Statistical hypothesis testing for dynamic scenes recovery from low-light binary photon streams

Single-photon avalanche diode (SPAD) sensors enable low-light imaging with single-photon sensitivity, but reconstructing dynamic scenes from their binary photon streams under photon starvation remains a significant challenge. A primary difficulty lies in the reliance of many existing methods on explicit motion estimation, which can become highly unstable when derived from extremely sparse photon data. To address this, we introduce a statistical hypothesis-testing approach that reframes the initial motion detection problem. Instead of directly estimating motion parameters, our method tests the temporal stationarity of photon arrival statistics at each pixel to identify regions undergoing motion. This detection provides a robust prior for a subsequent iterative process that integrates photon compensation and motion correction to recover the scene. We validate the method by reconstructing challenging dynamic scenes from binary streams, including rigid-body rotation under ultralow light (down to 0.03 photons per pixel) and transient non-rigid deformation (balloon rupture at ∼0.82 PPP). For scenarios lacking absolute intensity ground truth, we propose a suite of three complementary evaluation metrics: mixed entropy (global), temporal correlation analysis (regional), and single-pixel photon transients (pointwise). The results demonstrate that the proposed statistical perspective offers a viable and robust alternative for dynamic scene recovery under photon starvation, providing a data-efficient and training-free pathway valuable for low-light applications in biomedical imaging and industrial inspection.