基于孪生网络与证据推理规则的视频目标跟踪

Visual Tracking Based on Siamese Networks and Evidential Reasoning Rule

卷积神经网络(Convolutional Neural Networks,CNN)已经作为特征提取的方法被广泛用于视频目标跟踪。由于CNN每层卷积特征在给定的视频序列上具有不同响应以及随着层数的增加,提取的特征更为高级和抽象,所以仅使用最高层的卷积特征用于视频跟踪准确率可能大大降低。为解决这一问题,文章在孪生网络的基础上提出基于证据推理规则(Evidential Reasoning Rule,ER Rule)加权位置信息组合方法。描述了基于ER规则组合位置信息的过程,然后构建了位置信息组合模型。通过实际跟踪实验验证了模型的实际效果。

Convolutional Neural Networks(CNN) are widely used in visual tracking fields represented by computer vision and natural language processing. Since the convolutional features of each layer of CNN have different responses on a given video sequence. As the number of layers increases, the layer features are more advanced and abstract. Therefore, only using the highest layer of convolutional features for visual tracking,the accuracy may be reduced. To solve this problem, This paper proposes a weighted location information combination method based on Evidential Reasoning Rule(ER Rule) and the Siamese network.Describes the process of combining location information based on Evidential Reasoning Rule, then built the location information combination model.