HHMM Based Recognition of Human Activity Motion Trajectories in Image Sequences

A bstract In this paper, w e present a m ethod for recognition of hu m an activity as a series of actions from an im age sequ ence. T he diffi cu lty w ith the problem is that there is a chicken-egg dilem m a that each action needs to be extracted in advance for its recognition bu t the precise extraction is only possible after the action is correctly identifi ed. In order to solve this dilem m a, w e u se as m any m odels as actions of ou r interest, and test each m odel against a given sequ ence to fi nd a m atched m odel for each action occu rring in the sequ ence. F or each action , a m odel is designed so as to represent any activity containing the action. T he hierarchical hidden M arkov m odel (H H M M) is em ploy ed to represent the m odels, in w hich each m odel is com posed of a su bm odel of the target action and su bm odels w hich can represent any action, and they are connected appropriately. S everal experim ental resu lts are show n.