基于深度学习的EKF和C-BMIoU目标跟踪方法  

EKF and C-BMIoU Target Tracking Methods Based on Deep Learning

作　　者：翁培欣 吴林煌[1] 苏喆 WENG Peixin;WU Linhuang;SU Zhe(School of Advanced Manufacturing,Fuzhou University,Quanzhou 362200,China)

机构地区：[1]福州大学先进制造学院,福建泉州362200

出　　处：《电视技术》2023年第12期60-64,共5页Video Engineering

摘　　要：针对卡尔曼滤波预测的目标边界框与实际对象的边界框存在较大偏差,以及目标在长期遮挡下导致行人目标跟踪丢失问题,首先采用扩展卡尔曼滤波替换ByteTrack算法中的卡尔曼滤波,并直接将目标边界框的宽度和高度作为扩展卡尔曼滤波器的待预测状态。其次,使用C-BMIoU作为跟踪器的匹配规则,使得跟踪器能根据目标遮挡程度来扩展目标跟踪的匹配空间,从而改进被遮挡目标的跟踪效果。最后通过实验验证了所提方法能够有效提高目标跟踪过程中的准确度。In view of the large deviation between the target bounding box predicted by Kalman filter and the actual object bounding box,and the problem that pedestrian target tracking is lost due to long-term occlusion of the target,the extended Kalman filter is first used to replace the Kalman filter in the ByteTrack algorithm,and The width and height of the target bounding box are directly used as the state to be predicted by the extended Kalman filter.Secondly,using C-BMIoU as the matching rule of the tracker enables the tracker to expand the matching space of target tracking according to the degree of target occlusion,thereby improving the tracking effect of occluded targets.Finally,experiments show that the proposed method can effectively improve the accuracy of target tracking.

关 键 词：深度学习 目标跟踪 扩展卡尔曼滤波 匈牙利算法 神经网络 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]