[Introducing a Physics-Based Uniform Particle Space for 3D Motion Estimation from Consecutive Frames] Please use updated version: https://fordatis.fraunhofer.de/handle/fordatis/514

Object motion estimation is crucial for tasks like traffic monitoring and animal behavior analysis. However, existing methods typically rely on complex workflows involving spatial projections, coordinate transformations, camera calibration, and model stacking. This increases system complexity, reduces generalization capabilities, and leads to speed estimation susceptible to cumulative front-end errors. To address this, we propose the Physics-based Uniform Particle Space speed Estimation Network (PUPSENet), which predicts objects real-world speed frame-by-frame without requiring calibration or explicit geometric reconstruction. Furthermore, as existing datasets focus solely on planar vehicle motion, we introduce the AquaMotion synthetic dataset to expand the diversity of available resources for speed estimation research. Experiments on AquaMotion and real fish datasets demonstrate that PUPSENet significantly reduces error and enhances robustness, while it achieves competitive performance on the real vehicle scenario. Its lightweight, scene-agnostic end-to-end architecture operates stably across diverse environments, providing a practical and scalable solution for real-time motion analysis.