基于深度特征表达与学习的视觉跟踪算法研究

摘要

For the robustness of visual object tracking, a new tracking algorithm based on multi-stage convolution filtering feature is proposed by introducing deep learning into visual tracking. The algorithm uses the Principal Component Analysis (PCA) eigenvectors obtained by stratified learning, to extract the deeper abstract expression of the original image by multi-stage convolutional filtering. Then the Bhattacharyya distance is used to evaluate the similarity among features. Finally, particle filter algorithm is combined to realize target tracking. The result shows that the feature obtained by multi-stage convolution filtering can express target better, the proposed algorithm has a better inflexibility to illumination, covering, rotation, and camera shake, and it exhibits very good robustness in video sequence with such characteristics.%该文针对视觉跟踪中运动目标的鲁棒性跟踪问题，将深度学习引入视觉跟踪领域，提出一种基于多层卷积滤波特征的目标跟踪算法。该算法利用分层学习得到的主成分分析(PCA)特征向量，对原始图像进行多层卷积滤波，从而提取出图像更深层次的抽象表达，然后利用巴氏距离进行特征相似度匹配估计，进而结合粒子滤波算法实现目标跟踪。结果表明，这种多层卷积滤波提取到的特征能够更好地表达目标，所提跟踪算法对光照变化、遮挡、异面旋转、摄像机抖动都具有很好的不变性，对平面内旋转也具有一定的不变性，在具有此类特点的视频序列上表现出非常好的鲁棒性。