A Deep Learning Approach to Classify Drones and Birds
S. Sethu Selvi, S Pavithraa, R S Dharini, E Chaitra
Abstract
Owing to the rising threat of employing drones in unlawful acts, detection of drones has recently gained tremendous importance in the field of security and surveillance. Artificial neural networks do not support real time object detection as it requires multiple GPUs for training the models. Deep learning architectures try to resolve this issue by designing convolutional neural networks (CNN) that can operate in real time requiring only one traditional GPU for training. In this paper, suitable deep learning architectures are utilized for detection of drones and birds. You only look once (YOLO) based algorithms are considered for this application which are one-stage approaches that looks at an image only once and has a single neural network which is trained end-to-end, to directly provide the bounding box along with the class label and probability of detection. The models are trained on a custom dataset consisting of 664 drone images and 236 bird images. Simulation results of these models have shown that YOLOv4 and YOLOv5 achieved a F1-score of 98% and 94% with a detection speed of 54fps and 77fps respectively providing a mean average precision (mAP) of 97.4% and 95%. The finding is that YOLOv4 outperformed YOLOv5 in terms of mAP, while YOLOv5 outperformed YOLOv4 in terms of detection speed.