Gait Analysis¶
One-line summary: Recognizing individuals by their walking pattern — a soft biometric that works at a distance without subject cooperation, using silhouette-based or skeleton-based representations.
Modality: Gait
Related concepts: Deep Learning Architectures for Biometrics, Transformer Architectures for Biometrics, Multimodal Biometrics, Bias and Fairness in Biometrics, Real World Biometric Deployments, Biometric Datasets and Benchmarks
Last updated: 2026-04-04
Overview¶
Gait recognition identifies people by how they walk. It is the only mainstream biometric that operates at long range (50–100+ meters) without subject awareness or cooperation, making it valuable for surveillance and forensic applications.
Two dominant representation paradigms:
- Appearance-based (silhouette) — Extract binary silhouettes from video frames, aggregate into a compact temporal representation (Gait Energy Image), and match via CNN or metric learning.
- Model-based (skeleton/pose) — Estimate body joint positions per frame using pose estimation (OpenPose, HRNet, MediaPipe), then model temporal dynamics of the skeleton sequence using RNNs, GCNs, or Transformers.
Pipeline¶
- Video capture — Surveillance camera, depth sensor, or radar.
- Person detection + tracking — Detect and track the subject across frames.
- Silhouette extraction / Pose estimation — Background subtraction or segmentation for silhouettes; 2D/3D pose estimation for skeletons.
- Temporal representation — GEI (silhouette), or joint coordinate sequences (skeleton).
- Feature extraction — CNN on GEI; GCN/Transformer on skeleton sequences.
- Matching — Cosine similarity on embeddings.
Technical Details¶
Silhouette-Based Methods¶
| Method | Year | Key Idea |
|---|---|---|
| Gait Energy Image (GEI) | 2006 | Average silhouette over one gait cycle; simple and effective baseline |
| GaitSet (Chao et al.) | 2019 | Set-based: treats gait as an unordered set of silhouettes; no explicit temporal modeling |
| GaitPart | 2020 | Part-based: splits silhouette into horizontal strips and learns local features |
| GaitGL (Lin et al.) | 2021 | Global-local feature extraction; combines set-level and part-level features |
| OpenGait (Fan et al.) | 2023 | Unified open-source framework; reproduces GaitSet, GaitPart, GaitGL, and adds new baselines |
| GaitBase / GaitGCI | 2023 | Strong ResNet-based baseline that outperforms many specialized methods; shows backbone matters more than architecture tricks |
| DeepGaitV2 | 2024 | Explores diverse backbone architectures (ConvNeXt, Swin, etc.) for gait |
Skeleton-Based Methods¶
| Method | Year | Key Idea |
|---|---|---|
| PoseGait (Liao et al.) | 2020 | 3D pose features (joint angles, limb lengths) + FC network |
| GaitGraph (Teepe et al.) | 2021 | Graph Convolutional Network on skeleton sequences |
| GaitGraph2 | 2022 | Multi-scale graph convolutions + augmentation strategies |
| GaitMixer | 2023 | MLP-Mixer on skeleton sequences; competitive with GCN approaches |
| GaitTR | 2024 | Transformer on skeleton sequences; best skeleton-based results on GREW |
Covariate Challenges¶
- Clothing — Carrying bags, wearing coats, different shoes alter silhouette shape significantly.
- View angle — Cross-view matching is a core challenge; most methods learn view-invariant features.
- Speed — Walking speed affects gait cycle length and joint dynamics.
- Surface/terrain — Inclines, stairs, different flooring change gait patterns.
Datasets¶
| Dataset | Subjects | Setting | Notes |
|---|---|---|---|
| CASIA-B | 124 | Indoor, 11 views, 3 conditions (normal, bag, coat) | Classic benchmark; limited scale |
| OU-MVLP | 10,307 | Indoor, 14 views | Largest lab-based gait dataset |
| GREW (Gait Recognition in the Wild) | 26,345 | Outdoor, natural, unconstrained | Wild gait from street cameras; 4 subsets |
| Gait3D | 4,000 | Outdoor, 3D point clouds | 3D gait in the wild from multi-camera LIDAR |
| OUMVLP-Pose | 10,307 | Indoor, skeleton | Pose-estimated version of OU-MVLP |
| FVG (Frontal-View Gait) | 226 | Indoor, frontal view only | Frontal gait recognition |
Challenges¶
- In-the-wild performance — Lab-to-wild generalization remains the biggest gap; GREW benchmark is humbling even for SOTA methods (Rank-1 ~70–80% vs. >95% on CASIA-B).
- Silhouette quality — Background subtraction fails in crowded scenes, poor lighting, and dynamic backgrounds. Semantic segmentation (Mask R-CNN) helps but adds compute.
- Clothing and carrying conditions — Still the primary covariate degrading accuracy 10–30%.
- Occlusion — Partial body visibility in real scenes; part-based methods handle this better.
- Privacy and ethics — Gait can be captured covertly; raises surveillance concerns. See Bias and Fairness in Biometrics, Privacy Preserving Biometrics.
- Low uniqueness — Gait has lower discriminative power than face, iris, or fingerprint; typically used as a soft biometric or for re-identification rather than 1:N identification at scale.
State of the Art (SOTA)¶
As of early 2026: - CASIA-B (NM, BG, CL): Rank-1 ~98%, ~95%, ~88% respectively for top silhouette methods (GaitBase + augmentation). - OU-MVLP: Rank-1 ~93% (GaitGL, DeepGaitV2). - GREW (in the wild): Rank-1 ~75% (best published); significant room for improvement. - Gait3D: Rank-1 ~55–65% (challenging 3D outdoor benchmark). - Skeleton-based on GREW: Rank-1 ~60% (GaitTR); silhouette methods still dominate.
Open Questions¶
- Can multimodal (silhouette + skeleton + depth) fusion close the gap on in-the-wild benchmarks?
- Will self-supervised pre-training on large-scale unlabeled walking video improve generalization?
- How to build gait recognition that is truly invariant to clothing without losing gait-specific discriminative features?
- Can radar-based gait (through-wall, privacy-preserving) become practical for smart home or healthcare applications?
- Ethical framework: should gait recognition be regulated differently from face recognition given its covert nature?
References¶
- Chao, H. et al. (2019). GaitSet: Regarding Gait as a Set for Cross-View Gait Recognition. AAAI.
- Fan, C. et al. (2023). OpenGait: Revisiting Gait Recognition Towards Better Practicality. CVPR.
- Zhu, Z. et al. (2022). Gait Recognition in the Wild: A Benchmark. ICCV.
- Teepe, T. et al. (2022). GaitGraph: Graph Convolutional Network for Skeleton-Based Gait Recognition. ICIP.
Backlinks: Deep Learning Architectures for Biometrics, Transformer Architectures for Biometrics, Multimodal Biometrics, Bias and Fairness in Biometrics, Biometric Datasets and Benchmarks, Real World Biometric Deployments