光学学报, 2017, 37 (11): 1115005, 网络出版: 2018-09-07
基于多层卷积特征融合的目标尺度自适应稳健跟踪 下载: 1117次
Target Scale Adaptive Robust Tracking Based on Fusion of Multilayer Convolutional Features
机器视觉 稳健跟踪 深度学习 卷积特征 相关滤波 尺度估计 machine vision robust tracking deep learning convolutional features correlation filtering scale estimation
摘要
针对复杂跟踪条件下目标的稳健跟踪和精确尺度估计问题,提出了一种基于多层卷积特征融合的目标尺度自适应稳健跟踪算法。算法首先利用VGG-Net-19深层卷积网络架构提取目标候选区域的多层卷积特征,通过相关滤波算法构建二维定位滤波器,得到多层卷积特征并进行加权融合,从而确定目标的中心位置;然后通过对目标区域进行多尺度采样,提取其梯度方向直方图特征构建一维尺度相关滤波器,确定目标的最佳尺度。实验结果表明,与6种当前主流跟踪算法相比,该算法取得了最好的跟踪成功率与精度,同时在跟踪过程中较好地实现了对目标快速尺度变化的自适应跟踪,且具有较快的跟踪速率。
Abstract
For the problems about robust tracking and precision scale estimation of the visual objects in the complex tracking conditions, a target scale adaptive robust tracking algorithm based on the fusion of multilayer convolutional features is proposed. First, the multilayer convolutional features are extracted from the target candidate area using VGG-Net-19 deep convolutional network architecture. By constructing the two-dimensional location filters by correlation filtering algorithm and fusing the multilayer convolutional features, the center location of the target is determined. Then, through the multi-scale sampling of target, the histogram of oriented gradient features are extracted to construct the one-dimensional scale filter to achieve the optimal scale estimation. The experimental results show that the proposed algorithm gains the best success rate and precision compared with the six state-of-the-art methods. Meanwhile, this algorithm achieves an adaptive tracking to the fast scale changing of target effectively, and possesses a fast tracking speed.
王鑫, 侯志强, 余旺盛, 金泽芬芬, 秦先祥. 基于多层卷积特征融合的目标尺度自适应稳健跟踪[J]. 光学学报, 2017, 37(11): 1115005. Xin Wang, Zhiqiang Hou, Wangsheng Yu, Zefenfen Jin, Xianxiang Qin. Target Scale Adaptive Robust Tracking Based on Fusion of Multilayer Convolutional Features[J]. Acta Optica Sinica, 2017, 37(11): 1115005.