基于IoU约束的孪生网络目标跟踪方法

基于孪生网络的跟踪方法通过离线训练跟踪模型，不需要对跟踪模型进行在线更新，兼顾了跟踪精度和速度。现有孪生网络目标跟踪方法使用固定阈值选择正负训练样本易造成训练样本漏选问题，且训练时分类分支和回归分支之间存在低相关性问题，不利于训练出高精度的跟踪模型。为此，提出了一种基于交并比(IoU)约束的孪生网络目标跟踪方法。通过使用动态阈值策略根据预定义锚框与目标真实框的相关统计特征，动态调整正负训练样本的界定阈值，提升跟踪精度。所提方法使用IoU质量评估分支代替分类分支，通过锚框与目标真实框之间的IoU反映目标位置，提升跟踪精度，降低模型的参数量。在数据集VOT2016、OTB-100、VOT2019、UAV123上进行了对比实验，所提方法均有较好的表现。在VOT2016数据集上，所提方法的跟踪精度比SiamRPN方法高0.017，期望平均重叠率为0.463，与SiamRPN++相比仅差0.001，实时运行速度可达220帧/s。 

Object tracking method based on IoU-constrained Siamese network

The tracking method based on the Siamese network trains the tracking model offline. Therefore, it maintains a good balance between tracking accuracy and speed, which attracts the interest of a growing number of researchers recently. The existing Siamese network object tracking method uses a fixed threshold to select positive and negative training samples, which is easy to cause the problem of missing training samples, and such methods have low correlation between the classification branch and the regression branch during training, which is not conducive to training a high-precision tracking model. To this end, an object tracking method based on intersection over union (IoU)-constrained siamese network is proposed. By using a dynamic threshold strategy, the thresholds of positive and negative training samples are dynamically adjusted according to the relevant statistical characteristics of the predefined anchor boxes and the real boxes. Thereby improving the tracking accuracy. In addition, the proposed method uses the IoU quality assessment branch to replace the classification branch, and reflects the position of the target through the IoU between the anchor box and the target ground-truth frame, which improves the tracking accuracy and reduces the amount of model parameters. The proposed object tracking method based on the IoU-constrained Siamese network has been compared and tested on four datasets: VOT2016, OTB-100, VOT2019, and UAV123. Ideal results have been achieved in these datasets. The tracking accuracy of the proposed method in this paper is 0.017 higher than SiamRPN on the VOT2016 dataset. And with a real-time running speed at 220 frame/s, the expected average overlap rate is 0.463, which is only 0.001 worse than SiamRPN++.