Article: Dual-stream spatiotemporal fusion with dynamic feature mapping for gait-based identity recognition Journal: International Journal of Reasoning-based Intelligent Systems (IJRIS) 2026 Vol.18 No.17 pp.67 - 82 Abstract: Gait recognition offers the advantage of being contactless, but it is susceptible to interference from clothing and viewing angles. The key lies in extracting stable individual motion features. To address this, we propose a dynamic feature mapping framework based on skeletal keypoints. Unlike existing skeletonbased methods that directly use raw joint coordinates, our framework incorporates three key innovations: 1) an explicit kinematic feature mapping module that transforms raw coordinates into joint angles, angular velocities, and angular accelerations; 2) a dualstream spatiotemporal graph convolution architecture that separately processes positional and kinematic features; 3) a framewise spatial attention mechanism that dynamically reweights body parts according to input conditions. This framework employs dual-stream spatio-temporal convolutional networks to fuse features such as joint positions, angles, and angular velocities, and introduces attention mechanisms to adaptively weight the contributions of different body parts. On the CASIA-B dataset, the accuracy reached 95.8%, an improvement of 12.6 percentage points over GaitGraph; on Gait3D, the Rank-1 accuracy score reached 83.7%, an improvement of 25.0 percentage points over gait global-local model (GaitGL). The results demonstrate that dynamic features effectively capture differences in walking patterns and are robust to variations in appearance, providing a new approach for model-based gait recognition. Inderscience Publishers - linking academia, business and industry through research

Title: Dual-stream spatiotemporal fusion with dynamic feature mapping for gait-based identity recognition

Authors: Binge Quan; Beibei Zhang

Addresses: Anhui Police College, Hefei, 238000, China ' Anhui Police College, Hefei, 238000, China

Abstract: Gait recognition offers the advantage of being contactless, but it is susceptible to interference from clothing and viewing angles. The key lies in extracting stable individual motion features. To address this, we propose a dynamic feature mapping framework based on skeletal keypoints. Unlike existing skeletonbased methods that directly use raw joint coordinates, our framework incorporates three key innovations: 1) an explicit kinematic feature mapping module that transforms raw coordinates into joint angles, angular velocities, and angular accelerations; 2) a dualstream spatiotemporal graph convolution architecture that separately processes positional and kinematic features; 3) a framewise spatial attention mechanism that dynamically reweights body parts according to input conditions. This framework employs dual-stream spatio-temporal convolutional networks to fuse features such as joint positions, angles, and angular velocities, and introduces attention mechanisms to adaptively weight the contributions of different body parts. On the CASIA-B dataset, the accuracy reached 95.8%, an improvement of 12.6 percentage points over GaitGraph; on Gait3D, the Rank-1 accuracy score reached 83.7%, an improvement of 25.0 percentage points over gait global-local model (GaitGL). The results demonstrate that dynamic features effectively capture differences in walking patterns and are robust to variations in appearance, providing a new approach for model-based gait recognition.

Keywords: gait recognition; skeletal landmarks; kinematic features; identity authentication.

DOI: 10.1504/IJRIS.2026.154536

International Journal of Reasoning-based Intelligent Systems, 2026 Vol.18 No.17, pp.67 - 82

Received: 27 Apr 2026
Accepted: 25 May 2026
Published online: 02 Jul 2026 *

Title: Dual-stream spatiotemporal fusion with dynamic feature mapping for gait-based identity recognition

Keep up-to-date