改进关联策略的三维多目标跟踪算法

3D multi-object tracking algorithm with improved association strategy

为提高基于激光雷达的三维多目标跟踪准确度,提出一种基于检测的3D多目标跟踪算法。使用深度神经网络从激光点云中获取目标的位置信息后,对目标跟踪算法中的关联策略进行了优化。首先,估计出目标的速度信息,与位置信息一同纳入观测值,在BEV视角下使用卡尔曼滤波器对目标的状态进行预测与更新;然后,基于目标与激光雷达的距离来评估目标位置的不确定度,用于修正观测模型中的协方差矩阵;最后,在马氏距离中添加对目标测量的不确定性加权项,使用匈牙利算法进行数据关联。在大规模自动驾驶数据集Nuscenes上对所提算法进行了测试,得出其跟踪精度超过了现有的基线方法。消融实验结果表明,所提出的改进措施能有效提高三维多目标跟踪的性能。

In order to enhance the accuracy of three-dimensional(3D)multi-object tracking based on lidar,a detection-based 3D multi-target tracking algorithm is proposed.After using deep neural networks to obtain target position information from lidar point clouds,the correlation strategy in the target tracking algorithm is optimized.The velocity information of the target is estimated and incorporated into the observations along with the position information.A Kalman filter is used to predict and update the target states in the bird's eye view(BEV)perspective.The uncertainty of target position is evaluated based on the distance between the target and lidar,which is used to correct the covariance matrix in the observation model.An uncertainty weighting term for target measurement is introduced into the Mahalanobis distance,and the Hungarian algorithm is utilized for the data association.The proposed algorithm was tested on the large-scale autonomous driving dataset Nuscenes,and it was found that its tracking accuracy exceeded existing baseline methods.The results of the ablation experiments indicate that the proposed improvement measures can effectively improve the performance of 3D multi-object tracking.