Related Experiment Video
Updated: Aug 2, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Equirectangular Image Data Detection, Segmentation and Classification of Varying Sized Traffic Signs: A Comparison of
Heyang Thomas Li1, Zachary Todd1, Nikolas Bielski1
1School of Mathematics and Statistics, College of Engineering, University of Canterbury, Christchurch 8041, New Zealand.
Sensors (Basel, Switzerland)
|April 13, 2023
Summary
This study introduces a new pipeline for processing omnidirectional images, improving traffic sign detection and recognition (TSDR) in real-world conditions. The system achieves over 95% accuracy in classifying 12 traffic sign types, overcoming common image distortions.
Area of Science:
- Computer Vision
- Robotics
- Artificial Intelligence
Background:
- Mobile omnidirectional cameras face challenges like object distortion and occlusion in real-world scenarios.
- Accurate traffic sign detection and recognition (TSDR) is crucial for Advanced Driver Assistance Systems (ADAS).
- Existing deep learning methods often rely on ideal conditions, not real-world 'in the wild' data.
Purpose of the Study:
- To develop and evaluate a novel processing pipeline for extracting objects from omnidirectional images captured in uncontrolled environments.
- To demonstrate the pipeline's effectiveness within a Traffic Sign Detection and Recognition (TDSR) system.
- To address limitations of current TSDR systems when applied to real-world data.
Main Methods:
- Implemented a new processing pipeline for omnidirectional images.
- Compared Mask RCNN, Cascade RCNN, and Hybrid Task Cascade (HTC) instance segmentation models.
- Utilized RsNeXt 101, Swin-S, and HRNetV2p backbones with transfer learning for localization and segmentation.
- Conducted a multinomial classification experiment on 12 traffic sign classes.
Main Results:
- The proposed pipeline successfully extracts objects from omnidirectional images.
- Achieved over 95% accuracy in classifying detected traffic signs across 12 categories.
- Demonstrated robustness against common issues like object distortion and partial occlusion.
Conclusions:
- The developed pipeline enhances the performance of TSDR systems using omnidirectional images in real-world conditions.
- This approach enables greater situational awareness by processing a single, all-encompassing image source.
- Provides a viable method for utilizing omnidirectional imagery in safety-critical applications like ADAS.
