Abstract:

Aiming at the shortcomings of traditional manual features in target tracking algorithm based on kernel correlation filtering, this paper takes target tracking technology based on kernel correlation filtering as the research object, uses deep convolution neural network to automatically extract deep convolution features of target to be tracked to replace traditional manual features. The deep convolution feature extracted from different convolution layers is separately learned by the kernel correlation filter to obtain different feature maps, and the position of the target to be tracked in the video sequence is determined by weighted fusion of multiple feature maps, it improves the robustness of the tracking algorithm in complex interference background.

Key words: video target tracking, kernel correlation filtering, deep learning, deep feature

摘要：

针对核相关滤波目标跟踪算法中传统手工特征的不足，以核相关滤波方法的目标跟踪技术作为研究对象，利用深度卷积神经网络自动提取待跟踪目标的深度卷积特征，来代替传统的手工特征，利用从不同卷积层提取到的深度卷积特征分别经过核相关滤波器学习来得到不同的特征图，然后对多个特征图进行加权融合来确定待跟踪目标在视频序列中的位置，以此来提高跟踪算法在复杂干扰背景下的鲁棒性。

关键词: 视频目标跟踪, 核相关滤波, 深度学习, 深度特征