Let's Go Bananas: Beyond Bounding Box Representations for Fisheye Camera-Based Object Detection in Autonomous Driving

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving

Background:

• Object detection is crucial for autonomous driving, with pedestrian detection being a well-established area.
• Fisheye cameras offer wide-angle, near-field sensing but present unique challenges for standard object detection methods due to severe radial distortion.

Purpose of the Study:

• To explore and develop effective object representations for fisheye cameras in autonomous driving.
• To address the limitations of traditional bounding boxes in distorted fisheye imagery.

Main Methods:

• Implemented a YOLO (You Only Look Once) based framework to evaluate object representations on the WoodScape dataset.
• Investigated standard bounding boxes, oriented bounding boxes, ellipses, generic polygons, and proposed novel curvature-adaptive polygons and curved boxes.
• Incorporated vanishing-point constraints and a camera geometry tensor for improved distortion adaptation.

Main Results:

• The curvature-adaptive polygon improved mean average precision (mAP) by 3 points over standard bounding boxes.
• The proposed curved box representation, enhanced with vanishing-point constraints, outperformed standard bounding boxes by 3 mAP and oriented bounding boxes by 1.6 mAP.
• Further improvements of 1.4 mAP were achieved using a camera geometry tensor for non-linear distortion adaptation.

Conclusions:

• Traditional bounding boxes are inadequate for fisheye cameras due to significant distortion.
• Novel curved box representations offer a practical and accurate solution for object detection in fisheye images.
• The developed methods enhance the reliability of near-field sensing for autonomous vehicles.