Neural Computing and Applications Software
Evaluating human motion prediction: a framework for accuracy and realism
Abstract
Human Motion Prediction (HMP) is an essential research area for safety- and performance-critical applications in autonomous systems, assistive healthcare, and human–robot interaction. Predicting future human motion, especially full-body poses, remains challenging due to the intrinsic complexity of human dynamics. Although substantial progress has been achieved through advances in Deep Learning (DL), this has not been matched by equally rigorous and standardized evaluation methodologies. Existing protocols often rely on inconsistent metric choices, overemphasize frame-wise geometric accuracy, and provide limited insight into motion realism, hindering reproducible benchmarking and practically meaningful model selection. This work introduces a unified evaluation framework that jointly assesses three key criteria: (i) positional accuracy via Mean Per Joint Position Error (MPJPE), (ii) joint-rotation accuracy via Joint Angle Difference (JAD), and (iii) sequence-level motion realism through a new frequency-domain metric, Power Spectral Density Shift (PSDS). The latter quantifies the extent to which predicted motion preserves the characteristic frequency content of natural motion. To summarize multi-criteria behavior, we further propose a composite index, namely Human Motion Prediction Error (HMPE), together with a three-axis performance visualization that exposes accuracy-realism trade-offs in an interpretable manner. Extensive experiments on representative DL architectures across different datasets and frame rates show that relying on a single metric can yield misleading conclusions, including changes in apparent model ranking, and can mask accuracy-realism trade-offs. Moreover, the results highlight the limitations of widely used benchmarking methodologies and indicate the value of more diverse datasets for evaluating both accuracy and plausibility. Overall, the proposed protocol establishes a rigorous baseline for consistent comparisons and supports the development of more reliable HMP models. The framework is publicly available on GitHub at https://github.com/MichaelVanuzzo/HMP-Evaluation-Framework .