基于二值特征和结构化输出支持向量机目标快速跟踪算法.docVIP

基于二值特征和结构化输出支持向量机目标快速跟踪算法 　　摘要：复杂场景下基于判别式分类器的目标跟踪通常采用复杂的外观表示模型以提高跟踪精度，但影响了算法的实时性。为此，提出一种基于半色调的二值特征来描述目标的外观，在此基础上对结构化输出支持向量机（SVM）的核函数进行改进，实现了判别模型的快速更新和判别；同时提出一种基于分块匹配的判别模型更新策略，保证了跟踪过程中样本的可靠性。在Benchmark数据集上进行的测试实验中，与压缩跟踪（CT）算法、跟踪学习检测（TLD）算法和核化的结构化输出跟踪（Struck）算法相比，在跟踪速度上，该算法分别提高了0.2倍、4.6倍、5.7倍；在跟踪精度上，当重叠率阈值取0.6时，该算法的成功率达到0.62，而其他三种算法的成功率均在0.4以下，当位置误差阈值取10时，该算法的精度为0.72，而其他三种算法精度均小于0.5。实验结果表明该算法在发生光照变化、尺度变化、严重遮挡和突变运动等复杂情况下均具有很好的鲁棒性和实时性。 　　关键词：目标跟踪；结构化输出；二值特征；支持向量机；判别模型 　　中图分类号： TP391.41 　　文献标志码：A 　　Abstract： The object tracking algorithm based on discriminative classifier usually adopts complex appearance model to improve the tracking precision in complex scenes， which relatively influences the realtime performance of tracking. To solve this problem， a binary feature based on halftone was proposed to describe the object appearance and the kernel function of structured output Support Vector Machine （SVM） was improved， so as to realize fast updating and discriminating of discriminative model. In addition， a discriminative model updating strategy based on part matching was proposed， which can ensure the reliability of the training samples. In the experiments conducted on Benchmark， compared with the three algorithms including Compressive Tracking （CT）， Tracking Detection Learning （TLD） and Structured Output Tracking with Kernels （Struck）， the proposed algorithm had better performance in tracking speed with the increases of 0.2 times， 4.6 times and 5.7 times respectively. On the aspect of tracking precision， when overlap rate threshold was set to 0.6， the success rate of the proposed algorithm reached 0.62， which was higher than the success rates of the other three algorithms that were all less than 0.4； when the position error threshold was set to 10， the precision of the proposed algorithm reached 0.72，while the precisions of the other three algorithms were all less than 0.5. The experimental results show that the proposed algorithm obtains good robustness and realti