A compact 3D descriptor in ROI for human action recognition

In this paper, a new action recognition system is proposed, which employs 3D FAST corner detection in ROI, compact 3D descriptor to represent action information, and SOM to learn and recognize actions. Through detecting 3D FAST corners in ROI, action information of shape and motion can be obtained, and noise corners can be deleted at the same time. Furthermore, based on 3D HOG, we produce a simpler descriptor which is proposed by shortening the support region of interest points, combining symmetric bins after orientation quantization using icosahedron, and keeping the top value bin of quantized histogram. Compared with the descriptor before adjustment, our descriptor uses only 80 bins other than 960 bins to describe one interest point, which saves much computation time and memory. Our frame matching experiment on descriptor also certifies that our descriptor outperforms the previous one. Our descriptor is applied to recognize actions on KTH and Hollywood databases, and the results show that it performs well.