Video-based human posture recognition

We propose a human behaviour recognition system based on video sequences. Our aim is to identify one among several kinds of actions performed by a single person in a particular scenery. Each frame will be processed, detecting the moving objects and using a new statistical-based algorithm to erase shadows. The final step consists of the extraction of different kinds of other human contour points. Correspondences between each contour pattern and posture will be achieved by using classical pattern classifiers (K-neighbours and Mahalanobis distance), and also with a modified self organized map (SOM). We will analyze and compare the obtained results, combining the contour patterns with some information concerning temporal relationship in consecutive movements, in order to improve the correct action detection.