Primitive Human Action Recognition Based on Partitioned Silhouette Block Matching

This paper deals with the issue of recognizing primitive human actions through template matching with time series silhouette images. Although existing methods based on this simple approach can recognize a subject’s action from a low-resolution image sequence, which is a basic requirement for surveillance applications, their recognition accuracy decreases considerably for corrupted silhouettes due to occlusion. To deal with this problem while keeping algorithm simplicity, we propose a novel method, which integrates template matching results for temporally and spatially partitioned silhouette blocks. Experimental results indicate that our method outperforms the existing methods in the accuracy of action recognition for corrupted silhouettes.