Towards Ultrafast Depth Sensing Via Active Event-Based Stereo Vision.

Area of Science:

• Computer Vision
• Robotics
• Sensor Technology

Background:

• Conventional frame-based stereo imaging struggles with fast-motion scenarios, limiting high-speed depth sensing.
• Designing novel paradigms for ultrafast depth sensing remains a critical research challenge.

Purpose of the Study:

• To propose and validate a new approach for high-speed dense depth sensing using active event-based stereo vision.
• To develop and evaluate a novel neural network, ActiveEventNet+, for processing stereo event streams.

Main Methods:

• Developed a stereo camera prototype integrating binocular event cameras and an infrared 2D pattern projector.
• Created real-world and synthetic datasets with synchronized labels for training and evaluation.
• Proposed ActiveEventNet+, a lightweight neural network featuring novel cost volume and temporal consistency architectures.

Main Results:

• ActiveEventNet+ generates high-quality dense disparity maps from stereo event streams with low latency.
• The system achieves real-time processing up to 150 FPS, outperforming state-of-the-art methods.
• Demonstrated superior depth sensing performance compared to conventional frame-based stereo cameras in high-speed scenes.

Conclusions:

• Active event-based stereo vision offers a promising paradigm for future high-speed depth sensing systems.
• The proposed ActiveEventNet+ provides an effective and computationally efficient solution for ultrafast depth sensing.
• The developed datasets and code facilitate further research in event-based stereo vision.