| 具有空间调整和稀疏约束的相关滤波跟踪算法 |
| |
| 引用本文： | 田 丹,臧守雨,涂斌斌. 具有空间调整和稀疏约束的相关滤波跟踪算法[J]. 图学学报, 2021, 42(5): 755-761. DOI: 10.11996/JG.j.2095-302X.2021050755 |
| |
| 作者姓名： | 田 丹 臧守雨 涂斌斌 |
| |
| 作者单位： | 沈阳大学信息工程学院,辽宁 沈阳 110044 |
| |
| 基金项目： | 国家自然科学基金项目(61703285)；辽宁省自然科学基金项目(2019-MS-237)；辽宁省博士科研启动基金计划项目(2020-BS-263) |
| |
| 摘 要： | 因频繁遮挡、尺度变化、边界效应等因素的影响，进行目标跟踪时，时常难以达到较好的预期效果。再有，采用传统特征提取策略也会影响目标跟踪的鲁棒性。针对上述问题，提出一种具有空间调整和稀疏约束的相关跟踪算法。利用传统特征与深度特征的有效融合，适应目标表观变化；基于峰值旁瓣比判别目标在跟踪过程中是否被遮挡，若发生遮挡，则对滤波器进行稀疏正则化约束，提高模型对遮挡问题的鲁棒性；若未发生遮挡，则通过高斯空间调整惩罚滤波器系数，抑制边界效应的影响。实验利用 OTB 数据集中 5 组涵盖了严重遮挡和尺度变化等挑战因素的标准视频序列进行测试，定性和定量对比了算法与 4 种热点算法的跟踪效果。定性分析中基于视频序列的主要挑战因素进行比较，定量分析通过中心点位置误差和重叠率比较跟踪算法的性能。实验结果表明，算法对上述挑战因素更具鲁棒性。
|
| 关 键 词： | 目标跟踪 相关滤波 深度特征 稀疏约束 空间调整 |
Correlation filter tracking with spatial regularization and sparse constraints |
| |
| TIAN Dan,ZANG Shou-yu,TU Bin-bin. Correlation filter tracking with spatial regularization and sparse constraints[J]. Journal of Graphics, 2021, 42(5): 755-761. DOI: 10.11996/JG.j.2095-302X.2021050755 |
| |
| Authors: | TIAN Dan ZANG Shou-yu TU Bin-bin |
| |
| Affiliation: | School of Information Engineering, Shenyang University, Shenyang Liaoning 110044, China |
| |
| Abstract: | Due to the influence of frequent occlusion, scale variation, boundary effect and other factors, it is oftendifficult to achieve the desired results in target tracking. At the same time, the traditional feature extraction strategyaffects the robustness of target tracking. To address the above problems, we proposed a correlation filter trackingalgorithm with spatial regularization and sparse constraints, which utilized the effective fusion of traditional featuresand deep features to adapt to the changes of the target appearance. Based on the peak side lobe ratio, a judgment wasmade on whether the target is occluded in the tracking process. If occlusion occurs, sparse constraints are applied tothe filter for the improvement of robustness against the occlusion problem. Otherwise, the filter coefficients areadjusted in Gaussian space to suppress the influence of boundary effect. Five sets of standard video sequences in OTBdatasets including severe occlusion and scale change, were used to test the tracking performance of the proposedalgorithm, and four hot spot algorithms were compared qualitatively and quantitatively. In qualitative analysis, themain challenges of video sequences were compared. In quantitative analysis, the performance of tracking algorithmwas compared by center point position error and overlap success rate. Experimental results show that the proposedalgorithm is more robust to the above-mentioned challenges. |
| |
| Keywords: | target tracking correlation filtering deep feature sparse constraint space regularization |
| 本文献已被 万方数据 等数据库收录！ |
| 点击此处可从《图学学报》浏览原始摘要信息 |
|
点击此处可从《图学学报》下载免费的PDF全文 |
|
