基于卷积神经网络和核相关滤波的物体跟踪算法

Object tracking based on convolutional neural network and kernelized correlation filter

针对核相关滤波算法(KCF)在物体跟踪过程中因难以适应物体尺度变化、非刚性形变,容易出现跟踪丢失或者混淆的现象,提出一种结合卷积神经网络(CNN)深度模型和核相关滤波来进行跟踪的方法。首先,利用线下大量样本训练得到一个卷积神经网络分类模型;其次,目标跟踪过程中在卷积神经网络模型提取到的特征空间中用核相关滤波训练得到一个回归模型;最后,利用得到的回归模型对下一帧中的目标位置进行预测。理论上证明了在卷积神经网络特征空间进行核相关滤波操作的合理性,并通过跟踪对比实验表明相比利用原始像素或者方向梯度直方图(HOG)特征,新的跟踪算法在背景复杂情况下的平均预测偏移降低为原来的34.25%,并且通过利用GPU加速能达到实时的效果。

Object tracking with Kernelized Correlation Filter（ KCF） is difficult to adapt to target scale change and nonrigid deformation, which is prone to result in object missing or confusion during target tracking. In order to overcome this shortcoming, a new method based on Convolutional Neural Network（ CNN） and KCF was proposed. A neural network classification model was trained by a large amount of training samples off-line, and a ridge regression to predict the position of target in the next frame during the process of target tracking was obtained in the feature space extracted by the convolutional model. It is proved theoretically that it is reasonable to carry out kernelized correlation in the feature space of convolutional network and the final experiments show that the average prediction bias of the new algorithm is reduced to 34. 25% compared with the KCF in original pixel or Histogram of Oriented Gradients（ HOG） feature, which can achieve real-time results by acceleration of GPU.