Friday 21 March 2025
The quest for perfect vehicle detection has long been a challenge in urban traffic surveillance. With the rise of autonomous vehicles and smart cities, accurate and efficient object detection is more crucial than ever. A team of researchers has proposed a novel approach that leverages deep learning techniques to improve vehicle detection in fisheye camera images.
Fisheye cameras, which provide a 360-degree view of their surroundings, are commonly used in urban traffic surveillance due to their ability to capture a wide field of view. However, this unique perspective also introduces challenges such as light glare from vehicles and street lights, shadows, non-linear distortion, and scaling issues for vehicles.
The proposed approach uses a modified YOLOv5 object detection scheme, which is a deep learning-oriented convolutional neural network (CNN) based on the popular You Only Look Once (YOLO) algorithm. The researchers introduced several modifications to enhance performance in fisheye images, including a light-weight day-night CNN classifier and upsampling techniques.
The team trained their model using a combination of datasets, including BDD-100K, Pseudo dataset, and Fish- Day/Fish-Night datasets. They also employed an ensemble approach to combine the predictions from multiple models, which improved overall performance.
Experimental results on the VIP Cup 2020 dataset showed significant improvements over existing methods. The proposed scheme achieved a mean average precision (mAP) of 13.7% higher than YOLOv5 alone, and outperformed other state-of-the-art object detection algorithms.
The researchers also addressed an issue in the original dataset known as ground truth inconsistency, where some vehicles near the image circumference were not annotated. They found that this problem was more prevalent in night-time images, which affected overall performance.
This study demonstrates the potential of deep learning techniques for improving vehicle detection in fisheye camera images. The proposed approach can be applied to various applications, such as autonomous vehicles, smart traffic management systems, and surveillance cameras.
The team’s findings highlight the importance of addressing specific challenges in fisheye image processing, such as light glare and non-linear distortion. By developing tailored solutions for these issues, researchers can improve the accuracy and efficiency of object detection algorithms in urban traffic surveillance.
As the world continues to urbanize, efficient and accurate vehicle detection will play a crucial role in ensuring road safety and reducing traffic congestion. The proposed approach offers a promising solution to this challenge, paving the way for further advancements in autonomous vehicles and smart city technologies.
Cite this article: “Deep Learning-Based Vehicle Detection in Fisheye Camera Images”, The Science Archive, 2025.
Vehicle Detection, Fisheye Cameras, Deep Learning, Object Detection, Urban Traffic Surveillance, Autonomous Vehicles, Smart Cities, Yolov5, Convolutional Neural Network, Mean Average Precision
