Gait Biometrics Flashcards
(21 cards)
Properties of Gait
Gait is non-contact and users sequences
Advantages: perceivable at distance and hard to disguise. Works at low resolutions
Security/surveillance, immigration, forensic and medical applications
Gait in medicine
For disease diagnosis through video, optoelectronics, moving light displays, etc
Gait Databases
Southampton
CASIA
HiD
HiD Database
Human ID at Distance.
122 subjects
Cannon Progressive Scan 30fps
Included change in surface, shoe and luggage
Included evaluation protocol
Southampton Dataset
> 100 Subjects
Filmed indoors and outdoors
Included covariate data for 12 subjects
Also from HiD
CASIA Dataset
124 Subjects
11 Viewpoints
Different Clothing and carrying conditions
Gait extraction and description techniques
Silhouette descriptors (many)
- statistical analysis
- temporal symmetry
- velocity moments
- unwrapped silhouette
Modelling Movement
- pendular thigh motion
- coupled and forced oscillator
- anatomically-guided skeleton
Velocity Moments
Combine information about motion (velocity) and shape (moments).
Equate to some Velocity Function * some centralised moment for every pixel.
Moment output can dramatically change based on walking type.
Average silhouette
Most popular simple and effective gait technique, aka gait energy image or gait entropy image (newer version)
Average Silhouette Method
Background is taken from each frame and pixels thresholded resulting in binary image
Normalise silhouettes by height to account for distance
Add all silhouettes together and divide by the number of frames
Resulting image is the signature
HiD Baseline Analysis
Form Silhouette
Detect Gait periods
Similarity between centred silhouettes = Intersect over Union (IoU) per frame
Estimate correlation of frame similarity between sequences
Similarity is median of max correlation between gallery probe sequences
HiD Baseline Performance
Recognition accuracy is highest in controlled settings and drops with variation in appearance, surface or view.
Gait CASIA Method
Human Detection and Tracking:
- Background modelling
- Motion Segmentation
- Human Tracking
Feature Extraction:
- Silhouette Extraction
- 2D Silhouette Unwrapping
- 1D Signal Normalisation
Training or Classification:
- Eigenspace computation
- Projection
- Recognition
CASIA Silhouette Representation
Unwrapping the silhouette gives a signal
Signal is process for recognition (eigen analysis) and Gait (correlation)
Aspect Ratio (1st Row)
Background subtraction and PCA (2nd)
Autocorrelation
First derivative of correlation
Positions of peaks
CASIA Gait Performance
Wang & Tan’s CASIA gait method is both accurate and computationally efficient.
It performs similarly to other methods on CASIA and HiD datasets but highlights that viewpoint variation affects silhouette-based recognition
Modelling GAIT
Extended pendular thigh model, based on angle at the leg.
Uses forced oscillator/ bilateral symmetry/ phase coupling.
Gait Cycle is described in angles over time.
Gait Model Recognition
Accumulate evidence of Fourier descriptors.
Use to calculate Fourier Transform
Calculate magnitude spectrum across frequencies
Phase spectra shows when in a cycle a frequency occurs.
Product of phase and magnitude gives recognition measure
Covariate Data Gait
Covariate Data negatively impacts performance. Walking speed and footwear leading to largest performance spread.
Gait viewpoint Invariance
Can compensate for camera view for 25-144 degrees.
Various assumptions made.
Recognition Immune to view and covariates.
Semantically Mediated Biometric Fusion/ Soft Biometrics
Combines sensor data with information sources (semantic features) to build a classification system.
Useful in surveillance and forensics.
Mimics eye witness data and fuses it with sensor.
This increases recognition performance consistently.
Invariant Gait Results
The rectified method handles viewpoint variation and covariates (bags, clothing) far better than the silhouette-only approach, leading to much more consistent recognition performance across angles.
