Comparative Study of Markerless Vision-Based Gait Analyses for Person Re-Identification
Jaerock Kwon1, Yunju Lee2, Jehyung Lee3
1Department of Electrical and Computer Engineering, University of Michigan-Dearborn, Dearborn, MI 48128, USA.
Sensors (Basel, Switzerland)
|December 28, 2021
Summary
A Siamese Long Short Term Memory (LSTM) network effectively identifies individuals using 3D gait patterns, outperforming traditional feature-based methods. This advancement aids in applications like forensic gait analysis and neurological disorder detection.
Area of Science:
- Computer Vision
- Biomechanical Engineering
- Machine Learning
Background:
- Model-based gait analysis utilizes human kinematic characteristics for individual identification.
- Extracting gait features from 3D anatomical landmarks is crucial, but 2D images present challenges like viewpoint variance and occlusion.
- Estimating 3D joint positions from 2D sequences is essential for real-world gait analysis.
Purpose of the Study:
- To compare feature-based and spatiotemporal-based viewpoint-invariant person re-identification using gait patterns.
- To evaluate the efficacy of a Siamese LSTM network against traditional gait feature extraction methods.
- To assess the potential of gait analysis for applications in rehabilitation, forensics, and medical diagnostics.
Main Methods:
- A comparative study was conducted on two approaches: gait features from 3D joint positions and a Siamese LSTM network using 3D spatiotemporal changes.
- The Siamese LSTM network classifies individuals directly from gait patterns without explicit feature extraction.
- Experiments were performed on the MARS and CASIA-A open datasets to validate and compare the methods.
Main Results:
- The Siamese LSTM network demonstrated superior performance, outperforming gait feature-based approaches by 20% on the MARS dataset and 55% on the CASIA-A dataset.
- Feature-based gait analysis using 2D and 3D pose estimators was found to be premature.
- The study highlights the potential of deep learning models for robust gait recognition.
Conclusions:
- The Siamese LSTM network offers a more effective approach to person re-identification using gait patterns compared to traditional feature extraction.
- Further research is needed to develop large-scale gait datasets and improve 2D/3D joint position estimators for gait analysis.
- Advancements in gait analysis hold promise for diverse applications, including forensic science and early disease detection.
