Multi-Temporal Feature Integration: Learning Complementary Motion and Semantic Representations for Gait Recognition
Open Access
- 12 March 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 41529-41540
Abstract
Gait recognition is a promising biometric modality for the non-invasive, long-distance identification of individuals based on their unique walking patterns. Despite its potential advantages, real-world gait recognition is hindered by the unconstrained nature of gait acquisition, where factors such as camera viewpoint, clothing, and carried objects can distort body shape and gait dynamics. While most existing methods rely on 3D convolutions to extract spatiotemporal features, they typically operate at a fixed temporal resolution (TR), which limits their ability to jointly model fine-grained motion details and high-level spatial semantics. To overcome this limitation, we propose a novel method comprising two modules with a dual TR gait recognition strategy that integrates complementary features extracted from high TR (HTR) and low TR (LTR) video streams. The HTR feature extraction module captures subtle motion dynamics, while the LTR feature extraction module preserves rich spatial and semantic shape details. Additionally, an effective lateral fusion mechanism is employed to exploit multi-scale temporal information, enabling more discriminative feature representation for gait recognition. By fusing both features, the proposed method demonstrates strong robustness across diverse covariate conditions, including camera-view variations, carried objects, clothing variations, and real-world scenarios. Extensive experiments on four publicly available benchmark datasets (CASIA-B, CCPG, OULP, and Gait3D) validate the effectiveness of our proposed method, which achieves superior accuracy. On the CASIA-B dataset, the proposed method achieves average rank-1 accuracies of 98.9%, 96.2%, and 86.7% for the normal walking, carried object, and clothing variation covariate conditions, respectively. Our method also achieves average rank-1 accuracy on each dataset of 81.2% for CCPG, 99.1% for OULP, and 54.2% for Gait3D. The source code will be available at https://github.com/MDROKONMIA/HTR_LTR/.Keywords
Funding Information
- University Grant Commission, Bangladesh (37.01.0000.052.20.015.20-1493(30))
This publication has 50 references indexed in Scilit:
- Cross-view gait recognition based on discriminative global–local feature representation and learningEngineering Applications of Artificial Intelligence, 2025
- GaitSCM: Causal representation learning for gait recognitionComputer Vision and Image Understanding, 2024
- A spatiotemporal energy model based on spiking neurons for human motion perceptionCognitive Neurodynamics, 2024
- GaitAMR: Cross-view gait recognition via aggregated multi-feature representationInformation Sciences, 2023
- PoseMapGait: A model-based gait recognition method with pose estimation maps and graph convolutional networksNeurocomputing, 2022
- Gait Lateral Network: Learning Discriminative and Compact Representations for Gait Recognition2020
- Pose-Based Temporal-Spatial Network (PTSN) for Gait Recognition with Carrying and Clothing Variations2017
- Effective part-based gait identification using frequency-domain gait entropy featuresMultimedia Tools and Applications, 2013
- Gait flow image: A silhouette-based gait representation for human identificationPattern Recognition, 2011
- Chrono-Gait Image: A Novel Temporal Template for Gait Recognition2010