结合掩膜与孪生网络的目标跟踪方法研究

针对在舰船目标跟踪过程中由于视场角变化导致的跟踪精度下降的问题,基于全连接孪生神经网络的目标跟踪方法,融合了目标分割策略,提出了基于SiamMask的实时目标跟踪算法。该算法将二元掩模运算作为孪生网络的一个分支,以实现对目标的分割,在获取目标位置信息的同时,获取目标的外观信息,使得该网络的损失函数得到显著增强。由传统的轴对称目标跟踪框,改进为可根据目标形状、外观而自适应调整的可旋转矩形框。将该算法与传统基于孪生网络的跟踪算法SiameseNet,以及基于相关滤波的KCF跟踪算法进行实验仿真测评对比,通过计算各算法的中心误差,结果表明该算法的跟踪精度较两者分别提升了19.5%和24.5%,且运行速度可达30 fps,满足了舰船目标跟踪对于准确性和实时性的要求。

Research on Target Tracking Method Based on Mask and Siamese Network

Aiming at the problem that the tracking accuracy decreases due to the change of the field of view during the ship target tracking,a real-time target tracking algorithm based on SiamMask is proposed,which combines the target segmentation strategy with the method of full-connected Siamese neural network. In this algorithm,binary mask operation is taken as a branch of Siamese network to realize the segmentation of the target. While acquiring the location information of the target,the appearance information of the target is obtained,so that the loss function of the network is significantly enhanced. The traditional axis-symmetrical target tracking frame is improved to a rotatable rectangular frame which can be adaptively adjusted according to the shape and appearance of the target. Compared with traditional Siamese-network-based tracking algorithm SiameseNet and KCF-based tracking algorithm based on correlation filtering,the simulation results show that the tracking accuracy of the proposed algorithm is improved by 19.5% and 24.5% respectively,and the running speed can reach 30 fps,which meets the requirement of accuracy and real-time for ship target tracking.