基于改进空间正则化相关滤波器的运动目标跟踪研究

为降低光照、遮挡、尺寸变化等因素对目标跟踪过程的影响,提出在经典空间正则化核相关滤波基础上,增加遮挡检测及处理机制,并分别采用相似度计算与空间距离计算作为遮挡评判标准。在模型更新之前先判断遮挡是否存在,如有遮挡,则不更新模型;否则更新模型。提出搜索半径择优处理,分别以6种搜索半径进行目标跟踪,寻找最优搜索半径;进而提出特征择优处理,分别提取HOG特征、PHOG特征、Haar-like特征、LBP特征以及FHOG特征与改进算法结合,选取最佳特征。采用两组实验进行验证:分别采用经典KCF算法、Mean Shift算法、Fragment算法、DSST算法、经典SRDCF算法和改进SRDCF算法对Bolt2和Basketball两个视频中运动目标进行跟踪对比。实验结果表明:FHOG特征与改进空间正则化核相关滤波相结合,且在搜索半径为8个像素点时的跟踪性能最佳,优于其他经典跟踪算法,处理速度可达3. 7 fps。

Research on Target Tracking Based on Improved Spatial Regularization Correlation Filter

In order to reduce the poor effect of illumination,occlusion,size change and other factors on the target tracking process,it was proposed to increase the occlusion detection and processing mechanism on the basis of classical spatial regularization correlation filtering.The similarity calculation and spatial distance calculation were used as the criterion of occlusion detection.If the occlusion existed,the occlusion process was performed,that is,the update of the model was stopped;otherwise,the model was updated.Secondly,the radius selection was proposed,and the target tracking was carried out with six search radius,and the optimal search radius was found.Then,it was proposed the feature selection method,and the feature of HOG,PHOG,Haar,LBP and FHOG were combined with the algorithm to select the best features.Two groups of experiments were used for verification:classical KCF algorithm,Mean Shift algorithm,Fragment algorithm,DSST algorithm,classical SRDCF algorithm and improved SRDCF algorithm were used to track and compare moving targets in Bolt2 and Basketball.The experimental results show that FHOG feature combined with the improved spatial regularized correlation filtering has the best tracking performance when the search radius is 8,and it is superior to other classical tracking algorithms.The processing speed can reach 3.7 fps.