Multi-Scale Graph Convolutional Feature Fusion Network for Human Pose Estimation with Event Cameras

Event cameras exhibit superior performance in human pose estimation owing to their high temporal resolution, wide dynamic range, and rapid response. However, most of the existing event-based methods rely on the frame-based representations, which could not fully take advantage of the sparsity and high temporal resolution of asynchronous event data, hindering the real-time and low-power deployment. To address this issue, this paper introduces the cross attention framework to progressively fuse multi-scale graph features with a latent learnable query, where a lightweight off-the-shelf event-based dynamic graph convolutional network (EDGCN) is adopted as the backbone network to efficiently extract multi-scale graph features from voxelized event streams. The experimental results show that our method achieves AP scores of 85.96% on DHP19 and 71.04% on CDEHP, striking an excellent trade-off between performance and computational efficiency. These results highlight the potential of the proposed method for real-time and low-power applications. Overall, the method advances event-based human pose estimation by effectively leveraging sparsity and temporal dynamics of event data, paving the way for broader applications in resource-constrained environments.