核心概念
A highly efficient and accurate deep learning model for real-time gait-based person identification that can be deployed on low-power edge devices.
要約
This paper presents a lightweight convolutional neural network (CNN) model for real-time person identification using gait analysis. The key highlights are:
Dataset: The authors used the public whuGAIT dataset and augmented it with data collected from the four team members, resulting in a total of 24 classes.
Feature Extraction: The authors used spectral analysis to extract features from the raw accelerometer and gyroscope data, leveraging the richer information in the gyroscope data compared to the accelerometer.
Model Architecture: The authors designed a 4-layer CNN model that achieves 96.7% accuracy on the test set, while consuming only 5KB of RAM and 70ms inference time on an Arduino Nano 33 BLE Sense board.
Deployment: The authors successfully demonstrated real-time person identification using the model running on the Arduino board, as well as on a smartphone. They also converted the CNN model to a spiking neural network (SNN) and deployed it on the BrainChip Akida neuromorphic processor.
Challenges and Iterations: The authors faced challenges in hyperparameter tuning, data collection, and deployment, requiring numerous iterations to arrive at the final solution.
Advantages: The edge-based approach offers several benefits, including reduced bandwidth, low latency, cost-effectiveness, reliability, and privacy preservation.
Ethical Considerations: The authors discuss the importance of addressing privacy, informed consent, and bias in the development and deployment of such gait-based identification systems.
Overall, this work presents a highly efficient and practical solution for real-time person identification using gait analysis, with the potential for deployment in various applications, such as surveillance, security, and healthcare.
統計
Each person has a unique gait, influenced by factors like anatomy, musculoskeletal structure, and personal habits.
The final model achieves 96.7% accuracy and consumes only 5KB RAM with an inferencing time of 70 ms and 125mW power, while running continuous inference on Arduino Nano 33 BLE Sense.
引用
"Every person has a distinct gait, influenced by factors like anatomy, musculoskeletal structure, and personal habits. This distinctiveness makes gait analysis an intriguing and effective tool for identifying individuals in diverse settings, ranging from surveillance and security to healthcare and rehabilitation."
"Our model achieves 96.7% accuracy and consumes only 5KB RAM with an inferencing time of 70 ms and 125mW power, while running continuous inference on Arduino Nano 33 BLE Sense."