Core Concepts
A low-power, low-latency neuromorphic approach for accurate pupil tracking using event-based vision and spiking neural networks.
Abstract
This paper introduces a novel neuromorphic dataset and methodology for efficient eye tracking using event data captured by a Dynamic Vision Sensor (DVS). The key contributions are:
Ini-30 Dataset: The first event-based eye tracking dataset collected with two DVS cameras mounted on a glass frame, capturing natural eye movements in unconstrained settings.
Retina Model: A lightweight Spiking Neural Network (SNN) architecture based on Integrate-and-Fire (IAF) neurons, featuring only 64k parameters. Retina achieves a pupil tracking error of 3.24 pixels on a 64x64 DVS input, outperforming the state-of-the-art event-based method 3ET.
Neuromorphic Hardware Deployment: Retina is deployed on the Speck neuromorphic processor, demonstrating end-to-end power consumption between 2.89-4.8 mW and latency of 5.57-8.01 ms, making it suitable for energy-efficient, low-latency eye tracking applications.
The authors show that Retina's performance is superior to 3ET, while being 35 times more computationally efficient. This work paves the way for further development of neuromorphic solutions for real-world, event-based eye tracking.
Stats
The Ini-30 dataset contains event data from 30 volunteers, with recording durations ranging from 14.64 to 193.8 seconds and event counts varying from 4.8 million to 24.2 million.
The median sampling time step is 200 microseconds, with the number of events per timestamp ranging from 3 to 5,000.
Quotes
"Retina is the first eye tracking algorithm, suitable for deployment on neuromorphic hardware."
"Retina shows superior precision (-20% centroid error) and reduced computational complexity (-30x MAC) compared to 3ET."