基于增强RPN的孪生网络目标跟踪算法

A Siamese Network Object Tracking Algorithm Based on Enhanced RPN

目前孪生网络跟踪器已经具有比较良好的表现,但是对于卷积神经网络所提取的特征仍没有较好地利用其特点,同时孪生网络通过相似性学习进行跟踪的特性使跟踪器的准确性和鲁棒性存在不足。提出了一种金字塔式特征融合的方法,根据骨干网络特征提取层不同深度具有不同侧重的特点提高网络对目标的表征能力,然后使用注意力机制对区域推荐网络(Region Proposal Network,RPN)进行增强,最终实现更精准更鲁棒的跟踪。在OTB100数据集的实验中,新提出的SiamERPN(Siamese Enhanced RPN)算法分别得到了0.668的成功率和0.876的精度,测试结果好于基线算法和其他对比算法。

The current siamese network trackers have a relatively good performance,but still do not make good use of the features extracted by convolutional neural networks,while the characteristics of siamese network tracking through similarity learning lead to insufficient accuracy and robustness.This paper proposed a pyramidal feature fusion method,which can improve the backbone to represent the target because the layers with different depths have different focus.And then the attention mechanism is used to enhance the regional proposal network(RPN)to achieve more accurate and robust tracking.In the experiments on the OTB100 dataset,the Siamese Enhanced RPN(SiamERPN)proposed in this paper obtaines a success score of 0.668 and an accuracy score of 0.876,which are better than those of the baseline algorithm and other comparison algorithms.