EyeFi: Fast Human Identification Through Vision and WiFi-based Trajectory Matching.

Area of Science:

• Computer Vision
• Wireless Sensing
• Machine Learning

Background:

• Human sensing and motion tracking are crucial for applications like retail, surveillance, and smart cities.
• Existing methods often rely on facial recognition or multiple devices, lacking long-term re-identification.
• There is a need for robust, standalone systems for accurate individual identification.

Purpose of the Study:

• To introduce EyeFi, a novel system combining WiFi and camera for human sensing and identification.
• To overcome limitations of existing solutions, such as reliance on facial recognition or multiple hardware units.
• To achieve accurate long-term re-identification of individuals.

Main Methods:

• Integrating a WiFi chipset with an overhead camera in a standalone device.
• Fusing motion trajectories from both vision and radio frequency (RF) data.
• Employing a student-teacher model to train a neural network for WiFi Angle of Arrival (AoA) estimation using Channel State Information (CSI).

Main Results:

• EyeFi significantly improves WiFi CSI-based AoA estimation accuracy by over 30%.
• The system demonstrates a 3,800-fold increase in computational speed compared to state-of-the-art solutions.
• Achieves an average person identification accuracy of 75% in real-world environments with 2-10 individuals.

Conclusions:

• EyeFi offers a promising, efficient, and accurate solution for human sensing and identification.
• The fusion of WiFi and camera data enhances motion trajectory estimation and re-identification.
• The system's performance validates its potential for diverse applications in surveillance, access control, and smart environments.