Kernel weighted self-similarity descriptor for 3D motion trajectory recognition

3D motion trajectories provide rich spatiotemporal information in motion characterization of human, robots or other moving objects. An effective unique and invariant descriptor for representing motion trajectories can offer substantial advantages over raw data. However, the existing descriptors are not flexible to make tradeoff between the recognition accuracy and the computational cost. This paper first presents a new descriptor by exploring the local self-similarity of the motion trajectory over time, which is also locality, noisy stability and invariant to group transformation. For a given trajectory, instead of the traditional Euclidean norm, the kernel distances between all pairs of frames are computed to form a kernel weighted Self-Similarity Matrix (SSM). The kernel weighted SSM can be regarded as an image, which is shown to be stable under intra-class variations. We then explore the use of kernel weighted SSM with a support vector machine (SVM) classifier in the context of trajectory-based motion recognition. Finally, our method is validated on the Australian sign language (ASL) dataset. The experimental results show that our method outperforms the previous descriptors in terms of computational cost and has similar or superior performance in the context of recognition accuracy.