GPU-Accelerated Tensor Decomposition for Moving Object Detection from Multimodal Imaging

Moving object detection is a crucial component of video surveillance. With advances in sensory technologies, a multimodal image sensor can collect various signals such as infrared signals and depth signals for accurate moving object detection. Recent approaches regard these signals as a single tensor and apply tensor decomposition for classifying moving objects. However, the size of tensor is usually large from the multimodal image sensor, which makes the decomposition extremely slow in practice. In this paper, a parallel framework is proposed to take advantage of the graphics processing unit (GPU) for accelerating tensor decomposition. Compared with the original tensor decomposition in the central processing unit (CPU), our best GPU implementation achieves up to 380% speedup in moving object detection on public data sets.