Broadband Over-the-Air Voxel Fusion for Integrated Sensing and Edge AI

The emerging integrated sensing and edge artificial intelligence paradigm commits to enabling high-precision environment perception applications (e.g., autonomous driving) in edge networks. Fusing features from distributed sensors, an operation critical to perception, creates a multi-access bottleneck against hostile wireless channels. In this paper, we propose a task-oriented fusion framework for sparse voxel features, termed over-the-air voxel fusion (AirFusion). AirFusion is distinct from existing task-oriented edge inference techniques in exploiting the diversified sparsity across voxel features. This allows intelligently pairing each voxel with a matching subcarrier of the broadband channel, which is the key principle of AirFusion, to achieve boosted communication efficiency. Specifically, a voxel-carrier pairing and power allocation (VoCa-PPA) algorithm is developed to maximize the minimum receive SNR among voxels subject to instantaneous power constraints. VoCa-PPA incorporates optimal power allocation in closed-form into low-complexity greedy pairing. Experimental results on the OPV2V dataset show 98.4% error reduction and 5.74 × latency reduction compared against sparsity-agnostic and digital multi-access benchmarking schemes, respectively.