Human action and activity recognition has recently gained much importance as a research topic in the computer vision community. The first reason for its prominence is the various critical applications. One can quote security applications such as behavioral biometrics, human gait recognition; intelligent surveillance, social applications such as care for elderly and/or sick people, and more generally ambient intelligence; multimedia applications such as content-based video analysis, video summarization, animation, motion capture and video editing. The second reason is the availability of huge quantities of video data and the growing need for processing those using tools of artificial intelligence. As a case in point we can consider daily amount of uploaded videos to YouTube, films and television movies in databases, and the output of surveillance cameras.
Human action and activity recognition is a challenging problem due to such factors as the randomness and variability in the rate of the action, the style of the subject, the human pose and the viewing angle, changes in the background and, occlusions, resulting in significant intra-class variations. Our goal is to design a real-time embedded system robust against to both human-to-human and human-to-object interaction in realistic scenarios. Our scope mainly involves: i) unifying spatio-temporal descriptors and their geometrical configuration into a graph-based representation, intending to provide a real-time performance; ii) a depth-map assisted approach (exploiting human body part identification, pose estimation) which is now possible with a consumer camera, called Kinect; iii) handling complex human activities by higher-level processing, grammars and reasoning methods.