PRFNet-REDFormer: an Improved PillarNeXt and Transformer-Based Fusion Framework for 3D Object Detection

3D object detection plays a vital role in autonomous driving and intelligent transportation systems, where robustness and accuracy under diverse environmental conditions are critical. Recent advances in transformer-based multi-modal fusion frameworks have shown promising performance by integrating radar and vision features. However, existing radar backbones often fail to fully exploit the spatial and dynamic characteristics of millimeter-wave radar data, resulting in incomplete feature representations. To address this issue, we propose PRFNet-REDFormer, a transformer-based 3D object detection framework enhanced with an improved PillarNeXt backbone. Specifically, we design a receptive-field atrous spatial pyramid pooling (RF-ASPP) module as the neck of the radar backbone to capture multi-scale contextual features, thereby enhancing discriminability and robustness. Extensive experiments on the nuScenes dataset demonstrate that our approach outperforms baseline REDFormer by 1.4% in mAP and 1.8% in NDS, while effectively reducing false and missed detections under challenging conditions such as rain and night. The results validate the effectiveness of PRFNet-REDFormer in robust 3D object detection.