一种融合感知哈希的快速压缩跟踪算法

Fast Compressive Tracking Algorithm Based on Perceptual Hashing

针对快速压缩跟踪算法(FCT)分类器参数更新盲目、目标尺寸固定和未能跟踪目标完全遮挡再出现的问题,提出一种融合感知哈希的快速压缩跟踪算法(Fast compressive tracking algorithm based on perceptual hashing,PH-FCT).首先,使用压缩特性构建目标和背景的贝叶斯分类器,同时生成目标的感知哈希描述子;使用分类器获得下一帧响应值最高的样本,以样本为中心采集不同尺寸区域,计算它们与目标的汉明距离,若最小汉明距离小于阈值,则视当前尺寸区域为目标区域,更新目标信息(目标位置、尺寸和感知哈希描述子)与分类器参数,并标记当前帧检测到目标,否则不更新且标记当前帧未检测到目标.当上一帧被标记为未检测到目标,则当前帧使用全图等间隔采样,样本个数与FCT算法粗采样一致,使用分类器得出响应值最高的样本,再以该样本中心为圆心,半径为5的圆形区域遍历精确采样,得出最有可能是目标的样本,最后通过判断汉明距离决定是否更新参数.实验结果表明,该算法在抗遮挡性、有效性和鲁棒性上优于FCT算法,且拥有较好的目标自找回能力,为目标的快速跟踪提供一种新的方法.

According to the problems that classifier parameter update blindly,the size of target fixed and the target appearing again after completely occlusion is failed to track for fast compressive tracking algorithm （FCT）,this paper proposes a perceptual hashing fast compressive tracking algorithm（PH-FCT）.First of all,a naive Bayes classifier was set up by compressive characteristics of target and background,and perceptual hashing descriptor of target was generated； In the next frame,the highest response sample will be found with the classifier and based on its center and size some different size region samples will be got.Hamming distance of those samples with target is calculated.If the minimum Hamming distance is less than the threshold,the current size is regarded as the size of target,update the target information （location,size and perceptual hashing descriptor）,classifier parameters,and set the mark（target is detected in this frame） to true.If the mark of the last frame is false,the current frame use equal interval sampling from full image,the number of samples as same as rough sampling of FCT to obtain the highest response sample by the classifier.And then,we take sample accuratly in the circular region and obtain a sample of the most likely target.Hamming distance finally decides whether it is target and whether the parameters is updated.Experimental results show that PH-FCT is superior to the FCT in anti-occlusion,effectiveness and robustness,and has better ability to target self-retrieval,which provides a new method for fast tracking target.