基于孪生三分支神经网络的目标跟踪

Object Tracking in Siamese Network Based on Triplets Neural Network

为了在保持较高跟踪精度的前提下,提高智能监控摄像头对单目标跟踪的实时性,提出一种基于孪生三分支神经网络的跟踪方法。首先,基于DenseNet重新构建孪生网络的骨干部分,使用较少的参数量和计算量更加充分地利用不同层次的特征图;其次,添加掩膜分支,对目标模板图像和搜索图像特征间的相似性度量得分重新排序,直接根据最大得分对应的候选响应窗口生成和细化掩膜;最后,定义算法的损失函数。在OTB50/100以及VOT2018基准数据集上对所提算法进行评估,实验结果表明,所提算法相比原始的SiamMask算法在准确性有所提升的情况下更加轻量级,平均帧速提高了2倍,实时性更佳。

In order to improve the real-time performance of intelligent surveillance cameras in single target tracking while keeping high tracking accuracy a tracking method based on the triplets siamese neural network is proposed.Firstly the backbone of the siamese network is reconstructed based on DenseNet to make full use of feature maps of different levels with less parameters and computation amount.Secondly the mask branch is added to the siamese network.Then the similarity scores between the features of target template image and those of the search image are reordered and the mask is directly generated and refined according to the candidate response window corresponding to the maximum score.Finally the loss function of the algorithm is defined.The proposed algorithm is evaluated on OTB50/100 and VOT2018 benchmark data set.Experimental results show that compared with the original SiamMask algorithm the proposed method is more lightweight while its accuracy is improved the average frame speed is increased by two times and the real-time performance is better.