一种改进的基于孪生卷积神经网络的目标跟踪算法

Improved Target Tracking Algorithm Based on Siamese Convolution Neural Network

目标跟踪是计算机视觉的一个重要研究方向.为了对视频序列中单个目标进行准确定位和实时跟踪,本文采用孪生卷积神经网络解决深度神经网络模型更新不及时和训练数据不足的问题;同时在孪生卷积神经网络的特征提取子模块中加入SE-Net,先利用卷积层提取图像的空间特征信息,再利用特征通道间的相互依赖关系建模,强化有效通道特征,进一步提升网络的特征表征能力,从而提升特征提取的效果;最后通过区域推荐网络进行目标定位和边框微调.本文使用OTB2015数据集进行实验,以平均覆盖率和OPE方法作为评估标准,实验结果表明平均覆盖率为66.6%,OPE准确率图和成功率图也均显示跟踪效果优于其他算法.

Target tracking is an important research direction of computer vision.In order to accurately locate and timely track the single target in video sequence,the Siamese convolution neural network is adopted in this paper to solve the problem that the deep neural network can’t be updated in time and training data is insufficient.At the same time,SE-Net is added to the feature extraction submodule of Siamese CNN.The spatial feature information of the image is extracted by using the convolution layer,and the interdependencies between feature channels are used for modeling to strengthen the characteristics of the effective channels and further improve the network’s feature characterization ability,so as to improve the effect of feature extraction.Finally,the Region Proposal Network is adopted for target positioning and boundary fine-tuning.In this paper,OTB2015 dataset is used for experiment and average coverage and OPE method is used as evaluation criteria.The results reveal that average coverage is 66.6%and both success rate and accuracy rate diagrams show our approach is better than other algorithms.