Detecting Motion Patterns Based on Energy-Based Models with Energy Distance

In the context of intelligent surveillance and rehabilitation systems, motion pattern recognition remains a challenging task due to environmental variations and motion complexity.This paper addresses the problem by proposing a hybrid energy-based model that integrates two complementary components: a statistical energy term () and a geometric energy term data ().These components are combined through a tunable coefficient to form a hybrid energy function hybrid ().A probabilistic inference mechanism based on a normalized Gibbs distribution is then used to determine the motion type.The proposed model is evaluated on four representative motion patterns Linear, Oscillatory,Spiral, and Complex extracted from the UCF101 dataset.Experimental results demonstrate that the hybrid model achieves high stability and consistency across different values.The Spiral pattern consistently dominates the probability distribution, while the Oscillatory pattern appears least frequently.These findings suggest that the model effectively captures relative energy relationships and exhibits robustness under parameter variation, making it well-suited for deployment in real-world motion analysis applications.