Human action classification in partitioned feature space

Video surveillance plays a prominent role in law enforcement, personal safety, traffic control, resource planning and security of assets, etc. The need for such systems is increasing every day, with a number of surveillance cameras deployed in public places to analyze human actions. In this paper, a fast and a simple method is proposed to recognize human activities such as walking, running, jumping and bending by analyzing video sequences. Since, no pan, tilt and zoom camera is assumed, a simple background subtraction is used to extract the foreground region. Histogram projection technique is applied to remove shadow from the foreground image. The extreme points of the foreground region are detected using star skeletonization algorithm are then localized by partitioning them into equal sized blocks. The proposed method has been tested on Weizmann dataset and test video sequences and is found to process a frame at the rate of 0.066s and provides an accuracy of 96.87%.