In context
In 2024, 3D multi-object tracking (MOT) remained a critical perception task for autonomous driving and robotics, yet most methods optimized detection and data association separately, limiting accuracy under challenging conditions such as rapid rotation. This paper addressed that gap by proposing a joint optimization framework.
What was reported
Researchers from China University of Mining and Technology and Shanghai Jiao Tong University introduced FlowDet-Track, a 3D MOT framework that simultaneously optimizes object detection and scene flow estimation. The framework uses a shared feature extraction backbone based on PV-RCNN to output both detection and scene flow features, enabling coupled learning of single-frame detection and inter-frame association.
A key innovation is a detection-guided scene flow estimation module that filters out points from different objects during feature aggregation, reducing incorrect inter-frame associations. The authors also proposed a box transformation-based method for computing scene flow ground truth, which is more accurate than previous translation-only approaches, especially for rotating objects.
Experiments on the KITTI MOT dataset showed that for the vehicle category, the proposed method improved HOTA by 25.03% and DetA by 30.8% compared with PointTrackNet, demonstrating superior position tracking accuracy. Additional tests under extreme rotational motion confirmed the algorithm's robustness.
Why it mattered
By treating detection and association as a coupled problem rather than separate stages, this work pointed toward more reliable 3D tracking for autonomous vehicles and robots operating in dynamic, complex environments—an essential step for safe navigation and planning.
“To achieve the coupled learning of single-frame detection and inter-frame association, a 3D multi-object tracking framework is proposed based on the joint optimization of target detection and scene flow estimation, named FlowDet-Track.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2024-09-15 · “基于目标检测和场景流估计联合优化的3D多目标跟踪”