摘要： 针对多数传统目标特征无法实现复杂场景下的鲁棒视觉跟踪问题,提出一种新的视觉跟踪算法。采用卷积神经网络(CNN)提取目标更加鲁棒的深度特征,同时融合具有旋转不变性的局部二值模式纹理特征,弥补CNN深度特征在旋转适应性上的不足。根据CNN网络训练速度慢的问题,引入离线预训练方法,提高在线特征提取效率。实验结果表明,与DLT算法相比,该算法在跟踪测试集上的跟踪精度提高14.08%,运算效率提高10.47%,能够较好地适应目标表观变化,具有较强的鲁棒性和跟踪时效性。

关键词: 深度学习, 卷积神经网络, 纹理, 局部二值模式, 自适应融合, 视觉跟踪

Abstract: Most traditional features may not be good enough for robust visual tracking in complex environments.This paper proposes a new visual tracking algorithm using Convolutional Neural Network(CNN) to learn robust generic object feature representation,which fuses with Local Binary Pattern(LBP) texture to offset CNN’s shortcoming in rotation invariance.Aiming at the slow training speed of CNN,it uses offline pre-training on an auxiliary tiny image dataset to improve the efficiency of online feature extraction.Experimental results on an open tracker benchmark show that this algorithm is more accurate,improving tracking precision by 14.08% on average,and more efficient improving lomputaction efficiency by 10.47% on average,compared with DLT.It suits target changes and background influence,and has strong robustness and tracking efficiency.

Key words: deep learning, Convolutional Neural Network(CNN), texture, Local Binary Pattern(LBP), adaptive fusion, visual tracking

中图分类号: 

• TP391