融合峰旁比和帧差均值自适应模型更新的视觉跟踪

Adaptive Model Update via Fusing Peak-to-sidelobe Ratio and Mean Frame Difference for Visual Tracking

为了让相关滤波模型更加适应目标外观的变化,提高相关滤波跟踪算法的鲁棒性和实时性,根据相关滤波响应值、帧差均值和目标运动位移之间的关系,提出了一种单层卷积相关滤波实时跟踪模型的自适应学习率调整跟踪方法.该方法首先选取单个卷积层卷积特征,减少了卷积特征维度,然后使用单层卷积特征训练相关滤波分类器预测目标位置,用快速尺度预测方法估计跟踪目标的尺度,并采用稀疏的模型更新策略,提高跟踪的速度;最后利用相关滤波预测响应图的峰旁比估计预测位置的可信度,结合图像帧差均值和目标的运动位移量来评估目标的表观变化,并根据目标预测的可信度和表观变化情况自适应调整相关滤波模型更新的学习率,使模型快速学习目标的变化特征,提高了目标跟踪的精度.在OTB100数据集上对算法进行测试,实验结果表明,本文算法的平均距离精度达90.1%,优于实验中对比的9种主流算法,平均成功率值为79.2%,仅次于9种算法中的连续卷积跟踪算法,平均速度为31.8帧/秒,是连续卷积相关滤波算法的近30倍.

In order to adapt the correlation filter model to the change of the target appearance and improve the robustness and real-time performance of the correlation filter algorithm for visual tracking,an adaptive learning rate adjustment method for real-time tracking of a single-layer convolution filter is proposed,which is based on the relationship of the correlation filter response value,mean frame difference and the object movement displacement.This method selects the convolution features of a single convolution layer to train the correlation filter classifier that is used to predict the position of object,reducing the convolution feature dimension and improving the speed of visual tracking.Meanwhile it uses the fast-scale prediction method to estimate the object's scale,and adopts a sparse model update strategy.Besides,the Peak-to-Sidelobe Ratio(PSR)of convolutional response is used to estimate the credibility of the predicted location.The apparent change of the object is evaluated by combining the mean frame difference and the object movement displacement.And the learning rate of the correlation filter model update can be adjusted by these two terms adaptively,so that the change characteristics of the object can be quickly learned.The accuracy of visual tracking is improved by this method.The method is tested on the standard OTB-100 dataset.The results show that the average distance accuracy is 90.1%,which is better than the state-of-the-art algorithms in the experiment.And the average success rate is 79.2%,which is only smaller than the continuous convolution tracking algorithm(CCOT).But the average speed is 31.8 frames per second,nearly 30 times of the CCOT.