摘要:
为实现对运动目标的实时跟踪,提出中心对称局部相似数量( center symmetric -local similarity number ,CS-LSN)这一新的局部显著度纹理描述算子,并将其引入到目标表征模型。该算子通过分析中心像素及其8邻域像素之间的大小关系,在局部相似数量( local similarity number ,LSN)纹理算子的基础上,针对其无法区分同一局部显著度下的不同纹理结构的问题,增加以中心像素为对称点的局部梯度信息,提取出候选目标区域中具有CS-LSN主要模式的真实目标像素,有效地抑制了背景像素的影响并减少了后续目标表征模型的计算量;利用真实目标像素的CS-LSN纹理特征和色度特征构建直方图,完成目标表征;进而将其嵌入到均值漂移(mean shift,MS)框架完成跟踪。实验结果表明,该方法在目标与背景相似、部分遮挡、光照变化及物体形变等情况下均能完成鲁棒跟踪,目标大小为29像素×25像素时,处理速度约为25帧/s,可满足实时应用的需求。
Abstract:
In order to complete the task of real -time object tracking , this paper proposes a new local texture description operator , which is called center symmetric -local similarity number ( CS-LSN) .The operator analyzed the center pixel and its eight neighborhood pixels , on the basis of the local similarity number ( LSN);as the texture operator fails to distinguish the same local saliency of different texture structures , the authors added the local gradient information using the center pixel as the symmetric point , extracted the key pixels which correspond only to the five major patterns of the CS -LSN in the target candidate region , effectively restrained the influence of the background pixels and reduced the computation in the target representation model , and then represented target by CS-LSN texture feature and the chromaticity in the true target pixel , which was embedded into the mean shift (MS) tracking framework.The experimental results show that the proposed method can continuously track the target when the background and the color are similar under the condition of changing illumination and in the occlusion cases.The processing speed can reach 25 frames or so per second when the object is 29 pixel ×25 pixel, so it can satisfy the demand of real -time application .
下载: