中文摘要：

针对跟踪过程中运动目标的鲁棒性问题，提出了一种基于深度特征的跟踪算法.首先，利用仿射变换对每一帧图像进行归一化处理.然后，利用深度去噪自编码器提取归一化图像的特征.由于提取的特征维数巨大，为了提高计算效率，提出了一种高效的基于稀疏表示的降维方法，通过投影矩阵将高维特征投影到低维空间，进而结合粒子滤波方法完成目标跟踪.最后，将初始帧的目标信息融入到目标表观更新过程中，降低了跟踪过程中目标发生漂移的风险.实验结果表明，所提出的视觉跟踪算法在6段视频序列上获得了较高的准确度，能够在遮挡、光照变化、尺度变化和目标快速运动的条件下稳定地跟踪目标.

英文摘要：

To solve the robustness problem of the motion object in the tracking process,a tracking algorithm based on deep feature is proposed.First,each frame in the video is normalized by affine transformation.Then,the object feature is extracted from the normalized image by the stacked denoising autoencoder.Because of the large dimensions of deep feature,to improve the computational efficiency,an effective dimension reduction method based on sparse representation is presented.The high dimensional features are projected into the low dimensional space by the projection matrix.The object tracking is achieved by combing the particle filter algorithm.Finally,the object information of the first frame is integrated into the updating process of the object appearance to reduce the risk of object drift during the tracking process.The experimental results show that the proposed tracking algorithm exhibits a high degree of accuracy in six video sequences,and it can stably track the object under the circumstance of occlusion,illumination change,scale variation and fast motion.