Detecting Line Segments in Motion-Blurred Images With Events

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Motion blur significantly degrades line segment detection accuracy in visual SLAM and 3D mapping.
• Existing methods struggle with performance degradation under motion blur conditions.
• Event data offers high-temporal resolution, minimal blur, and edge awareness, complementing image data.

Purpose of the Study:

• To develop a robust line segment detection method that overcomes motion blur challenges.
• To leverage the complementary information from both image and event data for improved detection.
• To create new datasets for training and evaluating line segment detection under motion blur.

Main Methods:

• A general frame-event feature fusion network was designed, incorporating channel-attention and self-attention modules.
• The fused feature maps were processed using state-of-the-art wireframe parsing networks.
• Two datasets, synthetic FE-Wireframe and realistic FE-Blurframe, were created for training and evaluation.

Main Results:

• The proposed fusion network demonstrated effectiveness in combining image textures and event edges.
• The method achieved the highest detection accuracy compared to state-of-the-art approaches.
• The approach maintained comparable real-time performance and showed robustness to motion blur and high dynamic range scenes.

Conclusions:

• Combining image and event data provides a robust solution for line segment detection under motion blur.
• The developed fusion network and datasets contribute to advancing reliable line detection in challenging conditions.
• The method shows potential for practical applications requiring accurate line segment detection in dynamic environments.