基于CSK的目标跟踪稳健算法

发布时间：2019-03-20 15:45

【摘要】：近些年,针对视频帧图像的目标跟踪技术成为计算机视觉研究的热点,通过目标检测实现目标跟踪的方法备受青睐,机器学习的思想被应用于模型更新,基于核相关滤波理论的目标跟踪算法在跟踪精度和实时性等方面取得了突出的成果。然而,光照剧烈变化、目标尺度变化、目标被完全遮挡以及目标在帧间发生大位移等来源于复杂实际环境的不可控因素仍然是目标跟踪研究的难点与挑战。本文基于循环结构核(Circulant Structure Kernels,CSK)跟踪方法对上述问题进行深入研究,取得如下成果:(1)通过对图像特征进行研究分析,并对CSK跟踪算法在频域进行多通道扩展,使算法能应用梯度方向直方图(Histogram of Oriented Gridients,HOG)、颜色名(Color Name,CN)、局部二值模式(Local Binary Pattern,LBP)等更优秀的视觉特征,增强算法对目标的表观能力,削弱光学变化和几何变化对目标跟踪的影响。(2)通过对源图像缩放变换构建图像金字塔,然后分层提取HOG特征,构建基于HOG特征的金字塔样本集,训练金字塔核相关滤波分类器(Pyramid Kernel Correlation Filter,PKCF),实现目标尺度检测,并根据目标的尺度变化调整跟踪矩形框和采样窗口的尺度,减小目标模型的误差积累,提高目标跟踪精度,完成CSK的尺度自适应改进。(3)在CSK跟踪流程中引入Kalman滤波器,充分利用目标运动状态信息,对目标在下一帧中可能出现的位置进行初步预测,然后再利用PKCF在预测位置附近进行目标中心位置校准和尺度检测,实现目标检测自适应,改进CSK跟踪算法当前帧目标检测区域固定在上一帧目标中心位置附近的缺陷,解决目标被完全遮挡和帧间大位移的问题。(4)对于Kalman滤波器与PKCF的更新,将离线更新与在线更新相结合,实现目标模型和分类器参数的自适应更新。首先利用跟踪效果好的目标模型和分类器参数建立备选方案,当跟踪精度下降或目标被完全遮挡时,启用备选方案代替在线目标模型和分类器参数离线更新PKCF。Kalman滤波器对当前帧进行位置预测的状态输入是上一帧PKCF获得的校准目标位置,即利用上一帧PKCF的输出对当前帧Kalman滤波器进行状态更新。(5)将尺度自适应、检测自适应以及更新自适应的思想与遮挡处理机制相结合,提出本文的最终算法:基于预测—校准—更新的目标跟踪稳健算法。最后从标准测试集VOT和实景拍摄的视频集中选取几组具有光照变化、尺度变化、目标遮挡等不同挑战的视频进行对比实验。本文算法与CSK算法的对比实验结果表明,本文算法成功实现尺度自适应改进,在.一定程度上解决了目标被完全遮挡和帧间大位移的问题,另外,跟踪精度和成功率也大幅提高。本文算法与CSK、KCF、CN、MOSSE、TLD、Struck算法在整体性能上做对比实验,结果表明,文本算法在中心位置误差、跟踪精度和成功率方面表现最优,在跟踪帧率方面表现不足。
[Abstract]:In recent years, target tracking technology for video frame images has become a hot topic in computer vision research. The method of target tracking through target detection has been favored, and the idea of machine learning has been applied to model updating. The target tracking algorithm based on kernel correlation filtering theory has achieved outstanding results in tracking accuracy and real-time performance. However, the uncontrollable factors such as intense illumination, change of target scale, complete occlusion of target and large displacement between frames are still the difficulties and challenges in the research of target tracking. Based on the cyclic structure kernel (Circulant Structure Kernels,CSK) tracking method, the above problems are deeply studied in this paper. The results are as follows: (1) through the research and analysis of the image features, the multi-channel expansion of CSK tracking algorithm in the frequency domain is carried out. It enables the algorithm to apply better visual features such as gradient direction histogram (Histogram of Oriented Gridients,HOG), color name (Color Name,CN), local binary pattern (Local Binary Pattern,LBP), and enhance the apparent ability of the algorithm to the target. The influence of optical and geometric changes on target tracking is weakened. (2) the pyramid of the source image is constructed by scaling and transforming the source image, and then the HOG feature is extracted by layers, and the pyramid sample set based on HOG feature is constructed. The pyramid kernel correlation filter classifier (Pyramid Kernel Correlation Filter,PKCF (pyramid kernel correlation filter classifier) is trained to realize target scale detection and adjust the scale of tracking rectangle box and sampling window according to the target scale change to reduce the error accumulation of the target model. The precision of target tracking is improved and the scale adaptive improvement of CSK is completed. (3) the Kalman filter is introduced into the CSK tracking flow to make full use of the moving state information of the target to predict the possible position of the target in the next frame. Then we use PKCF to calibrate and measure the center position of the target near the predicted position to realize the self-adaptation of the target detection, and improve the defect of the CSK tracking algorithm that the detection area of the current frame target is fixed near the center position of the target in the previous frame. The problem of complete occlusion of target and large displacement between frames is solved. (4) for the updating of Kalman filter and PKCF, off-line updating and on-line updating are combined to realize adaptive updating of target model and classifier parameters. Firstly, an alternative scheme is established by using the target model with good tracking effect and classifier parameters. When the tracking accuracy is reduced or the target is completely occluded, Enabling alternatives instead of on-line target models and classifier parameters offline updates the status input of the PKCF.Kalman filter to predict the position of the current frame is the position of the calibrated target obtained by the previous frame of PKCF. That is to say, the output of the previous frame PKCF is used to update the current frame Kalman filter. (5) the idea of scale adaptation, detection adaptation and update adaptation is combined with occlusion processing mechanism. The final algorithm of this paper: robust target tracking algorithm based on prediction-calibration-update. Finally, several sets of video with different challenges, such as illumination change, scale change and object occlusion, are selected from the standard test set VOT and the real-time video set to carry on the contrast experiment. The comparison between the proposed algorithm and the CSK algorithm shows that the proposed algorithm has successfully implemented the scale adaptive improvement. To some extent, the problem of complete occlusion and large displacement between frames is solved. In addition, the tracking accuracy and success rate are also greatly improved. The performance of the proposed algorithm is compared with that of the CSK,KCF,CN,MOSSE,TLD,Struck algorithm. The results show that the text algorithm has the best performance in the center position error, tracking accuracy and success rate, and underperforms in the tracking frame rate.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】