Learning Motion-Aware Representations for World-Space Speed Estimation from Consecutive Frames

Abstract

Object speed estimation is fundamental for understanding dynamic scenes, from traffic monitoring to animal behavior analysis. Existing monocular RGB video approaches typically rely on scene-specific information, such as camera calibration parameters or depth maps, which is often unavailable or unreliable in real-world applications. Further, speed estimation is often decomposed into separate learning problems, such as detection, tracking, depth estimation, geometric projection, and final speed regression, causing errors to accumulate across stages. We therefore propose the Motion-aware Speed Estimation Network (MotiSpeed), a lightweight end-to-end framework that replaces this multi-stage formulation with a single learned model and estimates object-level world-space speed directly from consecutive monocular RGB frames. Evaluating such models, however, requires benchmarks beyond structured planar motion. Existing speed estimation datasets are predominantly vehicle-centric, whereas datasets for unconstrained 3D motion remain limited. Therefore, we introduce AquaSpeed3D, a synthetic underwater dataset for speed estimation with freely moving agents, non-rigid appearance, occlusion, and dense multi-agent interactions. Experiments on AquaSpeed3D, 3DZeF20, and VS13 demonstrate that MotiSpeed achieves competitive performance across underwater and vehicle scenarios while remaining parameter-efficient.