Real-time Video Motion Detection Algorithm Research Based On CUDA

In recent years, driven by the demand of high definition video market, various video encoding formats were generated. Interlacing algorithm is the key technology of these different encoding formats compatible with each other. In order to get the right picture, the motion phase of the video input should be computed precisely. Now the main motion detection algorithms are field-based, region-based or block-based. It is difficult to detect the videos with complicated background or multiple rigions with existing motion detection algorithms. Also, the real-time demand is hardly being met when using complicated algorithms.GPGPU provides a way of dealing with the high quality videos with mess data, meeting the real-time demand of complicated algorithms. Compared with CPU, GPU has larger data transfer bandwidth, more computing units, and more floating-point computing power. After the NVIDIA CUDA architecture was published, the processing time of GPGPU is reduced range from several times to hundred times, or even more.This article gives out an improved context admptive motion detector, which combine those field-based, region-based and block-based motion detector algorithms. This algorithm not only suits for those simple videos, but also can deal with complicated videos efficiently and precisely. Firstly, the differences between frames are computed, and the input is aparted by regions and macroblocks. Secondly, processing the input video with highpass filter, then get the context parameters through combing effect estimation, block detail estimation and single line detection algorithms. At last, lowpass filter and the parameters are used to detect the pricise motion in the video.In this aritical, the following work is done:1) Developing the HD (High Definision) video simulink platform. This platform is based on CUDA architecture, and provides multiple interfaces for algorithm controlling and processing. 2) Realizing the algorithm in this article on CUDA. Also, the parts in the algorithm with mess data and paiallel computing are optimized based on CUDA. By optimizing the bandwidth, data accessing and registers, twice optimum is done on CUDA code.3) Integrated simulation testing. The testing results show that the improced algorithm in this article can detect the motion in the video with complicated background and multiple rigions precisely, and the processing time is 22fps (frame per second), which practically meets the realtime demand.