基于改进数据关联的三维多目标跟踪算法

3D Multi-object tracking algorithm based on improved data association

针对传统三维多目标跟踪算法在复杂场景中出现的误关联、跟踪中断、适应性差等问题,在数据关联阶段进行了相应改进,提出了一种基于加权聚合关联代价和目标预测置信度的多目标跟踪算法。首先,结合目标的位置、外观、方向特征计算加权聚合关联代价以度量目标之间的差异性。然后,在关联代价矩阵中引入预测置信度的相关概念,并依据该置信度调整丢失目标的关联搜索域。最后,使用卡尔曼滤波器进行目标运动状态以及预测置信度的更新。在实测数据上的实验结果表明,所提出的算法能够提高点云遮挡、轨迹交叉情况下的跟踪正确率,在MOTA上达到了73.6%。

Aiming at the problems of misassociation,track interruption and poor adaptability in traditional 3D multi-object tracking algorithms in complex scenes,corresponding improvements is made in the data association stage,and a multi-object tracking algorithm based on weighted aggregation association cost and the prediction confidence of objects is proposed.Firstly,the weighted aggregation association cost is calculated by combining the location,appearance,and orientation features of the object to measure the difference between the objects.Then,the related conception of prediction confidence is introduced in the association cost matrix,and the association search domain of the missing target is adjusted according to the confidence.Finally,Kalman filter is used to update the object motion state and the prediction confidence.Experimental results on the measured data show that the proposed algorithm can improve tracking accuracy in the case of point cloud occlusion and trajectory intersection,and the MOTA reaches 73.6%.