基于卷积神经网络的鲁棒高精度目标跟踪算法

目标跟踪是计算机视觉中重要的研究领域之一.为了跟踪复杂场景中外观变化剧烈的目标，本文提出了一种基于卷积神经网络的目标跟踪算法.算法中的网络模型结构包括预训练的特征提取层和自适应更新的分类器层.在开始跟踪前，首先训练全连接层和分类器层的参数，以及目标的特征与位置之间的线性关系.其次，定义了评估跟踪结果可信度的标准.如果得到的跟踪结果的可信度较高，则根据跟踪结果的特征调整位置，提高跟踪结果的精确度.最后，在训练网络时，每次迭代都选择分类器得分的最高的负样本参与训练.该策略可以提高模型的分辨能力.在OTB50测试集中的实验结果表明，我们的算法取得了良好的跟踪结果.

A Robust and Accurate Object Tracking Algorithm Based on Convolutional Neural Network

Object tracking is one of the most important area of computer vision.In order to track the object whose appearance changes dramatically in complex scene,we propose a tracking algorithm based on the convolutional neural network.The network of our tracker has two parts:the feature extraction layer and the adaptive classifier layer.At the beginning,we train a fully-connected layer,a softmax layer and the linear relationship between feature and position of these samples.Next,we define a reliability of the tracking result.If the result is reliable,we will adjust the result location according to its features.Finally,in the network training process,we select the negative samples with max classifying scores in each iteration.The strategy could improve distinguishability of our tracker.Experiments on OTB50 show that our tracker could achieve state-of-the-art performance.