Gait Emotion Recognition via Uncertainty-oriented Class Discriminative Learning

Human gait conveys rich emotional cues, yet automated gait-based emotion recognition remains challenging due to subtle inter-class differences and large intra-class variations. These issues introduce significant sample uncertainty, which degrades model performance. In this paper, we propose Uncertainty-oriented Class Discriminative Learning (UCDL), a novel framework that explicitly models and mitigates sample uncertainty to enhance emotion classification. UCDL consists of three core components: (1) Spatial-Temporal Attention, which enables dynamic weighting of important frames and critical body joints in a stream-interaction manner; (2)Graph Contrastive Learning, which is designed to strengthen intra-class compactness and inter-class separability; and (3) Uncertainty Alleviation, which estimates uncertainty via class probability distribution and adaptively adjusts learning focus and loss weighting to mitigate the impact of uncertain samples. We further introduce a dual-stream architecture that processes posture and movement cues separately, enabling fine-grained gait pattern analysis. Extensive experiments show our method significantly outperforms state-of-the-art (SOTA) approaches on the benchmark datasets Emotion-Gait, E-Gait, and E-Walk. The source code is released athttps://github.com/one-ear/UCDLcodehttps://github.com/one-ear/UCDLcode.