孪生网络跟踪算法是利用离线训练好的网络提取目标特征并进行匹配,从而实现跟踪。而离线训练深度特征在表征任意形式目标时将目标从背景中分离开的性能较差。为此,提出一种基于目标感知特征筛选的孪生网络跟踪算法。将经过裁剪处理后的模板帧和检测帧送入到ResNet50的特征提取网络分别提取目标和搜索区域的浅层、中层、深层特征;在目标感知模块中,通过设计一个回归损失函数来学习对目标敏感的特征,根据反向传播的梯度来确定每个卷积核的重要性程度,并以此来激活相对重要的卷积核筛选较重要的目标感知特征;将筛选得到的特征送入到SiamRPN模块,进行目标、背景的二分类判别和边界框的坐标回归,从而得到一个精确的目标边界框。在OTB2015和VOT2018两个标准数据集上进行测试实验,结果表明该算法可以实现对目标的稳健性跟踪。

Tracking algorithms implemented in Siamese networks utilize an offline training network to extract features from a target object for matching and tracking. The offline-trained deep features are less efficient for distinguishing targets with arbitrary forms from the background. Therefore, we proposed a tracking algorithm for a Siamese network based on target-aware feature selection. First, the cropped template and detection frames were sent to a feature extraction network based on ResNet50 to extract the shallow, middle and deep features of the target and search regions. Second, in the target-aware module, a regression loss function was formulated for target-aware features and an importance scale for each convolution kernel was obtained based on backpropagated gradients. Then, the convolution kernels with large importance scales were activated to select target-aware features. Finally, the selected features were inputted into the SiamRPN module for target-background classification and the bounding box regression was applied to obtain an accurate bounding box of the target. Results of experiments on OTB2015 and VOT2018 datasets confirm that the proposed algorithm can achieve robust tracking of the target.