learned in advance in the object detection and parsing algorithm. Experimental results show that the proposed method gives good performance on publicly available datasets in terms of detection accuracy and recognition rate.