基于软特征理论的目标跟踪研究

Research of Object Tracking Based on Soft Feature Theory

针对目标跟踪过程中遮挡、形状与尺度变化导致目标易丢失的问题,提出了一种新的基于软特征(Soft Feature,SF)的目标前趋预测跟踪方法.该方法首先在视频图像中选取待跟踪目标区域,统计目标区域内的初始像素点,计算初始像素相邻时域图像中与其具有相同变化强度的像素点,滤掉分散的像素点并标记像素群;然后将离散的像素群质心坐标拟合成时域轨迹,计算时域轨迹的空间谱带和边缘谱带,合并谱带信息中具有可微分的相同变化强度的频率,得到软特征信息及软特征约束模型;最后,根据软特征及其约束模型对视频中运动目标进行跟踪,并以前趋冲击强度对目标运动状态和软特征进行前趋预测,限定目标检测范围并得到预测特征,以此实现目标前趋预测跟踪.该方法抓住了目标在形变过程中其前景区域的灰度特征具有可微分的同频率变化的显著特点,这是目标区别于复杂背景以及对形变目标进行长时间稳定跟踪(Long-term Tracking)的重要信息源.软特征的提取可以有效凸显目标区域和前趋信息,同时能有效抑制干扰信息.实验结果表明,软特征跟踪方法不仅可以克服遮挡、形状和尺度变化对目标跟踪的影响,而且具有较高的实时性、准确性和鲁棒性能.与现有的跟踪方法(State-ofthe-art Trackers)相比,软特征理论具有以下优点:采用软特征跟踪运动目标,对目标形状变化和尺度伸缩问题具有很好的抗干扰性;采用前趋冲击强度对目标前趋进行预测,可以有效解决因遮挡而导致目标丢失的问题;由于目标检测范围较小,软特征数据量较低,无需存储目标姿态模型,其计算复杂度和空间复杂度较低,跟踪速度较快.

A novel approach to object predictive tracking,which is based on the soft feature （SF）,is proposed in this paper to overcome the problems of occlusion,shape and scale changes, which cause the situation to lose object.Firstly,the object region is selected as the tracking region in the video to initialize the tracking process.Several image sampling segments are made ready for counting the initial pixels with the gray of a continuous gradient features pixels in the tracking region.Then the pixel groups are made up of initial pixels,which have the same gray-scale changes in intensity,and the discrete pixel group centroid coordinate is fitting for a time-domain trajectory curve.Edges of spectrum and track of spectrum are calculated by the time-domain trajectory curve.The frequency variation,which has the same intensity of differentiable in spectrum,is merged together,which is marked as soft feature and constructed as soft feature constraint model.Finally,the object is tracked with soft feature and its constraint model,the range of object detection is limited and tracked by the state of predictive motion and predictionnbsp;soft feature,which is predicted by precursor shock strength.Above is the concrete step of the implementation of soft feature theory,which is designed to achieve predictive tracking of moving objects.This method captures the salient features of an object,which has the same frequency variation and differentiable during its foreground gray features deformation.As an important feature to distinguish a moving object from the complex background,a long-term object tracking becomes achievable.It realizes an effective extraction of the soft feature by extracting the edges of spectrum,track of spectrum and constructing soft feature constraint model,which has the same frequency variation and differentiable in the object foreground region,it can highlight the differences between the object and background,then the object area is quite obvious,the interference region is inhibited.The effects of dense background information,object shape change and scale expansion on object tracking are solved effectively.The prediction of the former information of the moving object is carried out,which is done by extracting the precursor shock strength of moving object.This method solves the problem caused by large range occlusion. Experimental results show that the proposed approach has overcome the occlusion,shape and scale changes,and has good adaptability to non-rigid object tracking with real-time,high accuracy and robust tracking performance.Even compared with the state-of-the-art trackers,the advantages of this method are as follows,this method use soft feature and constraint model to track the moving object,it has better anti-jamming performance for the object shape variation and scale expansion, which makes the method tracks object with high accuracy,better stability and robust performance;It can solve the problem when the target is occluded by using the precursor shock strength to predict the object location;Soft feature has a higher tracking speed due to the small target detection range,low amount of data,and it does not need to store a variety of appearance model,the computation complexity and the space complexity is lower.