Detecting humans: Analysis and synthesis as a pattern recognition problem

Abstract

In this paper, we address a few important image analysis problems, which are fundamental to the design of Perceptual User Interface (PUI). We use an inexpensive stationary desktop camera to collect the video streams and use them as input to the system. We present an algorithm for segmenting moving foreground object of interest from a complex, but stationary background. This algorithm can cope with illumination changes due to shadows, Automatic Exposure Correction (AEC) and long term illumination changes in the environment. The segmentation is done real time and works well for both indoor and outdoor scenes. The detected foreground is recognized as human being based on head shoulder profile. The head shoulder profile is extracted as a pattern using an unique physics based approach, which is robust to noise and eliminates the need of median filtering to remove noise. We obtain Ring Data Vectors as features from the pattern, which is used to classify them into different clusters. The patterns which are similar form a cluster in the N dimensional vector space formed by the orthogonal basis of Ring Data Vectors. A data set for each possible cluster is formed by initial training and manual classification of different sets of patterns belonging to different human pose. Unknown new pattern extracted from the segmented foreground is synthetically recognized and classified into one of the existing template pattern. The system has been used in design of Visual Computer games and can be used in the design of PUI's for pervasive computing applications, surveillance etc.