Dark-Scene Neuromorphic Imaging for Human Pose Estimation

Abstract

Neuromorphic imaging sensors (event cameras) offer a promising paradigm for computational imaging and human pose estimation (HPE) under extreme illumination conditions. Nevertheless, dark-scene background activity originating from photodiode dark current and circuit thermal noise, together with hot-pixel noise, severely corrupts event streams and impedes reliable HPE in low-light scenarios. To address this issue, we propose an adaptive event denoising framework built upon a spatiotemporal Gaussian-weighted neighborhood model with a dynamic thresholding mechanism. It can effectively suppress background activity and hot-pixel noise while preserving edge and motion details critical for pose estimation. Leveraging this denoising front-end, we construct a complete dark-scene neuromorphic HPE pipeline by transferring the pre-trained MediaPipe model onto event-based time-surfaces. Quantitative and qualitative evaluations on public and self-collected datasets demonstrate that our approach outperforms state-of-the-art denoising methods with an improvement of over 20% on public benchmarks and over 30% on self-collected dark-scene data. We expect our work to pave the way toward reliable dark-scene human–robot interaction through robust neuromorphic pose estimation.