基于图像传感器的上下文快速压缩感知算法

摘要

针对快速压缩感知算法在目标被遮挡、光照变化较大时存在跟踪不稳定的问题，提出了基于图像传感器的上下文快速压缩感知跟踪（ FCT）算法。新算法首先在Haar-like特征中引入时空上下文特征，通过目标周围的空间信息和时间上的递推关系协助估计目标的位置。通过改进的随机测量矩阵同时提取目标的纹理特征和灰度特征，加强了特征的稳定性，提高跟踪的准确性。通过方差分类器预判定候选样本，减少判定的次数，并减少错误的候选样本。改进的FCT算法对光照、旋转、尺度缩放都有良好的不变性，且不易发生跟踪漂移。实验证明：改进的FCT算法优于压缩感知跟踪（ CT）算法和FCT算法。%Aiming at problems that fast compressive tracking( FCT ),algorithm has poor robustness in target occlusion and illumination changes,propose a FCT algorithm with context based on image sensors. First,the new algorithm introduces temporal and spatial context features in the Haar-like feature,and assists to estimate target position by spatial information around target and time recurrence relation. By improved random measurement matrix,extract simultaneously texture features and gray features of target,stability of feature is enhanced,accuracy of target tracking is improved. Pre-judge candidate sample by variance classifier,reduce number of decision,reduce number of wrong candidate samples. The improved FCT algorithm has good invariance for illumination,rotation and scale changes,tracking drift also not easy to happen. It can be proved that the improved FCT algorithm is superior to CT and FCT algorithms.