基于核相关滤波的TLD跟踪算法的研究

TLD Object Tracking Algorithm Based on Kernelized Correlation Filter

对TLD跟踪算法进行改进,以提高在跟踪目标发生尺度变化或被遮挡时的跟踪性能.首先使用KCF跟踪器替代TLD算法中原有的中值光流跟踪器,并在特征提取时增加目标的Lab颜色特征,在寻找目标位置时引入尺度估计,在模型更新阶段引入跟踪状态判别机制,通过设定跟踪器中输出响应最大值阈值、APCE阈值及检测器中随机蕨分类器阈值来判断跟踪器跟踪结果的可靠性,改善跟踪器在尺度变化、出现遮挡、光照变化等情况下的跟踪效果.针对TLD算法中的检测器,为了减少大量无意义的窗口,提升算法在存在遮挡时的精确性,在检测之前使用Kalman滤波预估出目标位置,在预估位置周围使用改进的级联分类器更精准地定位目标,改进的级联分类器的前两级仍采用方差分类器和随机蕨分类器,第三级则采用改进的KCF跟踪器.在OTB-50数据集上的实验结果分析表明,该算法跟踪性能优于其他算法,能够满足实时性.

TLD tracker is improved in this paper in order to increase the tracking performance when the tracked object scale is changing or in occlusion.For the tracker,KCF tracker is used to replace the original tracker of TLD algorithm,and KCF tracker is improved as follows:the Lab color feature of the object is added in feature extraction,scale estimation is introduced in search for the object location,in the model updating stage,and tracking state judgment mechanism is adopted to determine the reliability of the tracking results by setting the threshold value of maximum output response,the threshold value of APCE and the threshold value of the random fern classifier,so as to improve the tracking effect of the tracker under the conditions of scale variation,occlusion and illumination variation.For the detector,in order to reduce a large number of meaningless windows before detection,Kalman filter is used to estimate the object position,and then an improved cascade classifier is used to locate the object more accurately around the estimated position.Variance classifier and random fern classifier are still used in the first two levels of the improved cascade classifier,and the third level is classified by the improved KCF tracker.Qualitative and quantitative analysis of experimental results on the OTB-50 dataset show that the tracking performance of the proposed algorithm is superior to that of other tracking algorithms and that it can ensure the real-time performance.