Motion Saliency Detection using Depth Information for Human Action Recognition Applications

Video processing tasks require the analysis of an enormous amount of data. However, in most practical applications this analysis is not required on the whole frame or video but is limited to the regions where some action is taking place. If these regions can be identified with low effort, the total computation time required to process and analyze such videos can be drastically reduced. In this work a new motion saliency detection algorithm intended to be incorporated in a human action recognition pipeline is presented. The approach uses depth data to improve the performance of the difference of frames method which is computationally simple and efficient. Results show that the algorithm achieves a performance which is comparable to the state of the art while requiring a much lower computational time.