TL;DR
This paper introduces a novel, unobtrusive continuous authentication system for mobility scooters using rider posture data. By leveraging few-shot learning and a spatio-temporal graph convolutional network, the system achieves high accuracy with minimal enrollment data, addressing the security needs of mobility scooter users, particularly seniors.
What Problem Does This Paper Address?
Mobility scooters offer crucial independence, but their security is often lacking. Traditional authentication methods like physical keys or passwords are inconvenient and difficult for many users, especially seniors who may experience memory issues. Existing continuous authentication techniques often rely on specialized sensors or lengthy enrollment, making them unsuitable for this user group. The authors aim to bridge this gap by developing a seamless, user-friendly, and secure authentication system that protects mobility scooters from unauthorized use and takeovers.
Methods
The core of the proposed system is a deep learning architecture designed to extract unique features from a rider's upper-body posture. Here's a breakdown of the key techniques:
- Posture Data Extraction: The system uses user-facing cameras to capture keypoint coordinates of the rider's torso, shoulders, elbows, wrists, and hips. This data is processed from video frames.
- Spatio-Temporal Graph Convolutional Networks (GCNs): To model the complex relationships between body joints over time, the authors employ GCNs. These networks are adept at understanding how joints move in relation to each other both spatially within a frame and temporally across frames.
- Hierarchical Encoding: The architecture incorporates a body-part-based hierarchical encoding structure. This allows for specialized processing of different body segments (upper torso, lower torso, left arm, right arm) before combining their features.
- Few-Shot Learning with Triplet Metric Loss: To enable quick enrollment, the system utilizes few-shot learning. Triplet Metric Loss is employed to train the model to distinguish between embeddings of the same rider and different riders, ensuring that embeddings from the same user are close together in the feature space, while those from different users are pushed apart with a defined margin.
- Triplet Mining: To accelerate training, the authors use triplet mining, which strategically selects challenging positive and negative sample pairs to optimize the learning process.
Datasets
The authors collected a new dataset specifically for mobility scooter riding.
- Mobility Scooter Riding Dataset: This dataset comprises data from 42 individuals performing various riding tasks on campus, including forward/backward riding, turns, and rotations, on both pavement and grass. The total recorded footage is approximately 10 hours and 9 minutes, amounting to 1.1 million frames. This dataset is made publicly available.
Key Insights and Ideas
- Posture as a Robust Biometric: The research highlights that subtle variations in rider posture, captured by keypoint data, can serve as a unique and reliable biometric for continuous authentication.
- Effectiveness of Few-Shot Learning: The system's ability to achieve high accuracy with only a few enrollment samples is a significant advantage, making it practical for real-world deployment where users may not want to spend extensive time enrolling.
- GCNs for Spatio-Temporal Body Dynamics: The application of spatio-temporal GCNs proves effective in capturing the intricate movements and relationships of upper-body keypoints during scooter riding, which are distinct from full-body movements.
- Unobtrusive Security: By relying solely on camera-based posture data, the system offers a seamless and unobtrusive security solution that doesn't require additional hardware or user interaction beyond normal riding.
Key Results
The experimental results demonstrate the effectiveness of the proposed system:
- High Accuracy with Minimal Enrollment: Using the MoveNet pose estimation model, the system achieved an Area Under the Curve (AUC) of 0.981 with just 1 enrollment sample, and performance improved to 0.990 with 10 samples.
- Comparison of Pose Estimation Models: The study compared MediaPipe and MoveNet for pose estimation. MoveNet generally yielded higher AUC scores and less variability compared to MediaPipe.
- Practical Deployability: The system is designed for quick enrollment and easy deployment, making it a viable solution for enhancing mobility scooter security.
Limitations and Future Work
- Pose Estimation Model Variability: While MoveNet performed better, the paper notes that MediaPipe showed greater variability, suggesting that the choice of pose estimation model can impact performance.
- Environmental Factors: The paper doesn't extensively discuss how factors like lighting conditions, occlusions (e.g., clothing obscuring joints), or extreme weather might affect performance.
- Future Work: The authors suggest further research into improving the system's robustness against environmental variations and exploring more advanced deep learning architectures for even greater accuracy and efficiency. They also aim to expand the dataset and conduct more extensive real-world testing.
Summary updated