Inspection and intervention by drones in rescue operations have growing attention due to multiple causes, including natural and man-related events. Additionally, the rapid advancements in vision sensors, object detection models, and AI-based methods can boost the success of rescue scenarios. Empowering Search and Rescue through affordable and cheaper drone technology is the main motivation. Detecting the missing persons with drones is the key aspect in this context. Drone navigation involves object scale variations creating a computation load for the scene urge high-speed processing. To solve the two issues mentioned above, we propose the APH-YOLOv7t method that follows Holdout method. In this paper, we introduce a new Attention-based Prediction Head for YOLOv7-tiny. We also present the evaluation results of YOLOv7 the state-of-the-art convolutional neural network-based solution, here is used for robust object detection. In this context of drone navigation there is a need to perform detection of persons on land and sea surfaces allowing to reduce disaster, distress, identify and rescue them. Despite the higher success rate of object detection models, vision complexities make detection tasks on drone-captured images more challenging and this area remains under-explored. We used the existing three search and rescue datasets which are images acquired from drones specific to our objective. Results show that our APH-YOLOv7t method was the most robust attention-based YOLO and comprehensive person detection method for our application, demonstrating a consistently high level of performance in comparison to YOLOv7-tiny. Evaluation results on all three datasets are reported. With this solution, and conditional performance, we demonstrate to be able to satisfy our requirements of a mean average precision (mAP50) of over 0.80 for the person class and operational performance with over 125 fps on a single GPU Nvidia RTX2080Ti.
APH-YOLOv7t: A YOLO Attention Prediction Head for Search and Rescue with Drones
Lect. Notes in Networks, Syst.
Iberian Robotics conference ; 2023 ; Coimbra, Portugal November 22, 2023 - November 24, 2023
2024-04-27
13 pages
Article/Chapter (Book)
Electronic Resource
English