摘要
针对高清视频监控系统中精确运动检测的高实时性需求,提出一种基于计算统一设备架构(CUDA)的运动检测算法。采用一种改进帧差与背景差分相结合的方法,减少背景更新干扰,提升运动检测的精确性。在CUDA内进行视频运动检测计算,避免传统图形处理器硬解码后视频数据在显示内存与CPU之间传输的问题。运用块内多线程合并访问共享内存的方式,减少52.9%全局内存访问量,解决CUDA大规模访问全局内存延迟较大的问题。实验结果表明,该算法在保证准确性的同时,针对高清视频每秒可传输52.6帧,能够满足实时性要求。
Aiming at high real-time demand for accurate motion detection in high-definition video surveillance system, this paper proposes an efficient motion detection algorithm based on Compute Unified Device Architecture（CUDA）. By using an improved frame difference and background subtraction method of combining,background interference update is reduced and the accuracy of motion detection is enhanced. By performing video motion detection calculations in the CUDA, the traditional Graphic Processing Unit （ GPU ） hard decoded video data avoid storing in display memory transferred to CPU,which is the bottleneck problem. This paper uses multiple threads within a block coalesced access shared memory, reduces the amount by 52. 9% of the global memory access, solves large-scale access to global memory latency CUDA larger problem. Experimental results show that for high-definition video monitoring system,the proposed motion detection algorithm can reach 52. 6 frames per second,ensuring the accuracy as well,and can meet the real-time requirements.
出处
《计算机工程》
CAS
CSCD
2014年第11期288-291,296,共5页
Computer Engineering
基金
浙江省级重点科技创新团队基金资助项目(2011R09021-02)
关键词
视频监控
运动检测
帧差法
背景差分法
硬解码
合并访问
video surveillance
motion detection
frame difference method
background subtraction method
hard decode
coalesced access
