Improving Passenger Detection With Overhead Fisheye Imaging

Abstract

Detecting passengers within overhead, fisheye images presents a unique set of challenges. Traditional approaches rely on radially-aligned bounding boxes based on the assumption that people are consistently oriented along the image radius. This assumption simplifies the detection process but introduces limitations in terms of flexibility and detection accuracy. Additionally, these methods often require extensive pre and post-processing, significantly increasing the computational complexity. We propose an innovative, end-to-end, rotation-aware detection framework specifically designed for the accurate detection of passengers using angle-oriented bounding boxes. This study investigates a fully convolutional neural network (CNN) that performs direct orientation regression of each bounding box, enhanced by a scale and angle loss function that effectively accounts for the periodicity of angles, ensuring accurate and robust bounding box orientation predictions. Moreover, we present a new dataset tailored to in-cabin passenger detection and counting. Our experimental results show an improvement of 5.3% in average precision, compared to state-of-the-art methods in overhead people detection. Finally, we demonstrate results from real vehicle experiments in Copenhagen and Geneva, highlighting the importance of this work for public transport operators.

Keywords

• Artificial intelligence
• edge computing
• autonomous vehicles
• computer vision
• in-cabin monitoring
• passenger counting