基于后验概率度量的粒子滤波跟踪算法研究

 针对遮挡、光照变化、尺度变化等复杂环境中的视觉跟踪问题,提出一种基于后验概率度量的粒子滤波跟踪算法。由于后验概率指标与Bhattacharyya系数指标相比具有更强的峰值特性,采用后验概率指标作为相似性度量函数,通过粒子的更新、推广、观测、估计等步骤实现跟踪算法。通过对实际视频图像序列进行目标跟踪实验,实验结果表明:传统算法只有约50%的图像能够实现尺度自适应,而本文算法采用传统算法25%的粒子就能够收敛逼近目标的真实轨迹,达到更强的抗遮挡能力,90%以上的图像序列都能够实现良好的尺度自适应效果。     

基于最大后验概率的人脸平移与旋转跟踪方法研究

针对目前大多数视觉跟踪算法只能够跟踪平移运动人脸目标,而不能跟踪旋转运动人脸目标这一问题,提出一种基于最大后验概率的人脸目标平移与旋转跟踪方法。在基于最大后验概率的视觉跟踪算法基础上,采用了自适应角度模板进行人脸目标旋转跟踪,与传统的视觉跟踪方法相比,当人脸发生倾斜时,该方法能更精确地得到人脸的位置和方向。实验结果表明,该方法能够有效地跟踪具有平移与旋转运动的人脸目标。     

基于自适应非参数统计模型的彩色目标跟踪算法研究

针对复杂环境条件下的视觉跟踪问题,提出一种基于自适应非参数统计模型的彩色目标跟踪算法。利用目标和背景之间的强度差别,基于自适应核密度估计模型对运动目标进行了非参数统计建模。为了实现具有较大范围运动目标的跟踪,在充分考虑目标和背景之间的相关性前提下,采用目标特征统计的背景加权直方图对搜索区域进行了扩大。为了提高对环境变化的适应能力,根据目标和环境的变化自适应更新目标特征分布模型。通过对实际图像序列的实验,结果表明该算法能够有效跟踪运动目标,并且平均迭代次数比传统方法减少了37.28%。     

基于核相关的尺度自适应视觉跟踪

针对视觉跟踪中目标尺度变化对准确跟踪的不利影响,提出一种基于核相关的尺度自适应视觉跟踪算法。首先,通过建立核岭回归模型构建二维核相关定位滤波器,采用融合后的多通道特征对滤波器进行训练,提高目标定位的精度;然后,对目标区域进行多尺度采样,样本缩放后提取其特征,并构造为一维特征,以此构建一维核相关尺度滤波器,估计出目标的最佳尺度。在OTB2013平台上的实验结果表明,与8种当前主流的跟踪算法相比,本文算法的跟踪精度和成功率均有优势。在尺度变化条件下,本文算法在快速准确跟踪的同时,较好地实现了对目标尺度的自适应跟踪。     

带宽自适应的均值漂移红外目标跟踪算法

当目标尺度发生变化时,传统均值漂移跟踪因窗口尺寸不变导致跟踪目标丢失.为解决该问题,提出一种带宽自适应的均值漂移跟踪算法.该算法在均值漂移框架下提取目标的形状特征,根据目标形状变化自适应的修正核函数带宽,并更新目标模板.实验结果表明,改进算法能很好地适应尺寸变化的目标,能有效提高红外目标的跟踪准确度.     

结合目标检测的多尺度相关滤波视觉跟踪算法

为满足视觉跟踪算法对跟踪精度与跟踪速度的要求,提出一种结合目标检测的多尺度相关滤波视觉跟踪算法。所提算法基于深度学习的目标检测算法找出图像中目标的位置和尺寸,利用相关滤波算法对所给出的目标特征进行视觉跟踪,并在多个尺度中搜索最优响应;当检测到相关滤波响应值异常时,停止对模型更新;当连续数帧响应值异常时,则在全图范围内搜索目标位置和尺寸。所提算法通过对跟踪状态进行评估和模型更新率自适应调整,解决了传统相关滤波类算法跟踪误差随时间积累的问题,且具有较大的跟踪速度和较高的精度。结果表明:在Matlab平台下,所提算法的平均定位精度为0.593,平均交叠率精度为0.784,帧率为65.3 frame/s。     

基于自适应尺度的TLD目标跟踪算法

为提升TLD目标跟踪算法的每帧处理速度,以达到在更高分辨率视频中跟踪目标的实时性要求,在TLD算法框架的基础上,提出了一种基于自适应尺度检测学习的目标跟踪算法(AS-TLD)。当跟踪目标成功时,选取当前帧跟踪到的目标尺度及几个相邻的尺度作为下帧检测目标时滑动窗口尺度的选取范围;而当跟踪失败时,则选取在TLD算法初始化阶段,根据跟踪目标及视频图像大小选定的尺度来保障长时间跟踪目标,从而有效减少了平均每帧扫描的窗口数量。实验结果表明,该方法不仅有效地降低了检测模块的检测时间,显著提高了整体算法速度,而且通过动态选取尺度,在一定程度使得TLD各个模块更加协调,跟踪精确度得到提升。     

Mean Shift跟踪算法中核函数参数的评估与分析

Mean Shift是一种基于特征的对目标实现快速跟踪的算法,传统的Mean Shift算法由于跟踪中物体的尺度变化会使跟踪偏离目标乃至跟踪失败,并且原有的自适应地对跟踪窗宽的调整,是基于对核窗宽的改变来得到的。在＂固定跟踪窗宽—改变核窗宽＂的基础上对目标进行跟踪,对目标空间定位精度进行了评估与分析,通过实验结果表明改变核函数参数能改善目标跟踪的精度。     

自适应窗口选取的Mean-Shift目标跟踪

当目标尺度发生变化时,传统Mean-Shift跟踪算法因跟踪窗口尺寸不变容易导致跟踪目标丢失,为解决此问题,本文提出一种带宽自适应算法对目标尺度变化进行检测,从而实现模板更新.该算法分别将模板图像与当前帧目标图像分割成等间隔半径的若干同心圆,通过计算模板图像与当前帧图像不同环层之间相似性度量,根据相应环层之间相似性度量关系确定当前帧模板带宽更新参量,最后利用kalman滤波完成模板尺度更新,从而实现目标稳定跟踪.实验证明,当目标尺度发生变化时,目标模板自动更新,能够实现目标稳定跟踪;相对传统Mean-Shift跟踪算法,目标跟踪可靠性能得到了提高.     

基于自适应Kalman滤波的机器人运动目标跟踪算法

针对足球机器人比赛时的模型变化及其环境噪声先验估计不准确的问题,提出一种基于自适应卡尔曼滤波的足球机器人视觉跟踪算法。该算法将一种基于减背景的运动目标识别的方法与自适应卡尔曼滤波跟踪模型进行结合,对背景进行实时更新,并通过形态学滤波去除残留的小区域,从而准确的识别运动目标,通过自适应的在线调整运动模型参数来保证模型预测值的准确性,进而提高了目标跟踪时的匹配效率,实现了目标的精准、迅速跟踪。通过实验证明,该算法是很有效的,具有推广价值。     

An adaptive window object tracking algorithm based on variable resolution

This paper presents an adaptive window object tracking method based on variable resolution. It copes with the change in size of the object during visual tracking. On the basis of the visual tracking algorithm, based on maximum posterior probability, we analyze the posterior probability index on the inside and outside panes of the object window, then build a mathematical model for adjusting object size with an adaptive window. Since the resolution changes according to the size of the object, this thesis uses a statistical sampling method of the feature by variable resolution. The resolution of the statistical feature is correspondingly changed in object tracking with an adaptive window. The resolution of a larger object is decreased, which realizes an object tracking method with adaptive window based on variable resolution.     

A coarse-to-fine kernel matching approach for mean-shift based visual tracking

Mean shift is an efficient pattern match algorithm. It is widely used in visual tracking fields since it need not perform whole search in the image space. It employs gradient optimization method to reduce the time of feature matching and realize rapid object localization, and uses Bhattacharyya coefficient as the similarity measure between object template and candidate template. This thesis presents a mean shift algorithm based on coarse-to-fine search for the best kernel matching. This paper researches for object tracking with large motion area based on mean shift. To realize efficient tracking of such an object, we present a kernel matching method from coarseness to fine. If the motion areas of the object between two frames are very large and they are not overlapped in image space, then the traditional mean shift method can only obtain local optimal value by iterative computing in the old object window area, so the real tracking position cannot be obtained and the object tracking will be disabled. Our proposed algorithm can efficiently use a similarity measure function to realize the rough location of motion object, then use mean shift method to obtain the accurate local optimal value by iterative computing, which successfully realizes object tracking with large motion. Experimental results show its good performance in accuracy and speed when compared with background-weighted histogram algorithm in the literature.     

成像目标跟踪目标建模技术综述

由目标跟踪的数学模型得出,影响目标跟踪性能的三个主要因素为目标状态转移模型、滤波算法和目标建模技术.对目标建模技术进行了综述和分析,分别从特征选择、特征的统计建模和相似性度量三个方面进行了阐述.以畸变不变性、目标/背景分辨能力作为性能评价手段定性地比较了国内外文献中提出的多种目标表征模型.指出了目标跟踪中目标表征模型自...     

Implementation of opto-digital stereo object tracking system

In this paper, a new opto-digital stereo object tracking system using the variable window mask and the optical binary phase extraction joint transform correlator (BPEJTC) is proposed. At the first step, with the distance information from the stereo camera to the tracking object easily acquired by the structural elements of a stereo vision system, the area of the tracking object can be digitally extracted by using the variable window mask. And, at the second step, by carrying out the optical BPEJTC between this reference image obtained from the variable window mask and the stereo input image, the coordinates of the tracking object's location can be acquired, and then with these values, the convergence angle and the pan/tilt of the stereo tracking camera can be finally controlled. From some experimental results, the proposed system is found to be able to effectively extract the area of the target object from the input image having the background noises by using the variable window mask. And, with the location values of the tracking object obtained by using the optical BPEJTC, the convergence angle and the pan/tilt of the stereo cameras can be controlled. Finally, a feasibility of real-time implementation of the adaptive stereo object tracking system using the proposed algorithm is also suggested.     

Multiple small objects tracking based on dynamic Bayesian networks with spatial prior

This paper proposes an end-to-end algorithm for multiple small objects tracking in noisy video using a combination of Gaussian mixture based background segmentation along with a Dynamic Bayesian Networks (DBNs) based tracking. Background segmentation is based on an adaptive backgrounding method that models each pixel as a mixture of Gaussians with spatial prior and uses an online approximation to update the model, the spatial prior is constructed for small objects. Furthermore, we create observation model with hidden variable based on multi-cue statistical object model and employ Kalman filter as inference algorithm. Finally, we use linear assignment problem (LAP) algorithm to perform the models matching. The experimental results show the proposed method outperforms competing method, and demonstrate the effectiveness of the proposed method.     

基于动态目标建模的粒子滤波视觉跟踪算法

提出一种根据场景变化动态建立目标模型的粒子滤波视觉跟踪算法.该方法首先选择简单且具有互补性的特征描述当前图像,并统一采用直方图法对这些特征进行建模;然后在粒子滤波框架下,根据巴塔恰里亚测度评价各个目标特征和背景特征之间的可区分程度,动态调整特征间的置信度;并对各个特征似然函数的噪音参量进行在线估计和更新,使其似然函数的度量标准达到统一.分析和实验表明,该算法性能优于仅仅采用多特征融合进行粒子滤波视觉跟踪的方法,对摄像机运动、混淆干扰、遮挡及目标外观大小的改变具有更强的鲁棒性.