基于改进SiamFC的实时目标跟踪算法

A Real-Time Target-Tracking Algorithm Based on Improved SiamFC

针对全卷积孪生神经网络SiamFC在目标跟踪速度以及网络判别能力上有待提升的问题,提出了一种基于改进SiamFC的实时目标跟踪算法。将原网络结构中的第二层卷积层替换为深度可分离卷积,通过减少参数计算量,提高了跟踪速度;为了提高网络判别能力,第三层卷积层使用混合深度卷积,通过不同尺寸的卷积核提取特征,实现多特征融合,提取到鲁棒性更强的特征;采用预处理后的ILSVRC2015数据集,使用随机梯度下降法对网络进行训练,并在OTB2015、VOT2016、ILSVRC2015数据集上对算法性能进行测试。实验结果表明,该算法和SiamFC算法相比,在跟踪成功率、跟踪精度以及跟踪速度上都有一定的提升,并能够满足实时跟踪要求。

A real-time target tracking algorithm based on improved SiamFC(a fully convolutional Siamese network)is proposed herein to improve the target tracking speed and the network discrimination ability of traditional SiamFC.First,after the first conventional convolution layer,a depth-wise separable convolution is set to improve the tracking speed by reducing the amount of parameter calculation.Second,the third convolution layer is set as a mixed depth-wise convolution(MixConv)to improve the recognition ability of the network.We extracted features from the convolution kernel of different sizes and concatenated them in the channel to achieve a multifeature fusion and extract more robust features.Finally,the preprocessed ILSVRC2015 data set was used to train the network using the random gradient descent method,and the performance of the algorithm was tested on the OTB2015,VOT2016,and ILSVRC2015 data sets.Experimental results show that compared with the SiamFC algorithm,our algorithm shows a certain improvement in the tracking success rate,tracking accuracy,and tracking speed,and can meet real-time tracking requirements.