Sažetak
Deep learning has become widely utilized and extensively studied in various fields such as autonomous vehicles and face recognition. However, its application in domains like land mine (LM) clearance and explosive objects clearance, particularly in non-military contexts, is relatively uncommon. This article builds upon prior research that employed the YOLO algorithm for detecting unexploded ordnance (UXO) in thermal images and extends it to the near real-time detection of annotated explosive objects in a thermal video sequence.
This work is based on the UXOTi_NPA dataset [1], which encompasses 11 distinct explosive targets, along with an original thermal video captured from a height of 3, 5, and 7 meters with very high ground sampling distance.
YOLO, known for its speed and accuracy, demonstrated the capability to detect explosive objects in over 40 frames per second (FPS), making it a viable solution for thermal videos operating at 25 FPS or 30 FPS. Currently, no automated systems are available for surface UXO detection using thermal video that can facilitate large-scale area surveys. Consequently, this research represents a significant step towards addressing this gap and paving the way for such automated solutions. Visible RGB imaging, picture, and video alike are thoroughly researched, proven by thousands of research results published on object detection, classification, recognition, and identification using various
machine learning algorithms, so these are vast pools of possible solutions.
Less researched because of the lack of publicly available datasets and more complex data processing are ground penetrating radar (GPR), magnetometer, and hyperspectral imaging sensors and future solutions would seek an application of deep learning algorithms for real-time detections with these sensors and their fusion for higher probabilities of detection.
Ključne riječi
YOLO UXO thermal video