MSF-A2: Improving 3D Object Detection with Multi-Scale Feature Integration

To address the limitations of the Part-A2 3D point cloud object detection model in multi-scale feature extraction and global feature fusion, which hinder the simultaneous capture of fine-grained and coarse-grained feature information leading to incomplete feature representation and insufficient accuracy. The model also lacks an effective mechanism to integrate local and global features, thereby affecting the accuracy of moderate and hard target detection. The MSF-A2 (Multi-Scale Feature-A2) model, based on the Part-A2 anchor-base model, by incorporating the Spatially-adaptive Feature Modulation (SAFM) module and the Large Kernel Attention Modulation (LKAM) module. These enhancements improve the model’s ability to perceive objects of different scales and types, effectively extract useful spatial and channel information, and enhance global feature extraction. The improved algorithm in this paper was compared with the benchmark network Part-A2 on the KITTI dataset. For car detection in moderate difficulty scenarios, the 3D detection accuracy improved by $2.88 \%$. For pedestrian and cyclist detection in hard difficulty scenarios, the 3D detection accuracy improved by $6.93 \%$ and $2.9 \%$, respectively. On average, the 3D accuracy in moderate and hard detection levels increased by approximately $1.51 \%$ and $3.38 \%$, respectively, demonstrating the effectiveness of the proposed algorithm. However, the experiments were conducted using only the KITTI dataset. Future work should evaluate the algorithm’s generalization ability and adaptability to more complex targets and scenes. Further optimization of the model architecture and introduction of new data augmentation methods could enhance the MSF-A2 model’s performance in a wider range of application scenarios.