摘要
针对基于深度神经网络视频实时目标跟踪的研究主要集中在骨干网优化方面,设计及训练均比较复杂的问题,本文提出一种新的表观增强的深度目标跟踪算法。首先,通过引入简单且易于实现的传统表观特征,直接与深层语义特征融合,增强了类内目标的区分能力;其次,通过投票机制及自适应搜索模块,增强了算法跟踪的鲁棒性。在VOT系列数据集上的测试结果表明:本文算法相对基准算法在平均重叠期望(EAO)上均有2%~4%的提升,准确性及鲁棒性指标达到甚至部分超过了现有复杂优化算法。
Aiming at the problem that the research of video real-time target tracking based on deep neural network mainly focuses on the optimization of backbone network, and the design and training are relatively complex, a new apparent enhanced depth target tracking algorithm with apparent enhancement was proposed. Firstly, by introducing simple and easy-to-implement traditional apparent features, it is directly fused with deep semantic features, which enhances the discrimination ability of objects within a class. Secondly, through voting mechanisms and adaptive search modules, the robustness of tracking algorithm is enhanced. The test results on the VOT series data set show that compared with the benchmark algorithm, the average overlap expectation(EAO) of the proposed algorithm has been improved by 2%~4%, and the accuracy and robustness have reached or even partially exceeded the existing complex optimization algorithms.
作者
王侃
苏航
曾浩
覃剑
Kan WANG;Hang SU;Hao ZENG;Jian QIN(Institute of Southwest Electronic Technology,Chengdu 610036,China;Microelectronic and Communication College,Chongqing University,Chongqing 400030,China)
出处
《吉林大学学报(工学版)》
EI
CAS
CSCD
北大核心
2022年第11期2676-2684,共9页
Journal of Jilin University:Engineering and Technology Edition
基金
天奥基金项目(2020154704)。
关键词
计算机视觉
孪生网络
表观信息
投票机制
自适应搜索
computer vision
siamese network
apparent information
voting mechanism
adaptive search