Abstract:

In recent years, the correlation filtering and deep learning theory has developed rapidly and has been widely used in target tracking. However, there are problems in accuracy or speed. The method based on siamese network structure can balance the accuracy and speed, and gradually became the mainstream method of target tracking. This paper introduces the basic concepts of target tracking technology, analyzes the development and shortcomings of traditional methods such as correlation filtering. Then structure of the siamese network and design principles and latest development of tracking algorithm based on siamese network structure are emphasized, and the performance of the related methods is compared. Aiming at the shortcomings of existing tracking methods based on siamese network structure, the future development trend is prospected.

Key words: target tracking, siamese network, deep learning

摘要：

近年来相关滤波和深度学习理论快速发展，在目标跟踪中得到广泛应用，但在精度或者速度方面存在问题，基于孪生网络结构的方法能够在精度和速度之间取得平衡，逐渐成为了目标跟踪的主流方法。介绍了目标跟踪技术的基本概念，分析相关滤波等传统方法的发展及其存在的不足。着重阐述孪生网络的结构和基于孪生网络结构的跟踪算法的设计原理及其最新进展，并对比相关方法的性能。针对现有基于孪生网络结构的跟踪方法的不足，展望未来的发展趋势。

关键词: 目标跟踪, 孪生网络, 深度学习