EAEPnet: multi-level representation enhancement and weighted box fusion network for multi-modal 3D object detection

Abstract

With the rapid advancement of intelligent driving, LiDAR and cameras provide complementary geometric and semantic information. Therefore, fusing these two sensors is a mainstream approach for 3D object detection. However, existing methods still face two key challenges: effectively exploiting complementary information among candidate boxes during post-processing and balancing detection accuracy with computational efficiency. To address these challenges, we first propose an aggregated Euclidean distance weighted box fusion (AED-WBF) method, which aggregates complementary information from multiple candidate boxes during post-processing to improve bounding-box selection and localization accuracy. We further develop a hybrid deformable half-conv (HDHC) module that jointly enhances global and local feature representations through hierarchical offset prediction and local neighborhood attention. By integrating half-conv with a separable self-attention mechanism, HDHC reduces computational complexity while maintaining detection accuracy. Based on AED-WBF and HDHC, we construct EAEPNet, an efficient multilevel LiDAR–camera fusion network for 3D object detection. Extensive experiments are conducted on the KITTI and nuScenes datasets. On the KITTI test set, EAEPNet improves the mean average precision (mAP) by 2.73% over the baseline network. On nuScenes, EAEPNet achieves a mAP of 72.5% and an nuScenes detection score (NDS) of 74.4% on the validation set, as well as a mAP of 73.2% and an NDS of 75.3% on the test set. These results validate the effectiveness of EAEPNet in multi-sensor 3D object detection and spatial measurement. Its strong performance across multiple datasets further demonstrates its potential for intelligent driving and real-time high-precision spatial measurement. The code is available at: https://github.com/juanmao73/EAEPNet.