Behavior recognition using Pictorial Structures and DTW

In recent years, there has been an increasing interest in monocular human behavior recognition system. The first step in behavior recognition is the measurement stage. We use an extended Pictorial Structure to speed up the detection. This extension adds a temporal term to the global energy function of the framework. We use a simple to complex approach in action recognition by decomposing it to its basic elements. The human body parts motions are tracked and classified individually. The body parts motions are matched using an adapted Dynamic Time Warping (DTW) that use a multilevel approach that projects a solution from a coarse resolution and refines the projected solution. The results of the DTW matching are used to activate hierarchical Petri Nets, or to act as input to Neural Network, used to classify the behavior.