一种易于初始化的类卷积神经网络视觉跟踪算法

An Easily Initialized Visual Tracking Algorithm Based on Similar Structure for Convolutional Neural Network

该文针对视觉跟踪中运动目标的鲁棒性跟踪问题,基于主成分分析(PCA)和卷积神经网络(CNN),提出一种易于初始化的类CNN提取深度特征的视觉跟踪算法。该算法首先利用仿射变换对原始图像进行处理,然后对归一化尺寸的图像进行分层PCA学习,将学习得到的PCA特征向量作为CNN结构中的各阶滤波器,完成特征提取网络的初始化,再利用特征提取网络获取目标的深层次表达。最后结合粒子滤波,利用一个简单的逻辑回归分类器通过分类估计实现目标跟踪。结果表明,利用这种易于初始化的CNN提取到的深度特征能够有效地区分目标和背景,具有很好的可区分性,提出的视觉跟踪算法对光照变化、尺度变化、遮挡、旋转和摄像机抖动等都具有良好的适应性,在许多视频序列上表现出了较好的鲁棒性和准确性。

On the issues about the robustness in visual object tracking, based on Principal Component Analysis（PCA） and Convolutional Neural Network（CNN）, a novel visual tracking algorithm with deep feature, which is acquired from a easily initialized CNN structure, is proposed. First, the original image is processed by affine transformation. Next, layered PCA learning is used to process the normalized size image, the eigenvectors learned by PCA are used to be the filters of a CNN structure to realize initialization. Then, the deep expression of the object is extracted by this CNN structure. Finally, combining particle filter algorithm, a simple logistic regression classifier is used to realize target tracking. The result shows that the deep feature acquired from the easily initialized CNN structure has a better expressivity, it can distinguish the object and background effectively. The proposed algorithm has a better inflexibility to illumination, occlusion, rotation and camera shake, and it exhibits a good robustness and accuracy in many video sequences.