一种低参数的孪生卷积网络实时目标跟踪算法

A Low Parameter Real-Time Target Tracking Algorithm Based on Siamese Convolutional Network

针对基于深度学习的目标跟踪算法模型参数多、难以部署于嵌入式设备上的问题,提出一种改进的孪生卷积网络实时目标跟踪算法。设计一个非对称卷积模块来构建整个网络框架,通过非对称卷积模块的压缩层减少模型参数量,利用非对称层进行特征融合,以在保证精度的同时压缩模型大小。使用三元组损失函数代替逻辑损失函数进行模型训练,在输入不变的情况下提取表达性更强的深度特征,从而完成目标跟踪任务并提高模型的跟踪精度。在GOT-10K、OTB100和VOT2016基准上对算法性能进行测试,结果表明,该算法能够将模型大小降为3.8×106,且速度与精度均优于SiamFC、KCF和DAT等跟踪算法。

The existing target tracking algorithms based on deep learning face deployment problems on embedded devices due to their large number of parameters.To address the problems,this paper proposes a real-time target tracking algorithm based on low parameter siamese convolutional network. By designing an asymmetric convolution module to build the entire network framework,the compression layer of the asymmetric convolution module is used to reduce the number of parameters,and the asymmetric layer is used for feature fusion to reduce the size of the model while maintaining accuracy. At the same time,the triple loss function is used to replace the logical loss function for model training,and the more expressive depth features are extracted for target tracking with the same input,which improves the tracking accuracy of the model. The performance of the algorithm is tested on the GOT-10 K,OTB100 and VOT2016 benchmarks.Experimental results show that the algorithm reduces the model size to 3.8×10~6,and outperforms SiamFC,KCF,DAT and other tracking algorithms in terms of speed and accuracy.