针对复杂场景下单个特征的稳健性差,以及目标存在背景干扰和目标遮挡时跟踪失败的问题,提出一种基于自适应特征融合和模型更新的相关滤波跟踪算法。该算法在核相关滤波的基础上,通过对不同特征的响应图采用平均峰值-相关能量的方法进行加权求和,实现了响应图层面的自适应特征融合。根据响应图的峰值特性计算自适应权重,以其作为置信度确定模型的更新率,进而设计自适应模型更新方法。实验结果表明,该算法能够很好地适应背景干扰、目标遮挡、旋转运动等复杂场景,与近年来优秀的相关滤波跟踪算法相比,所提算法的平均距离精度比其中最优的算法提高了2.64%,平均重叠精度提高了1.54%。

To address the poor robustness of single feature in a complex scene and tracking failure caused by background interference and object occlusion, this study proposes a correlation filter tracking algorithm that combines adaptive feature fusion and adaptive model update. Based on kernel correlation filtering, the proposed algorithm performs weighted summation on the response maps of different features by adopting the average peak-correlation energy method to realize adaptive feature fusion of response maps. The adaptive weight is calculated as the confidence according to the peak characteristics of the response maps to determine the update rate of the model,thereby realizing the design of an adaptive model updating method. Experimental results demonstrate that the algorithm can adapt to complex scene changes, such as background disturbance, object occlusion, and rotational motion. Compared to popular correlation filtering tracking algorithms, the proposed algorithm increases the average distance and overlapping precision by 2.64% and 1.54%, respectively.