自适应时空正则化的相关滤波目标跟踪

Correlation filter tracking for adaptive spatiotemporal regularization

针对正则化滤波器预先定义正则化项,但无法实时抑制非目标区域学习的缺点,提出了一种自适应时空正则化的新方法,从而提高算法在目标跟踪过程中适应外观变化的鲁棒性。首先在目标函数中引入空间局部响应变化量,使滤波器专注于学习对象中值得信任的部分,从而得到响应模型;其次根据全局响应变化决定滤波器的更新率;最后引入卷积神经网络进行深度特征提取,为减少高维数据存储过大,采用主成分分析算法进行降维处理,既保留主要特征又加快计算速度。在数据集OTB2013和OTB2015上的平均精确率和平均成功率相较于时空正则化相关滤波器算法分别提高了4.7%和12.7%。大量实验证明,该算法在复杂背景、物体遮挡、快速运动等多种场景下基本满足实时性需求。

The regularization filter defines the regularization term in advance,but it can not suppress the learning of nontarget region in real time.Aiming at this shortcoming,a new adaptive spatiotemporal regularization method is proposed to improve robustness of the algorithm to adapt to appearance change in the process of target tracking.The method firstly introduces the spatial local response variation into the objective function,so that the filter can focus on the trusted part of the learning object,obtaining the response model.Then,the update rate of the filter is determined based on the change of global response.Finally,the convolutional neural network is introduced for deep feature extraction.In order to reduce the storage capacity of high-dimensional data,the principal component analysis algorithm is employed to reduce dimension,which not only retains the main features,but also speeds up the calculation.Comparing with the spatialtemporal regularized correlation filtering algorithms,the average accuracy and success rate of data sets OTB2013 and OTB2015 are respectively increased by 4.7% and 12.7%.A large number of experiments show that the algorithm basically meets the real-time requirements in complex background,object occlusion,fast motion and other scenes.