Vision-Based Detection of Low-Emission Sources in Suburban Areas Using Unmanned Aerial Vehicles

Area of Science:

• Environmental Science
• Computer Vision
• Robotics

Background:

• Detecting emission sources in urban areas is crucial for environmental monitoring.
• Traditional methods may struggle with complex low-rise building environments.
• Unmanned aerial vehicles (UAVs) offer a flexible platform for aerial surveillance.

Purpose of the Study:

• To develop and evaluate a method for detecting emission sources (smoke) over low-rise buildings using UAVs.
• To analyze the impact of different data acquisition scenarios on detection feasibility.
• To create an automated system for generating training data for deep learning classifiers.

Main Methods:

• Utilized stationary video sequences from a drone in a hover position with a nadir-facing camera.
• Employed differential frame information from stabilized video sequences.
• Integrated the YOLOv7 classifier for object detection and a convolutional neural network (CNN) for building roof detection, trained on a custom dataset.

Main Results:

• The proposed method demonstrated effectiveness in detecting smoke objects in both single images and video sequences.
• The YOLOv7 classifier showed robust performance, accurately recognizing objects even when initial assumptions (e.g., camera angle) were not strictly met.
• The system enables automatic generation of training datasets for deep neural networks.

Conclusions:

• The drone-based YOLOv7 method is effective for detecting emission sources in low-rise building areas.
• The approach shows resilience to variations in image acquisition parameters.
• The automated training data generation contributes to advancing deep learning applications in environmental monitoring.