Event-based semantic segmentation of space targets with SNN-HRNet

Le résumé fourni par la source

Space-target perception is essential for rendezvous, on-orbit servicing, and debris removal. To improve foreground-background segmentation in event-based space scenes, we propose an SNN-HRNet framework for space-target semantic segmentation. We use an SNN-based HRNet backbone to extract multi-scale spatiotemporal features while preserving high-resolution representations, and introduce an LIF-based membrane-potential integration and temporal averaging module to bridge multi-step spiking features and continuous-domain prediction. A CNN prediction head is then used to generate binary segmentation results. To evaluate the proposed method, we construct a simulated dataset in Unreal Engine 5 and generate event streams with v2e. Experimental results show that our method achieves an mIoU of 0.988 on the test set, outperforming an SNN model with a U-Net backbone, which achieves 0.978. Qualitative results further show that the proposed method produces more accurate boundaries and better preserves slender target structures. These results demonstrate the effectiveness of combining high-resolution multi-scale representation learning with spiking temporal modeling for event-based space-target segmentation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.