Multilevel Embedding and Dropout for Improved 3D Object Detection in Autonomous Driving Systems

Abstract

The balance between performance and model complexity in terms of time and storage is crucial fordeploying effective and practical autonomous driving systems. However, some recent advancements in 3D LiDAR object detection come at the cost of significant increases in computational time or additional reliance on multisensor data fusion, limiting their potential deployment. In this paper, we introduce an innovative framework aimed at harnessing the latent information within a broad range of region‐of‐interest (RoI)‐based 3D object detection architectures to enhance their performance, while adding minimal increase in their time consumption and model complexity. More specifically, we develop a multilevel embedding and dropout strategy to enhance the detection head's capability to generate predictions across different levels of RoI features. Those predictions are utilized to train the overall 3D object detector network in the training stage and vote for the output in the testing stage. In this way, multilevel information in the RoI features is deeply explored and also the robustness of the prediction is improved. We conducted extensive experiments on three widely‐used datasets: KITTI, Waymo, and ONCE, to thoroughly assess the effectiveness of our method. The results demonstrate the superiority of our approach, consistently outperforming the mainstream point‐based, voxel‐based, and point‐voxel‐based 3D detectors in mean average precision by 1.21% on average. Remarkably, this consistent improvement comes at the cost of only a 3.5 ms increase in computational time and a 2.7 MB increase in model size.