Low Cost and Latency Event Camera Background Activity Denoising.

Area of Science:

• Computer Vision
• Sensor Technology
• Machine Learning

Background:

• Dynamic Vision Sensor (DVS) cameras generate background activity (BA) noise, especially in low light.
• Current denoising algorithms struggle with high noise levels and lack standardized quantitative comparison methods.

Purpose of the Study:

• To propose a novel framework for quantitatively evaluating BA denoising algorithms for DVS event cameras.
• To introduce and compare three new low-cost denoising algorithms.

Main Methods:

• Developed a framework using receiver operating characteristics (ROC) with known signal/noise mixtures.
• Created new datasets for stationary and moving DVS camera scenarios (surveillance, driving).
• Implemented and compared three novel algorithms: fixed window, pixel-proportional memory, and multilayer perceptron (MLP).

Main Results:

• Algorithm 1 (fixed window) is effective for stationary cameras.
• Algorithm 2 (pixel-proportional memory) shows improved correlation checking and better noise/signal balance than existing methods.
• Algorithm 3 (MLP classifier) achieved the highest accuracy across all tested datasets.

Conclusions:

• The proposed ROC-based framework offers a robust method for quantifying DVS BA denoising performance.
• The new algorithms, particularly the MLP-based one, demonstrate significant improvements in noise reduction and signal preservation.
• Shared code and data (DND21) facilitate further research and development in DVS event camera denoising.