Real-time Motion Detection in Extremely Subsampled Compressive Sensing Video

Compressive sensing (CS) has shown promising results in different areas of signal processing as it provides an elegant framework for simultaneous signal acquisition and compression. Iterative CS reconstruction algorithms limit the practical applicability of CS due to high computational complexity. In this paper, a real-time reconstruction method based on deep neural networks is presented and applied to spatial video CS. In order to show feasibility of learning-based CS approach in real-world applications, we perform motion detection on videos reconstructed from extremely sub-sampled measurements. Experimental results performed on a synthetic dataset show a comparison between performance of motion detection algorithms in the original and the compressively sensed video. The results confirm that most of the information used by standard motion detection algorithms is preserved in the low-dimensional measurement space. Inspired by the obtained results, we propose an adaptive sampling scheme in which CS video camera operates at extremely low measurement rate when there is no motion in the scene. Otherwise, when motion is detected, measurement rate is increased accordingly.