Multiview angle UAV infrared image simulation with segmented model and object detection for traffic surveillance

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Infrared (IR) imaging Unmanned Aerial Vehicle (UAV) technology is advancing rapidly, with applications in various fields.
• Obtaining real-world aerial images can be challenging due to flight limitations and costs, necessitating computer simulation.
• Existing simulation methods may lack the fidelity required for robust model training.

Purpose of the Study:

• To propose an improved method for simulating UAV infrared aerial images.
• To enhance the simulation quality of infrared images through 3D segmented model processing.
• To construct a comprehensive infrared traffic scene simulation dataset (IR-TSS) for object detection research.

Main Methods:

• An improved infrared aerial image simulation method based on open-source AirSim was developed.
• The simulation quality was enhanced using 3D segmented model processing.
• The Infrared Traffic Scene Simulation (IR-TSS) dataset was constructed, featuring seven object types.
• An efficient EfficientNCSP-Net was proposed and trained on the IR-TSS dataset.

Main Results:

• The improved simulation method generated high-quality infrared aerial images of traffic scenes from various viewpoints.
• The proposed EfficientNCSP-Net achieved a mean Average Precision (mAP50) of over 96% for object detection on the IR-TSS dataset.
• Comparative experiments demonstrated superior performance of EfficientNCSP-Net compared to existing methods.

Conclusions:

• The study presents a significant advancement in infrared aerial image simulation for traffic scenes.
• The developed IR-TSS dataset and EfficientNCSP-Net offer valuable resources for the research community.
• The findings have implications for improving object detection in aerial imagery and other simulation applications.