On double-level kernel self-similarities for 3D motion trajectory description and recognition

Abstract

3D Spatiotemporal trajectories can provide compact and informative clues in motion analysis of human bodies, robots and moving objects. This paper proposes a new framework for 3D motion trajectory-based recognition, which can achieve satisfactory performance in both accuracy and efficiency. Motion trajectories are firstly transformed into the Double-level Kernel Self-similarity Matrices (DKSM). The DKSM representation is constructed by investigating the pair-wise kernel distance within each trajectory itself at the trajectory level and the component level respectively, which has shown strong invariance ability and descriptive power. As each matrix in the DKSM representation can be viewed as a gray-scale image, the well-proven Histogram of Oriented Gradients (HOG) descriptors extracted from the DKSM are concatenated as the final DKSM-HOG descriptor. Next, we train a SVM classifier for multiple class recognition with the training DKSM-HOG descriptors as the input. Finally, extensive motion trajectory recognition experiments are conducted on two public datasets to demonstrate the effectiveness of the proposed method.