基于相关性滤波的鲁棒视觉目标跟踪算法研究

发布时间：2018-04-12 16:35

本文选题：相关性滤波 + 接力跟踪　；参考：《华中科技大学》2016年硕士论文

【摘要】：视觉目标跟踪是计算机视觉领域的重要课题。为处理复杂跟踪场景,越来越多跟踪算法将跟踪与检测相结合。其中比较突出的是判别式相关性滤波(Discriminant Correlation Filter,DCF)跟踪算法。但此类跟踪算法大多以全局特征构建目标表观模型,忽略尺度变化,并使用单一的跟踪器对目标表观更新。在面对遮挡和较严重形变等跟踪场景时,算法的鲁棒性大大降低。针对全局特征KCF算法在目标尺度变化或遮挡时性能下降问题,提出基于显著性检测分块的多尺度多线索KCF跟踪方法(Salient Patch-based visual Tracking with Multi-cues Integration,SPMCI)。分析了低层、中层和高层结构对目标表观的影响,采用中层目标块构建表观模型。常用均匀分块方法会产生过多目标块,引入不必要背景干扰,同时增大计算量。为此使用目标显著图作为分块的先验信息,控制目标块(patch)的分布和数目。进而结合块的图像金字塔对目标尺度估计。为进一步提高跟踪算法准确性,融合目标块表观、空间分布和运动轨迹线索进行目标定位。实验表明SPMCI算法有效提高了跟踪的准确性和对不同场景的适应能力。如何处理目标长时遮挡带来的表观污染和较大形变是跟踪算法的一大难点。以SPMCI算法为基础,提出了基于表观变化检测的多跟踪器接力跟踪算法(Apparent Change Detection based Visual Tracking with Multi-trackers Relay,MTR)。MTR使用颜色直方图匹配进行表观变化检测,采用PSR(Peak to Sidelobe Ratio)指标对目标形变和遮挡进行区分。根据表观变化更新模板或对跟踪器进行替换与选择。通过不同跟踪器接力跟踪,应对目标较大形变和长时全遮挡等恶劣场景。实验表明MTR算法提高了对长时遮挡和较大形变场景的适应能力。
[Abstract]:Visual target tracking is an important subject in the field of computer vision.In order to deal with complex tracking scenarios, more and more tracking algorithms combine tracking with detection.The discriminant Correlation filter tracking algorithm is prominent.However, most of these tracking algorithms are based on global features to construct the target visual model, ignore the scale changes, and use a single tracer to update the target view.The robustness of the algorithm is greatly reduced when tracking scenes such as occlusion and severe deformation are faced with.Aiming at the performance degradation of global feature KCF algorithm when the target scale changes or shading, a multi-scale and multi-clue KCF tracking method based on salience detection block is proposed.The influence of low level, middle layer and high level structure on the target's appearance is analyzed, and the surface model is constructed by using the middle level target block.The common uniform partitioning method will produce too many target blocks, introduce unnecessary background interference, and increase the computational complexity at the same time.In order to control the distribution and number of target patchs, the target saliency map is used as a priori information.Then the image pyramid of the block is used to estimate the target scale.In order to further improve the accuracy of the tracking algorithm, target location is carried out by fusion of target block representation, spatial distribution and motion trajectory clues.Experiments show that the SPMCI algorithm can effectively improve the accuracy of tracking and adaptability to different scenes.How to deal with the apparent pollution and large deformation caused by long time occlusion is one of the most difficult problems in tracking algorithm.Based on the SPMCI algorithm, a new relay tracking algorithm named Apparent Change Detection based Visual Tracking with Multi-trackers Relayn MTRN based on apparent change detection is proposed to detect the apparent change using color histogram matching.The target deformation and occlusion are distinguished by PSR(Peak to Sidelobe Ratio.Update templates based on apparent changes or replace and select trackers.Through the relay tracking of different tracker, it can deal with the bad scene such as large deformation of target and long time complete occlusion.Experiments show that the MTR algorithm improves the adaptability to long time occlusion and large deformation scenes.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】