应用LTRNet卷积特征的ECO目标跟踪算法改进

The application of LTRNet convolution features in the improvement of ECO

中文关键词  目标跟踪  应用高效卷子算子(ECO)  具有线性两步方法性质的残差网络(LTRNet)  特征压缩  目标优化

英文关键词  object tracking  ECO  LTRNet  feature compression  objective optimization

中文摘要

      本文提出了一种基于应用高效卷子算子(ECO)改进的LRECT跟踪算法. 首先, 为了增强网络所提取特征的 识别能力, 堆叠线性两步(LT)残差结构设计具有32层的线性两步方法性质的残差网络(LTRNet), 并且融合该网络浅 层与深层卷积特征信息形成跟踪算法的特征提取模块; 其次, 采用投影矩阵压缩LTRNet提取的高维特征, 将压缩特 征通过插值处理后, 与当前滤波器在傅里叶域进行卷积定位确定目标位置; 最后, 使用高斯牛顿算法和共轭梯度算 法求解以响应误差和惩罚项之和为优化目标的优化问题, 实现滤波器和投影矩阵的更新. 在OTB2015标准数据集 上进行测试实验, 结果表明本文所提算法可以实现较高精度的稳健性跟踪.

英文摘要

      An improved LRECT tracking algorithm based on ECO (efficient convolution operators for tracking) is proposed in this paper. Firstly, in order to enhance the extracted features for object recognition, a 32-convolution layers LTRNet (linear two-step residual network) composed of LT (linear-two step) residual structures is designed as the features extraction module of LRECT, and its shallow and deep features information are fused. Secondly, the projection matrix is used to compress the high-dimensional features extracted by LTRNet, and the compressed features are interpolated, then they are convoluted with the current filter to get the object location in the Fourier domain. Finally, filter and projection matrix are updated by Gauss-Newton and conjugate gradient methods which are applied to solve the optimization problem, and the optimization goal is the sum of response errors and penalties. The results of experiments on the OTB2015 dataset show that the proposed algorithm can achieve robust tracking with a higher accuracy.