Semantically Enhanced Multi-Object Detection and Tracking for Autonomous Vehicles.

Tao Wen,Nikolaos M. Freris
DOI: https://doi.org/10.1109/tro.2023.3299517
IF: 7.8
2023-01-01
IEEE Transactions on Robotics
Abstract:Accurate ambient perception via multi-object detection and tracking is instrumental for autonomous vehicles. This article addresses two main challenges when operating solely on 3-D light laser detection and ranging (LiDAR) point clouds: the classification of objects with similar geometric structures and tracking under the commonplace setting of low-frequency sensing. First, we design a semantically enhanced feature aggregation module that fuses features learned from two branches with different resolutions and depths. Subsequently, the extracted semantic information combined with our proposed Margin Loss allows the re-identification module to extract time-invariant geometric features. These features are fused with the positional information provided by the detector by a cluster-based Earth's mover distance algorithm along with conflation to improve the tracking stability. Extensive experiments on nuScenes demonstrate that our proposed model outperforms the state-of-the-art methods for both LiDAR-based 3-D object detection and tracking. In particular, we report an increase of 1.1% in average multi-object tracking accuracy, as well higher mean average precision for detection by 6.2% and 7.5% on motorcycle and bicycle, respectively.
What problem does this paper attempt to address?