Recognizing postures and head movements from video sequences

Video analysis is a fast growing field of study, with applications ranging from surveillance systems to face detection. Volvo Group Trucks Technology wishes to apply some of the innovative solutions discovered in recent times to their own study data to minimize the need of manual video processing. This master's thesis, is an initial step in this direction. Video data of drivers driving in a virtual environment has been gathered from frontal and side view. The frontal view data has been used to estimate gaze direction of the driver while the side view has been used to identify posture. Gaze has been estimated mainly using the popular Viola-Jones face detection algorithm combined with extracting feature points which are tracked frame to frame using the Kanade-Lucas-Tomasi feature tracker. For posture identification, marks were placed on the side of the driver and tracked using the Kanade-Lucas-Tomasi feature tracker. Training data was used to create clusters using the k-medoids method and driver postures were identified using mark coordinate data from sample videos using mainly a k-nearest-neighbors approach. Results are promising both for gaze estimation and posture identification even though ambiguity in matching continuous data to discrete evaluations lead to a noticeable amount of errors.