Identification of a Person by Gait Kinematics in Three-Dimensional Space Based on Invariant Features
Year: 2026
Type: article
Abstract: A system for data acquisition, preprocessing, and analysis for gait-based personal identification using machine learning and image analysis methods is proposed. The system includes a preprocessing mechanism for normalizing biometric motion data and a post-processing stage for extracting features invariant to orthogonal transformations of space and changes in camera pose, including perspective effects. A computationally efficient gait recognition method using a single stereo camera is implemented. The feature space is analyzed to determine the most informative characteristics, and the employed feature set is comparatively assessed. Several machine-learning architectures, including deep learning models, are evaluated with respect to both accuracy and computational efficiency. System performance is experimentally studied on various computing devices, and processing time characteristics are measured. The results enable comparison of applicability and efficiency depending on available computational resources. The input data are formed as time series of 3D coordinates of keypoints reconstructed from depth data using 2D pose detections (COCO, 17 keypoints), where missing detections are encoded by a zero vector. For gait identification, head keypoints are excluded as uninformative, while missing values are filled and the resulting trajectories are subsequently smoothed. An invariant feature space is built from the skeleton geometry (segment lengths, joint angles, and their derivatives); then these features are aggregated into a single vector using a sliding temporal window, i.e., each vector describes motion over an interval of several consecutive frames. At the identification stage, threshold-based filtering and aggregation of predictions at the frame level are applied. It is shown that both accuracy and processing speed depend on the chosen machine-learning model and the available computational resources.
Source: Informatics and Automation
DOI: 10.15622/ia.25.3.5
Institution: Institute for Information Transmission Problems
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7204470341