基于多实例回归模型的视觉跟踪算法研究

目前大部分基于检测的跟踪算法将跟踪任务看作是一个类别分类的任务，当目标发生形变或者遇到相似物体的干扰时，容易导致模型漂移。为此该文提出一种多实例回归跟踪算法。在该算法中，跟踪任务被认为建立在实例模型之上更为合适，为此该文利用一帧图像建立实例模型，并在时间序列上建立多实例模型集合表征目标的最近状态；为使跟踪算法能够适应目标的形变，利用逻辑回归将实例模型作为隐变量，由最近若干帧建立的正负样本集作为训练集，共同构建多实例回归跟踪模型。由于跟踪模型在整体上对多个实例模型建模，把它们紧密地联系在一起，故能有效应对目标的形变；由于模型漂移仅会影响当前帧的实例模型，各个实例模型之间互相独立，故跟踪算法能够有效减轻模型漂移对鲁棒跟踪的影响。实验中，OTB 2013数据库和UAV 123数据库被用来验证该文算法，DeepSRDCF, Siamese-fc等算法作为对比算法，实验结果表明，该文算法不仅充分发挥了基于多实例回归模型进行跟踪的优势，在形变等属性上具有很好的性能，而且在整体性能上优于各类先进算法3%~5%。

Visual Object Tracking Based on Multi-exemplar Regression Model

Most of the tracking-by-detection algorithms treat the tracking task as a category classification task, when the target experience deformation or encounter similar objects interference, the model drift is prone to occur. In this paper, a multi-exemplar regression tracking algorithm is proposed. In this algorithm, the exemplar model is considered to be more appropriate for tracking task, the exemplar model is set up by a frame image information, and the multi-exemplar model established in the time series can represent the target current state; in order to make the tracking algorithm adapt to the target deformation, the exemplar model is considered as the hidden variable by logistic regression model, together with the training sets from several recent frames sampling, can jointly build multi-exemplar regression tracking model. As the tracker builds multi-exemplar model on the whole, linking them together closely, it can effectively deal with the target deformation. Since the model drift only affects the exemplar model at current frame, each exemplar model is independent of each other, so the tracking algorithm can effectively reduce the influence of model drift on robust tracking. In the experiment, OTB 2013 benchmark and UAV 123 benchmark are used to verify the algorithm, DeepSRDCF, Siamese-fc and other algorithms act as the contrast algorithms, the experimental results show that the proposed tracker not only gives full play to the advantages of tracking based on multi-exemplar regression model, but also has good performance in deformation and background blur scene, and achieves three to five percent more than other advanced algorithms in the metrics of success rate and precision.