摘要
高效安全的多目标跟踪技术是智能汽车行驶过程中的重要环节,然而目前许多方法忽略了误检目标可能对行驶安全性造成的潜在影响。为了减少误检目标的出现,提出了一种基于多传感器融合的双重关联机制,首先将轨迹与点云域和图像域中同时检测到的目标相关联并使用卡尔曼滤波进行更新,其次将未关联的轨迹与仅出现在点云域中的目标相关联,其中第一步未关联的目标定义为新轨迹,而第二步未关联的目标删除,所提方法可以极大地减少智能车辆行驶过程中误检目标的出现,从而显著提升行驶的安全性。同时,针对一些采用非线性卡尔曼滤波器的方法中在转弯过程中目标框偏移的问题,提出了一种改进的容积卡尔曼滤波器。该方法利用IMU数据来判断车辆的行驶状态,并自适应地调整估计误差矩阵,有效消除了车辆转弯对目标行驶状态估计的负面影响。在Kitti多目标跟踪数据集上进行测试的结果显示,所提算法有很高的优越性,HOTA(High Object Track Accuracy)达到78.00,MOTA(Multi-Object Track Accuracy)达到88.85,FPS达到200,在保持高精度的同时能很好满足实时性要求。
Efficient and Secure Multi-Object Tracking Technology is a crucial part of intelligent driving in smart vehicles.However,many existing methods overly focus on improving tracking precision while neglecting the potential impact of false positive detections on driving safety.To enhance the safety performance of intelligent vehicles,a dual-association mechanism based on multi-sensor fusion is proposed.Firstly,the proposed approach associates trajectories with simultaneously detected objects in both point cloud and image domains,and Kalman filtering is used for updates.Secondly,unassociated trajectories are linked with objects detected only in the point cloud domain.Among them,the first-step unassociated detections are considered new trajectories,while the second-step unassociated detections are removed.This method significantly reduces false positive detections during the intelligent vehicles operation,thus substantially improving driving safety.Furthermore,the issue of object bounding box deviation during turns in some methods that employ nonlinear Kalman filters is addressed.An improved Cubature Kalman filter is proposed,which utilizes IMU data to determine the vehicles driving status and adaptively adjust the estimation error matrix.This effectively eliminates the negative impact of vehicle turns on the estimation of object states.The algorithms performance is evaluated by using the Kitti multi-object tracking dataset.The results demonstrate its superiority,achieving a HOTA score of 78.00,MOTA score of 88.85,and FPS of 200,satisfying real-time requirements effectively.
作者
刘德儿
程健康
刘峻廷
LIU Deer;CHENG Jiankang;LIU Junting(Jiangxi University of Science and Technology,The School of Civil Engineering and Surveying&Mapping Engineering,Ganzhou Jiangxi 341400,China)
出处
《传感技术学报》
北大核心
2025年第7期1253-1261,共9页
Chinese Journal of Sensors and Actuators
基金
国家自然科学基金面上项目(4271434)
江西省自然科学基金面上项目(20202BAB202025)。
