Segmentation and tracking of body parts for human activity recognition

• Main Author: Kristo.
• Other Authors: Chua Chin Seng
• Format: Final Year Project
• Language: English
• Published: 2009
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: http://hdl.handle.net/10356/17880
• Institution: Nanyang Technological University

Motion is an important cue for the human visual system. Current major goal of computer vision research is to recognize and understand human motion, activities and continuous activity. The research that was started on tracking a single person has nowadays grown into tracking, recognizing and understanding interactions among several people, like the scenes seen in MRT station or food courts. Interpreting such a scene with multiple interacting individuals is complex, because similar configurations may have different context and meanings. Understanding the complete meaning of an image sequence involving human interactions requires the thorough knowledge about the sequence, including the task of monitoring, inferring intention and ultimately interpreting the sequence. In the event of multiple interacting individuals, problems may arise in the process of segmenting the body parts into semantically meaningful parts. These problems are caused by the high degree of freedom (DOF) of the human body and irregular shape deformation caused by loose clothing. However, the biggest challenges are caused by the mutual occlusion and the presence of shadows that are inevitable in situations that involve multiple humans. This report presents a method for segmentation and tracking of body parts in a bottom-up fashion, by using appearance-based method for combining multiple free form blobs in color video sequences, dissimilarity comparison between blobs and human body model to assist in tracking the body parts.