Abstract
Gait recognition is an emerging biometric method where an individual is identified according to their unique walking pattern, which is recorded with minimal intrusion through vision-based systems. Vision-Based Gait Recognition: Operational Framework, Key Characteristics, Challenges, Applications, and Adversarial Attack Vulnerabilities is a report that explores the operational framework of vision-based gait, its key characteristics, challenges, applications and adversarial attack vulnerabilities. The report contains an analysis of the process from data acquisition to matching to illustrate the strengths and weaknesses of this technology. It discusses applications in security, healthcare, and innovative environments and the enormous risks from adversarial attacks that threaten system reliability.
Introduction
Gait recognition leverages the distinctive walking patterns of individuals as a biometric identifier. Gait recognition is noninvasive; unlike fingerprints, another traditional biometric, or facial recognition, it can be conducted without direct contact with the subject and even at a distance using imaging sensors. The versatility and applicability in real-life situations have propelled vision-based gait recognition systems that capture gait data through cameras to a relatively mature stage. This paper presents a complete survey of vision-based gait recognition systems, covering their operational principles, characteristics, challenges, applications, and ability to be robust to adversarial attacks.
1- Overview of Gait Recognition Process
It is arguable that the stages of the vision based on the gait recognition process are all critical in the evolution of accurate identification. These stages are outlined in the following subsections (Gait Recognition, 2020).
1.1- Data Acquisition
Imaging sensors, generally RGB cameras, depth cameras, or infrared sensors, are used to collect gait data. Usually, these devices record video sequences of an … Factors such as camera resolution, frame rate, and environmental conditions (e.g., lighting, occlusions) all determine the quality of data. A specific example of this regard is high-resolution cameras to get precise body movements and depth cameras for 3D gait information, while robustness is further improved (Nixon et al. 2010).
1.2- Preprocessing
Preprocessing of raw video data for the relevant gait features is performed. The first step is to subtract the background from the subject to isolate the walking figure and then extract the silhouette to indicate the shape of the walking figure. The third step is normalisation so that data from different conditions are standardised (i.e., taking the walking figure at a distance from the camera). Gaussian filtering is applied to the data to improve data quality, such as noise reduction (Makihara et al., 2015).
1.3-Feature Extraction
The distinct gait characteristics, such as stride length, walking speed, and joint angles, are extracted as features. Models-based and appearance-based are the two primary approaches. Skeletal model-based methods compute joint movements for the subject’s body in a model, and appearance-based methods are based on the analysis of silhouette shape or motion patterns. The standard features consist of Gait Energy Image (GEI) averaging silhouettes for a gait cycle and Frequency-Domain Entropy that describes temporal variability (Han & Bhanu, 2006).
1.4 Matching and Identification
Machine learning or deep learning algorithms, such as Convolutional Neural Networks (CNN) or Support Vector Machines (SVM), are used to compare extracted features with a database of known gait signatures. Similarity scores (e.g., Euclidean distance) are computed, and the system determines a match. The system identifies the input gait as the most similar in identification mode or confirms whether the input belongs to the claimed identity (Gait Recognition 2020).
2. Characteristics and Challenges
Vision-based gait recognition systems are unique to other biometric modalities and have unique challenges.
2.1 – Key Characteristics
- Non-Intrusiveness: Gait data can be collected without subject cooperation and is suitable for covert identification.
- Distance Capability: Systems can operate at a distance (up to 50 meters), unlike contact-based biometrics (Nixon et al., 2010).
- Robustness to Appearance Changes: The robustness to appearance changes such as clothing and hairstyle does not significantly affect gait quality.
- Temporal Dynamics: Gait data offer rich information for analysis, combining spatial (body shape) and temporal (motion), and being temporal at once (Makihara et al., 2015).
2.2- Challenges
- Environmental Variability: The accuracy or merging miss rate may be degraded by lighting changes, occlusions, and background clutter (Han & Bhanu, 2006).
- View Angle Sensitivity: Gait features are highly sensitive to the camera’s range of view (front vs. side) and thus require multi-view or view-invariant algorithms.
- Covariate Factors: The clothing, footwear, or carried objects (such as bags) can influence gait patterns, resulting in difficulties in identification.
- Computational Complexity: Deep learning models provide excellent predictive power but require considerable computational resources, making them infeasible for real-time applications.
- Data Scarcity: The lack of large, diverse gait datasets hinders model training and generalisation (Makihara et al., 2015).
3 – Applications of Gait Recognition
There are many applications of gait recognition concerning various domains due to its noninvasive and robust nature.
3.1- Security and Surveillance
Security systems that recognise suspects’ gait in public areas, such as airports or train stations, use gait recognition quite widely. It helps to covertly monitor where faces are covered (Nixon et al., 2010).
3.2- Healthcare
Gait analysis helps diagnose neurological disorders (such as Parkinson’s) since it can detect abnormal walking patterns in medical settings. It also supports rehabilitation by monitoring patient progress (Gait Recognition, 2020).
3.3- Smart Environments
In smart homes or offices, gait recognition allows seamless user authentication through access control. It also facilitates personalised retail services in type switching, for instance, by identifying returning customers by their gait (Makihara et al., 2015).
3.4- Forensics
Gait recognition is used in forensic investigations to identify people from surveillance footage and to distinguish them from other biometrics when any other one doesn’t exist (Han, Bhanu, 2006).
4- Adversarial Attacks on Gait Recognition Systems
The security and reliability of gait recognition systems are threatened by adversarial attacks, particularly when the models are based on deep learning.
4.1 – Nature of Adversarial Attacks
Adversarial attacks are when a subtle perturbation is added to input data (video frames) to deceive the recognition system. These perturbations, often imperceptible to a human, can cause discrimination or false negatives. An attacker could tamper with a gait video to confuse the system as to which person has a valid identity or even that it is not (Goodfellow et al., 2015).
4.2 – Attack Mechanisms
- Input Perturbation: Attackers add noise to video frames to disrupt feature extraction. Adversarial patches put on clothing can affect GEI-based silhouette features (Szegedy et al., 2014).
- Model Poisoning: Attackers manipulate training data to introduce biases in the model with the capacity to generalise compromised.
- Physical Attacks: Real-world attacks include physical attacks, such as wearing specific clothes or attempting to mimic another person’s gait to fool the system (Goodfellow et al., 2015).
4.3- Impact and Mitigation
Adversarial attacks are shown to break the reliance on gait recognition on a trustworthy model in important situations such as security and forensics. False positives or negatives result in unauthorised access and missed identifications. Adversarial training to mitigate the effects includes using perturbed inputs to train models and robust feature extraction techniques that prioritise invariant gait characteristics. Additional security can be gained from regular model updates and multi-modal biometrics (e.g., combining gait and facial recognition) (Szegedy et al., 2014).
Conclusion
Vision-based gait recognition systems are becoming very attractive because they provide a nonintrusive, powerful, and biometric solution for security, healthcare, and bright environments. The approach is based on low-level advanced imaging and
machine learning for reliable identification, except for data acquisition and matching. Nevertheless, performance is limited by environmental variability, the sensitivity to view angle, and computational demands. Further, adversarial attacks are a concern as they pose a significant threat. Future work on gait recognition systems will involve continued research on view-invariant algorithms, large-scale datasets and adversarial defence mechanisms to improve and add to the security of gait recognition systems.
Sample of Assignment Writing are below
Venezuela military intervention & Venezuela crisis
Venezuela & Venezuelan Political Crisis of Maduro
Chevrolet by General Motors in USA
General Motors in USA
What is the Metaverse, and why should care?
The Impact of Climate Change on Pollinators and Ecosystem Services
Best Market of Toyota in USA
