Spatio-temporal motion field descriptors for the hierarchical action recognition system

Spatio-temporal motion field descriptors for the hierarchical action recognition system have been developed. At the lower level, blurred motion fields representing local features such as speed and direction are estimated from video sequences. At the higher level, feature vectors are calculated by applying a parallel updating scheme to measure the similarity between the input video sequences and so-called template patches, which are selected from learning samples. Instead of traditional patch descriptors that applied computationally expensive methods such as PCA to reduce the size of feature vectors, we extended previous hardware-based ideas of Averaged Principal-Edge Distribution (APED) in face detection by incorporating temporal information. In addition, since the system only employs simple computations such as summation and Boolean operators, it can be implemented on VLSI chips without much effort to achieve real-time performance. We tested our system on a popular action recognition database and promising recognition performance has been reported.