中文摘要：

视觉跟踪中,高效鲁棒的特征表达是解决复杂环境下跟踪漂移问题的关键。该文针对深层网络预训练复杂费时及单网络跟踪易漂移的问题,在粒子滤波框架下,提出一种基于自适应深度稀疏网络的在线跟踪算法。该算法利用Re LU激活函数,针对不同类型目标构建了一种具有自适应选择性的深度稀疏网络结构,仅通过有限标签样本的在线训练,就可得到鲁棒的跟踪网络。实验数据表明：与当前主流的跟踪算法相比,该算法的平均跟踪成功率和精度均为最好,且与同样基于深度学习的DLT算法相比分别提高了20.64%和17.72%。在光照变化、相似背景等复杂环境下,该算法表现出了良好的鲁棒性,能够有效地解决跟踪漂移问题。

英文摘要：

In visual tracking, the efficient and robust feature representation is the key factor to solve the problem of tracking drift in complex environments. Therefore, to solve the problems of the complex and time-consuming of the pre-training process of deep neural network and the drift of the single network tracking, an online tracking method based on an adaptive deep sparse network is proposed under the tracking structure of particle filter. A deep sparse neural network architecture, which can be adaptively selected according to different types of targets, is constructed with the implementation of the Rectifier Linear Unit （ReLU） activation function. The robustness of deep tracking network can be easily achieved only through the online training of limited labeled samples. The results of experiments show that, compared with the state-of-the-art tracking algorithm, the average success ratio and precision of the proposed algorithm are both the highest, and they are raised by 20.64% and 17.72% respectively contrasted with the Deep Learning Tracker （DLT） algorithm based on deep learning. The proposed method can solve the problems of tracking drift efficiently, and shows better robustness, especially for the complex environment such as illumination changes, background clutter and so on.