Exploiting Lightweight Statistical Learning for Event-Based Vision Processing.

Area of Science:

• Computer Vision
• Machine Learning
• Embedded Systems

Background:

• Event-based vision systems capture visual information using dynamic vision sensors (DVS).
• Data is represented as asynchronous streams of pixel events indicating intensity changes.
• Existing algorithms like SCNNs and BoE have limitations in speed and complexity.

Purpose of the Study:

• To present a lightweight statistical learning framework for low-cost event-based vision.
• To improve processing speed and simplify online learning for DVS data.
• To achieve state-of-the-art classification accuracy on benchmark datasets.

Main Methods:

• Utilized a simple random ferns classifier.
• Employed randomly selected patch-based binary features to categorize pixel event flows.
• Developed a lightweight statistical learning framework for DVS data.

Main Results:

• Achieved state-of-the-art classification accuracy on MNIST-DVS, Poker-DVS, Posture-DVS, and CIFAR10-DVS datasets.
• Demonstrated 2x faster processing than BoE methods and >100x faster training than SCNNs.
• Framework shows high processing speed and simple online learning.

Conclusions:

• The proposed framework is highly suitable for low-cost event-based vision systems.
• Hardware estimations indicate preference for low-cost embedded system implementations.
• Offers a competitive alternative to existing event-based processing algorithms.