为了减少有效卷积算子(ECO)跟踪算法的特征提取网络参数量和计算量, 采用了一种基于端侧神经网络(GhostNet)改进的ECO目标跟踪算法。首先, 采用GhostNet网络作为主干特征提取网络提取图像浅层与深层的卷积特征, 运用全局平均池化对卷积特征下采样增加特征对图像的表征能力; 其次, 将卷积特征与手工特征插值后, 与当前滤波器在傅里叶域进行卷积计算实现目标定位; 最后, 采用共轭梯度算法优化响应误差与惩罚项之和的损失函数实现滤波器更新。在上述提出的算法和OTB2015与VOT2018数据集上进行了理论分析和实验验证, 取得了目标跟踪的对比实验数据。结果表明, 相对于基于ResNet特征提取网络的ECO算法, 该算法在实现高精度跟踪时, 卷积特征提取过程计算量减少了95.75%, 参数量减少了79.69%, 跟踪过程速度提升了160%。这些结果为轻量级目标跟踪算法的研究提供了参考。

In order to reduce the amount of feature extraction network parameters and computation of effective convolution operator (ECO) tracking algorithm, the improved eco target tracking algorithm based on GhostNet was adopted. Firstly, the GhostNet network was used as the main feature extraction network to extract the convolution features of shallow and deep layers, and the global average pooling was adapted to downsampling convolution features to improve the image representation ability. Secondly, after interpolating the convolution feature with the manual feature, convolution calculation was performed with the current filter in the Fourier domain to realize the target localization. Finally, conjugate gradient algorithm was used to optimize the loss function of the sum of response error and penalty term to update the filter. Theoretical analysis and experimental verification were carried out on the proposed algorithm and OTB2015 and VOT2018 datasets, then the comparative experimental data of target tracking were obtained. The results show that compared with the ECO algorithm based on ResNet feature extraction network, the proposed algorithm can achieve higher precision tracking, the convolution feature extraction process reduces 95.75% of computation and 79.69% of parameters, and the tracking speed increases 160% at the same time. These results provide a reference for the research of lightweight target tracking algorithms.