Negative space analysis for human action recognition

• Main Author: Shah Atiqur Rahman
• Other Authors: Leung Karhang, Maylor
• Format: Theses and Dissertations
• Language: English
• Published: 2012
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/50706
• Institution: Nanyang Technological University

In this thesis a novel region-based method is proposed to recognize human actions. Other region-based approaches work on the silhouette of the human body which is termed as the positive space according to art theory. On the contrary, we investigate and analyze regions surrounding the human body, termed as the negative space for human action recognition. This concept takes advantage of the naturally formed negative regions that can always be described/approximated using simple shapes, simplifying the job for action classification. Negative space is less sensitive to segmentation errors, overcoming some limitations of silhouette based methods such as leaks or holes in the silhouette caused by background segmentation. The proposed method also addresses the problem of long shadows which is one of the major challenges of human action recognition. Some systems attempted to suppress shadows during segmentation process but our system takes input of segmented binary images where the shadow is not suppressed or discarded. This makes our system less dependent on the segmentation process. Nonetheless, the system also tackles the problem of video sequences retrieval by query of complex human actions. The system consists of some hierarchical processing blocks: histogram analysis on segmented input image, motion and shape feature extraction, pose sequence analysis by employing Dynamic Time Warping and at last classification by Nearest Neighbor classifier. We evaluated our system by most commonly used datasets and achieved higher accuracy than the state of the arts methods. The proposed system is found relatively robust to partial occlusion, variations of clothing, noisy segmentation, illumination change, deformed actions etc. Furthermore, good accuracy can be achieved even when half of the negative space regions are ignored. This makes the proposed system robust with respect to segmentation error and distinctive from other approaches. In future, this system can be combined with positive space based methods to make an improved system in terms of accuracy and computational costs.