基于背景感知与自适应特征融合的相关滤波跟踪算法

Correlation Filter Tracking Based on Background-aware and Adaptive Feature Fusion

背景感知相关滤波算法能有效利用背景信息,在解决边界效应问题的同时保持较高的跟踪速度.针对背景感知相关滤波算法在尺度变化、遮挡、目标形变时容易跟踪失败的问题,提出一种基于背景感知与自适应响应融合的相关滤波算法.本文算法以波动面积比为参考指标,在响应层实现自适应特征融合,提高特征表达能力,并利用主成分分析法分别对方向梯度直方图特征和颜色名特征进行降维,提升跟踪速度.引入尺度滤波器,提升跟踪速度的同时准确估计目标尺度变化.采用根据响应置信度动态调整学习率的模型更新策略,降低模型漂移风险.实验结果表明,本文算法在OTB100数据集上跟踪精确度为0.842,跟踪速度为46.81帧每秒,相较于背景感知相关滤波算法分别提高了2.9%和28.85%.

Background-aware correlation filters tracking algorithm can effectively use the background information to solve the boundary effect while maintaining a high tracking speed.Aiming at the tracking failure of background-aware correlation filters tracking algorithm caused by scale changes,occlusion and target deformation,this paper propose a correlation filter tracking based on background-aware and adaptive feature fusion.This algorithm takes fluctuation area ratio as a reference index to achieve adaptive feature fusion in the response layer,and improve the feature expression ability.Principal Component Analysis is used to reduce the dimension of the histogram of oriented gradient feature and color name feature to improve the tracking speed.The scale filter is introduced to improve the tracking speed and estimate the target scale change accurately.The model updating strategy which adjusts learning rate dynamically according to response confidence is adopted to reduce the risk of model drift.Experimental results show that in the OTB100 evaluation benchmark the proposed algorithm has a precision of 0.842 and a tracking speed of 46.81 frames per second which is 2.9%and 28.85%higher than background-aware correlation filters tracking algorithm.