Related Experiment Videos
Spiking Optical Flow for Event-Based Sensors Using IBM's TrueNorth Neurosynaptic System
IEEE Transactions on Biomedical Circuits and Systems
|July 12, 2018
Summary
This study presents a low-power, spike-based neural network for optical flow estimation using dynamic vision sensors. The system achieves accurate velocity inference with under 80 mW power consumption.
Area of Science:
- Neuromorphic Engineering
- Computer Vision
- Artificial Intelligence
Background:
- Optical flow estimation is crucial for understanding motion in dynamic scenes.
- Traditional methods often require significant computational power and energy.
- Dynamic vision sensors offer event-based data with high temporal resolution.
Purpose of the Study:
- To develop a fully spike-based neural network for optical flow estimation.
- To create a low-power embedded implementation using a dynamic vision sensor and a neuromorphic system.
- To evaluate the system's accuracy and power efficiency.
Main Methods:
- Utilized an asynchronous time-based image sensor to generate event-driven spikes.
- Processed spikes using a spiking neural network on IBM's TrueNorth Neurosynaptic System.
- Implemented a variant of the Barlow Levick method for optical flow calculation.
Main Results:
- Achieved accurate optical flow estimation with an average endpoint error of 11%.
- Demonstrated a low-power embedded implementation with a power budget under 80 mW.
- The system effectively infers velocity by analyzing spike timing from neighboring pixels.
Conclusions:
- A fully spike-based approach is viable for efficient optical flow estimation.
- The integrated system offers a low-power solution for real-time motion analysis.
- This work advances the application of neuromorphic computing in computer vision.