A biologically inspired system for human posture recognition

We present a biologically-motivated system to recognize human postures in realtime video sequences. The system employs event-based temporal difference image between video sequences as input and builds a network of bio-inspired Gabor-like filters to detect contours of the active object. The detected contours are organized into vectorial line segments. After feature extraction, a classifier based on simplified line segment Hausdorff distance combined with projection histograms is implemented to achieve size and position invariant recognition. 86% average recognition rate is achieved in the experiment. Compared to state-of-the art bio-inspired categorization methods shows great computational savings, and is an ideal candidate for hardware implementation with event-based circuits.