A compact descriptor CHOG3D and its application in human action recognition

In this paper, we propose a new method to calculate local features. We extend the FAST corner detector to the spatiotemporal space to extract the shape and motion information of human actions. And a compact peak-kept histogram of oriented spatiotemporal gradients (CHOG3D) is proposed to calculate local features. CHOG3D is calculated in a small support region of a feature point, and it employs the simplest gradient, the first-order gradient, for descriptor calculation. Through parameter training, the proper length of the CHOG3D is determined to be 80 elements, which is 1/12.5 times the dimension of HOG3D in the KTH dataset. In addition, it keeps the peak value of quantized gradient to represent human actions more exactly and distinguish them more correctly. CHOG3D overcomes the disadvantages of the complex calculation and huge length of the descriptor HOG3D. From a comparison of the computation cost, CHOG3D is 7.56 times faster than HOG3D in the KTH dataset. We apply the algorithm for human action recognition with support vector machine. The results show that our method outperforms HOG3D and some other currently used algorithms.