Deep learning models for detection of explosive ordnance using autonomous robotic systems: trade-off between accuracy and real-time processing speed

The study focuses on deep learning models for real-time explosive ordnance detection (EO). This study aimed to evaluate and compare the performance of YOLOv8 and RT-DETR object detection models in terms of accuracy and speed for EO detection via autonomous robotic systems. The objectives are as follows: 1) conduct a comparative analysis of YOLOv8 and RT-DETR image processing models for explosive ordnance (EO) detection, focusing on accuracy and real-time processing speed;2) to explore the impact of different input image resolutions on model performance for identifying the optimal resolution for EO detection tasks;3) to analyze how object size (small, medium, large) affects detection efficiency for enhancing EO recognition accuracy; 4) to develop recommendations for EO detection model configurations; 5) to propose methods for enhancing EO detection model performance in complex environments. The following results were obtained. 1) The results of a comparative analysis of YOLOv8 and RT-DETR models for EO detection in the context of speed-accuracy trade-offs. 2) Recommendations for EO detection model configurations aimed at improving the efficiency of autonomous demining robotic systems, including optimal camera parameter selection. 3) Methods for improving EO detection model performance to increase its accuracy in complex environments, including synthetic data generation and confidence threshold tuning. Conclusions. The main contribution of this study is the results of a detailed evaluation of the YOLOv8 and RT-DETR models for real-time EO detection, helping to find trade-offs between the speed and accuracy of each model and emphasizing the need for special datasets and algorithm optimization to improve the reliability of EO detection in autonomous systems.