前景判别的局部模型匹配目标跟踪

目的 在复杂背景下,传统模型匹配的跟踪方法只考虑了目标自身特征,没有充分考虑与其所处图像的关系,尤其是目标发生遮挡时,易发生跟踪漂移,甚至丢失目标。针对上述问题,提出一种前景判别的局部模型匹配(FDLM)跟踪算法。方法 首先选取图像帧序列前m帧进行跟踪训练,将每帧图像分割成若干超像素块。然后,将所有的超像素块组建向量簇,利用判别外观模型建立包含超像素块的目标模型。最后,将建立的目标模型作为匹配模板,采用期望最大化(EM)估计图像的前景信息,通过前景判别进行局部模型匹配,确定跟踪目标。结果 本文算法在前景判别和模型匹配等方面能准确有效地适应视频场景中目标状态的复杂变化,较好地解决各种不确定因素干扰下的跟踪漂移问题,和一些优秀的跟踪算法相比,可以达到相同甚至更高的跟踪精度,在Girl、Lemming、Liquor、Shop、Woman、Bolt、CarDark、David以及Basketball视频序列下的平均中心误差分别为9.76、28.65、19.41、5.22、8.26、7.69、8.13、11.36、7.66,跟踪重叠率分别为0.69、0.61、0.77、0.74、0.80、0.79、0.79、0.75、0.69。结论 实验结果表明,本文算法能够自适应地实时更新噪声模型参数并较准确估计图像的前景信息,排除背景信息干扰,在部分遮挡、目标形变、光照变化、复杂背景等条件下具有跟踪准确、适应性强的特点。

Foreground discrimination in local model-matching tracking

Objective Under a complex background, a majority of the traditional model-matching tracking methods only consider the characteristics of the moving target without fully utilizing the relationship between the moving target and the image for object tracking, especially when the target was occluded during the process of object tracking. Consequently, these methods allow the moving target to drift easily; as a result, the moving target is sometimes lost. To solve these problems, a novel object-tracking approach based on foreground discrimination of local model matching is proposed. Method First, the algorithm selects previous m frames of the image frame sequences for tracking training, and each image frame is divided into superpixel blocks. Second, the vector cluster is composed of all superpixel blocks, and the object model that contains superpixel blocks is established by the discrimination appearance model. Finally, the algorithm takes the object model as a matching model, adopts expectation maximization to estimate the foreground information, and utilizes foreground discrimination to match the local model. Hence, the tracking object is determined. Result Compared with other excellent tracking algorithms, the proposed target-tracking algorithm can accurately and effectively adapt to the complex changes in the target states of a video scene through foreground discrimination and local model matching and can adequately solve the problems of tracking drift under various uncertain factors. This algorithm can also achieve the same or even higher tracking accuracy compared with existing model-matching tracking methods. For the video sequences of Girl, Lemming, Liquor, Shop, Woman, Bolt, CarDark, David, and Basketball, the average center errors are 9.76, 28.65, 19.41, 5.22, 8.26, 7.69, 8.13, 11.36, and 7.66, respectively, and the tracking overlap ratios are 0.69, 0.61, 0.77, 0.74, 0.80, 0.79, 0.79, 0.75, and 0.69, respectively. Conclusion Experiment results indicate that the proposed target-tracking algorithm can adaptively update noise model parameters in real time, accurately estimate the foreground information of images according to different image sequences, eliminate background information interference, and achieve tracking accuracy and adaptability under the conditions of partial occlusion, target deformation, illumination changes, and complex background.