【摘要】 针对核相关滤波算法(KCF)难以处理目标尺度变化、旋转、遮挡等问题,本文在KCF的框架下提出了一种融合多层深度特征的抗遮挡目标跟踪算法。首先,在频域中,利用岭回归分类器训练VGG-2048上的conv3和conv6两层深度特征,分别得出置信度,将两者特征加权相连,替代原KCF的方形梯度直方图(HOG)特征,同时引入第1帧目标的残留信息,获得更为出色的位置响应输出。然后,针对遮挡问题,提出一种响应峰值判断抗遮挡机制。最后,通过双线性插值建立深度特征尺度池,解决目标尺度问题。在测试集(OTB-100)上的实验结果表明,改进后的算法能解决复杂环境下的目标跟踪问题,算法具有鲁棒性。更多还原

【Abstract】 In view of the difficulty of dealing with the scale change, rotation and occlusion of the target in the kernel correlation filter(KCF), an anti-occlusion target tracking algorithm based on multi-layer deep features is proposed in this paper under the framework of KCF. Firstly, in frequency domain, ridge regression classifier is used to train conv3 and conv6 deep features on VGG-2048, and the confidence is obtained. The two features are weighted together to replace the original KCF’s HOG features. At the same time, residual information of the first frame is introduced to obtain better position response output. Then, aiming at the occlusion problem, a response peak judgment anti-occlusion mechanism is proposed. Finally, the depth feature scale pool is established by bilinear interpolation to solve the target scale problem. The experimental results on test set(OTB-100) show that the improved algorithm can solve the target tracking problem in complex environment, and the algorithm has robustness.更多还原