提出了一种基于深度频谱卷积神经网络的视觉目标跟踪算法。该算法在深度模型训练阶段采用谱池化替代深度卷积神经网络中的最大池化过程, 用贝叶斯分类器替代softmax损失层计算最大分类值, 并将其整合到深度神经网络跟踪框架中, 通过新网络计算输入正负样本的概率分布预测目标位置。该算法充分利用谱池化在频域下降维到任意维度且计算高效的优点, 克服了最大池化采样造成大量空间信息丢失的不足, 提升了计算速度。在权威多场景视频标准测试库上对所提算法进行验证, 结果验证了该算法兼顾了效率和跟踪精度, 有效提高跟踪器的性能, 在相同测试条件下, 文中算法性能优于同类对比算法。

The visual target tracking algorithm based on deep learning spectrum convolutional neural networks was presented. The spectral pooling was adopted instead of max pooling in the deep convolutional neural network, then the softmax loss layer was replaced with Bayesian theorem to compute maximum classifier score, and integrated it into the deep neural network tracking framework. The location of the target can be obtained by calculating the probability distribution of the input samples. The advantages of feature dimension reduction at random with spectral pooling and computation efficiency was taken to avoid much spatial information lost, which also helped to improve the computation speed. Compared with the original algorithm and other state-of-the-art methods, the proposed tracking method shows excellent performances on test baseline dataset.