基于CUDA的并行视觉背景提取算法

摘要

为了提高视觉背景提取(ViBe)算法的检测效果和检测速度,对其进行改进和并行化处理.该算法基于像素及其邻域像素之间的差值构建最终的背景样本,减少边缘噪声对于检测精度的影响;背景样本更新阶段中使用自适应的时间二次抽样因子,加快消除单帧初始化所形成的“鬼影”,提高算法的健壮性.同时分析了ViBe算法的并行点,并用CUDA语言实现该算法.实验结果表明,该算法能有效减少边缘噪声,快速消除“鬼影”,且检测速度较CPU端有大幅提升.%Improvements are made to ViBe (Visual Background extractor) algorithm for enhancing its test effect and test speed.The algorithm utilizes several pixel values from consecutive frames to train and construct the final samples of background during initialization,thus to decrease the ghost's effects on precision in detecting an object and partially remove the noises on the edge of background objects.In the process of pixel matching,a classified threshold and a match threshold for scene complexity of the pixel are adjusted dynamically to adapt complicated background scenes and diminish the disturbances of incorrect foreground pixels.In order to reinforce the robustness of the improved algorithm,an adaptive timesubsampling factor is utilized to accelerate elimination of the ghost caused by the departure of background objects and recovery of the background samples.Experiments show that the algorithm presented here can remove the ghost and noises effectively and is appropriate for complicated background scenes detections.