Kalman Filter Aided Depth-based Motion Saliency Detection in Human Activity Recognition Applications

The computational resources required by video processing tasks in human activity recognition (HAR) applications can be reduced if the input data is limited to only those regions which contain motion. This demands accurate identification of these regions to maximize the computational savings while maintaining similar performances of the video processing tasks. This work explores the use of Kalman filtering in combination with depth-based motion saliency to find the regions of human activity within video content. The Kalman filter tracks salient regions throughout the RGBD (Red, Green, Blue plus Depth) video and facilitates merging of regions pertaining to the same object. Experimental results show that a superior performance is achieved when compared to the state of the art BSUV-Net 2.0 algorithm.