多目标跟踪用补偿-自适应卡尔曼滤波器设计

Design of Compensation-Adaptive Kalman Filtering for Online Multi Object Tracking

现有卡尔曼滤波(Kalman Filter,KF)多目标跟踪(Multi Object Tracking,MOT)方法难以应对相机运动、检测失真和非线性运动等问题。针对这些问题,引入自适应参量,包括运动相似度代价和检测置信度,用于动态调节KF的过程和观测协方差。此外,使用稀疏光流和随机采样一致性算法,估计前后帧的仿射变换矩阵,据此校正KF的预测值。综合以上两点,提出补偿-自适应卡尔曼滤波(Compensation-Adaptive Kalman Filter,C-AKF)。将其与BYTE框架结合,以实现在线的MOT。在MOT17数据集上的实验结果显示,所提方法能够实现68.9的高阶跟踪准确度和81.1的身份F-1分数,且关联速度满足实时要求。

The existing Kalman Filter(KF)based Multi-Object Tracking(MOT)methods are facing such challenges as camera motion,detection distortion and non-linear motion.To address these issues,adaptive parameters,including motion similarity cost and detection confidence,were introduced to dynamically adjust the KF process and observation covariance.Additionally,sparse optical flow and random sample consensus algorithm were used to estimate the affine transformation matrix between consecutive frames to correct the predicted values of KF.Based on the above two points,Compensation-Adaptive Kalman Filter(C-AKF)was proposed.It was combined with the BYTE framework to achieve online MOT.Experimental results on the MOT17 dataset showed that the proposed method achieved high-order tracking accuracy of 68.9 and identity F-1 score of 81.1.Moreover,its associated speed meets real-time requirements.