基于堆栈式消噪自编码机的分块目标跟踪

在视觉目标跟踪系统中, 特征的表达和提取是重要的组成部分. 本文提出基于多个自编码机网络相联合的特征提取机, 通过对输入数据进行一定程度的重组, 采用深度学习的理论对其局部特征进行描述并对结果进行联合决策. 结合该网络结构, 本文提出一种融合局部特征的深度信息进行目标跟踪的算法. 将输入图像分块使得大量的乘法运算转化为加法和乘法的混合运算, 相对于全局的特征表达, 大幅降低了运算复杂度. 在跟踪过程中, 目标候选区的各分块权重能够根据相应网络的置信度进行自适应的调整, 提升了跟踪器对光照变化、目标姿态和遮挡的适应. 实验表明, 该跟踪算法在鲁棒性和跟踪速度上表现优秀.

Local patch tracking algorithm based on stacked denoising autoencoder

The expression and extraction of the feature plays the most important role in a visual tracking system. Based on the theory of deep learning, we propose a feature extractor based on multiple ensemble autoencoders which can decide the result by jointly describing the data input. Based the proposed network architecture, a novel tracking method applying deep features of various local patches is established. The process of breaking the input images into patches decreases the calculation complexity from amounts of multiplications to the combination of relatively less multiplications and some additions, thus reducing the time complexity. In the tracking process, the weights of different patches change according to the reliability of the corresponding ones, which improves the robustness of the tracker to conduct some challenging situations, such as light change, target posture change and occlusion. Experiments on an open tracking benchmark show that both the robustness and the timeliness of the proposed tracker are promising.