用基于二值化规范梯度的跟踪学习检测算法高效跟踪目标

为提高复杂环境下TLD(Tracking-Learning-Detection)算法的跟踪精度和速度,提出基于二值化规范梯度(BING)的高效TLD目标跟踪算法。在跟踪器中引入基于时空上下文的局部跟踪器失败预测方法和全局运动模型评估算法,提高了跟踪器准确度和鲁棒性;用BING算法取代滑动窗口搜索策略,结合级联分类器实现目标检测,减少了检测器的检测范围,提高了检测的处理速度;将训练样本权重整合到在线学习过程中,改进级联分类器的分类准确度,解决了目标漂移问题。对不同的图片序列实验结果表明:本算法的跟踪正确率达85%,帧率达19.79frame/s。与原始TLD算法及其他主流跟踪算法相比较,该算法在复杂环境下具有更高的鲁棒性、跟踪精度及处理速度。

Efficient target tracking by TLD based on binary normed gradients

To improve the tracking precision and processing speed of the Tracking-Learning-Detection(TLD) algorithm under a complex environment, an efficient TLD target tracking algorithm based on BInary Normed Gradient(BING) algorithm was proposed. The local tracker failure predicting method based on spatial-temporal context and the global motion model estimation algorithm was introduced into the tracker to improve its precision and robustness. Then, the BING algorithm was used to replace a sliding window for searching the target to detect the candidate target by combining with a cascaded classifier, so that to reduce the search space and improve the processing speed of the detector. The sample weight was integrated into the online learning procedure to improve the accuracy of the classifier and to alleviate the drift to some extents. The experimental results on variant sequences demonstrate that the accurate rate and the frame rate of the improved TLD are 85% and 19.79 frame/s, respectively. Compared with original TLD and state-of-the-art tracking algorithm under the complex environment, the improved TLD has the superior performance on robustness, tracking precision and tracking speeds.