A Saccade Based Framework for Real-Time Motion Segmentation Using Event Based Vision Sensors.

Area of Science:

• Robotics
• Computer Vision
• Sensor Technology

Background:

• Motion segmentation is crucial for autonomous robotic systems in cluttered environments.
• Event-based sensors offer low power consumption and high temporal resolution but present challenges in motion segmentation due to data asynchrony.
• Existing object tracking methods using neuromorphic sensors are limited when the sensor is static or object models are unknown.

Purpose of the Study:

• To develop a generalized motion segmentation technique for event-based sensors that overcomes limitations of prior approaches.
• To enable robust object tracking in dynamic scenes with a moving camera.

Main Methods:

• Introduced platform micromotion, inspired by saccadic eye movements, to separate static and dynamic scene elements.
• Developed a novel 'spike-groups' methodology to partition spatio-temporal event data for scene statistics computation and object characterization.

Main Results:

• The proposed algorithm successfully performs motion segmentation using spatial statistics across time frames.
• Experimental results demonstrated a maximum accuracy of 92% in classifying dynamic objects with a moving camera.

Conclusions:

• The developed technique provides a robust solution for motion segmentation with event-based sensors, even under challenging conditions like camera motion.
• This advancement facilitates improved object tracking for autonomous robotic systems in complex environments.