Real-Time Object Detection and Classification by UAV Equipped With SAR.

Area of Science:

• Remote Sensing and Surveillance
• Artificial Intelligence and Machine Learning
• Aerospace Engineering

Background:

• Real-time object detection and classification are crucial for modern surveillance and reconnaissance.
• Synthetic Aperture Radar (SAR) provides all-weather, day-and-night imaging capabilities essential for UAV-based operations.
• Existing methods often face challenges in balancing accuracy, computational complexity, and data transmission for onboard processing.

Purpose of the Study:

• To develop and evaluate advanced real-time object detection and classification methods for UAV-mounted SAR systems.
• To improve the accuracy and localization of detected objects compared to traditional algorithms.
• To assess the feasibility of onboard processing to reduce data load on ground control stations.

Main Methods:

• Implementation and testing of classic image analysis algorithms for SAR data.
• Application and evaluation of convolutional neural networks, specifically YOLOv5, for object detection.
• Development of a hybrid approach combining YOLOv5 with classic image analysis for enhanced post-processing.

Main Results:

• The novel hybrid system demonstrated significant improvements in both classification accuracy and object localization.
• Onboard implementation on a military-class UAV's mobile platform proved effective for online SAR image analysis.
• Low-computational complexity detection algorithms were shown to effectively reduce the size of SAR scans transmitted.

Conclusions:

• The integrated YOLOv5 and classic image analysis method offers a superior solution for real-time SAR object detection on UAVs.
• Onboard processing of SAR data using efficient algorithms is viable and beneficial for reducing communication bandwidth.
• This research advances the capabilities of autonomous aerial surveillance systems through AI-driven image analysis.