Voxel-wise Attention Fusion: A Novel Methodology for 3D Occupancy Networks in Autonomous Driving

The perception of the 3D environment is pivotal for ensuring safe autonomous driving. While traditional methods for 3D object detection excel in object localization, they often overlook crucial geometric and environmental contexts. To address this challenge, 3D semantic occupancy tasks have been developed. The key to this technology is the effective fusion of multiple sensors to help overcome the individual limitations of each sensor through enhanced connectivity. This study introduces a novel Voxel-wise Attention approach, incorporating learnable parameters to discern the nuanced differences between sensor modalities, thereby refining the fusion process. Our method has demonstrated a significant performance improvement, particularly in recognizing specific categories, with a 28% increase and an overall performance boost of 3.48% on the OpenOccupancy dataset. These enhancements highlight the potential of the self-attention module to develop multimodal perception algorithms in autonomous driving further.