Detecting Every Object From Events.

Area of Science:

• Computer Vision
• Robotics
• Sensor Technology

Background:

• Class-agnostic object detection (CAOD) is crucial for autonomous driving but faces challenges with unknown object categories.
• Traditional RGB cameras in CAOD systems exhibit high latency and limited dynamic range, posing safety risks in extreme conditions (e.g., darkness, fast motion).

Purpose of the Study:

• To introduce a novel approach for high-speed, class-agnostic object detection using event-based vision.
• To address the limitations of RGB cameras in autonomous driving perception systems.

Main Methods:

• Proposed Detecting Every Object in Events (DEOE), an approach utilizing event-based vision.
• Employed a recurrent vision transformer backbone to process event data.
• Incorporated spatial and temporal consistency analysis to identify potential objects.
• Developed a disentangled objectness head for improved foreground-background classification and novel object discovery.

Main Results:

• DEOE demonstrated superior performance in both open-set and closed-set scenarios.
• The method outperformed existing strong baseline methods in class-agnostic object detection.
• Event-based vision proved effective for robust object detection under challenging conditions.

Conclusions:

• Event-based vision offers a promising alternative for reliable object detection in autonomous driving.
• The DEOE approach enhances the generalization and accuracy of class-agnostic object detection systems.
• The proposed method improves safety by enabling detection in extreme environmental conditions.