LiDAR point clouds are sparse, irregular, and unevenly distributed, which makes representative feature extraction challenging for 3D object detection. Currently, Mamba modules have been increasingly applied to 3D object detection due to their ability to efficiently model global spatial dependencies. However, preserving geometric structures while modeling global spatial dependencies remains challenging in Mamba-based frameworks. In view of this, this paper proposes a single-stage 3D object detection framework for LiDAR-based point clouds. First, this paper designs a hierarchical multiscale structure called Multiscale Voxel–Point Alternating Fusion (MVPF) Module. Within this module, the 2DMamba module is introduced into the feature extraction stage at each scale to model global spatial dependencies in the BEV plane through selective state-space scanning. Second, we design a Z-to-Channel cross-dimensional reorganization strategy that merges the voxel Z dimension into the channel dimension, yielding a BEV-form feature representation suitable for 2DMamba processing. Third, this paper proposes a Local Voxel Feature Enhancement (LVFE) module composed of a Point-to-Voxel Feature Aggregation (PVFA) module, a Voxel Densification Module (VDM), and a coordinate-indexed Voxel-to-Point Mapping (VPM) module, which enhances local voxel representations while maintaining point–voxel spatial correspondence. On the KITTI validation set, Mamba-BEV achieved higher detection accuracy and inference speed than the baseline. Under the Moderate difficulty level of AP_R11, the 3D AP for Car, Pedestrian, and Cyclist improved by 0.6%, 1.5%, and 0.4%, respectively, compared to the baseline, while the inference speed increased from 44.05 FPS to 67.11 FPS. These results demonstrate the effectiveness of the proposed Mamba-BEV for point cloud-based 3D object detection.

To effectively address the challenge of balancing detection accuracy and efficiency in helmet-wearing detection, this paper proposed an improved YOLOv5 framework called SDIM-YOLOv5. Firstly, the CSPDarknet53 in the YOLOv5s benchmark structure was replaced by SimRepCSP to reduce model size. Secondly, the DualConv module was introduced into the neck network of YOLOv5s to construct the C3_DualConv modules. Thirdly, MPDIoU is combined with InnerIoU to construct the InnerMPDIoU loss function for bounding-box regression. Experimental results on the custom dataset show that our proposed model SDIM-YOLOv5 could achieve about 95.5% recall and 97.2% mAP@0.5, respectively, outperforming the YOLOv5s baseline. Meanwhile, SDIM-YOLOv5 only contains 5.89 M parameters with a smaller model size of 12.1 MB compared with the YOLOv5s model. Furthermore, compared with lightweight YOLO models with comparable model complexity, SDIM-YOLOv5 demonstrates competitive detection performance. Finally, deployment experiments on an embedded device demonstrate the feasibility of SDIM-YOLOv5 for edge-based helmet-wearing detection.