光学学报, 2020, 40 (3): 0315001, 网络出版: 2020-02-17
自适应特征融合的多尺度核相关滤波目标跟踪 下载: 1354次
Multi-Scale Kernel Correlation Filter Algorithm for Visual Tracking Based on the Fusion of Adaptive Features
机器视觉 计算机视觉 目标跟踪 核相关滤波 自适应特征融合 多尺度估计 machine vision computer vision target tracking kernel correlation filter adaptive features fusion multi-scale estimation
摘要
为了提高复杂场景中目标跟踪的稳健性,解决由光照变化、目标形变、尺度变化和遮挡等导致的目标跟踪失败问题,提出一种自适应特征融合的多尺度核相关滤波目标跟踪算法。该算法首先通过2种不同的特征分别训练2个核相关滤波器,利用这2个滤波器响应的峰值旁瓣比和相邻两帧的响应一致性获得融合权重,同时采用自适应加权的融合策略将这2个滤波器的响应结果进行融合,完成目标的位置估计;然后以此为中心进行多尺度采样,构建尺度金字塔,并通过贝叶斯估计的方法确定目标的最优尺度;最后依据目标跟踪的置信度进行跟踪模型更新,以避免模型退化。选取51组视频序列进行测试,并与近年来性能优异的目标跟踪算法进行对比。实验结果表明,所提算法能有效降低光照变化、目标形变、尺度变化和遮挡等因素影响,对测试视频序列取得了较高的跟踪精度和成功率,整体性能优于对比算法。
Abstract
In this study, we propose a multi-scale kernel correlation filter algorithm for visual tracking based on the fusion of adaptive features to promote the robustness of visual tracking in complex scenarios and tackle the tracking failure problems that can be attributed to illumination variation, target deformation, scale variation, occlusion, etc. First, two kernel correlation filters are separately trained using two different features. Then, the peak side-lobe ratio of the responses and the correlation filter response consistency of two consequent frames are considered to be the weight factors for feature fusion. Meanwhile, an adaptive strategy is adopted to fuse two responses for estimating the position. Next, multi-scale image patches are sampled to construct a scale pyramid based on the estimated position center, and the Bayesian method is employed to estimate the optimal scale of the target. Finally, the tracking model is updated according to the confidence of the tracking result to prevent the deterioration of the model. 51 video sequences are selected for conducting tracking evaluation, and the visual tracking algorithms that exhibited excellent performances in recent years are compared with our proposed algorithm. The experimental results demonstrate that the proposed algorithm effectively reduces the interferences, including the illumination variation, target deformation, scale variation, and occlusion. High tracking accuracy and success rate can be achieved using the aforementioned sequences, and the overall performance of our algorithm is observed to be better than those of the comparison algorithms.
陈法领, 丁庆海, 常铮, 陈宏宇, 罗海波, 惠斌, 刘云鹏. 自适应特征融合的多尺度核相关滤波目标跟踪[J]. 光学学报, 2020, 40(3): 0315001. Faling Chen, Qinghai Ding, Zheng Chang, Hongyu Chen, Haibo Luo, Bin Hui, Yunpeng Liu. Multi-Scale Kernel Correlation Filter Algorithm for Visual Tracking Based on the Fusion of Adaptive Features[J]. Acta Optica Sinica, 2020, 40(3): 0315001.