Drone Detection using Deep Learning

Abstract: Drone intrusions have been reported more frequently these years as drones become more accessible in the market. The abuse of drones puts threats to public and individual safety and privacy. Traditional anti-drone systems use radio-frequency sensors widely to get the position of drones. In this thesis, deep-learning-based detection algorithms on surveillance cameras have been investigated to be integrated into the RF anti-drone system. The objective of the thesis is to evaluate state-of-the-art models and training strategies for drone detection. The main challenges in this thesis were detecting small drone targets at long distances and running the model in real-time. It is difficult to find a publicly available dataset of small drones online, so a real-world small drone dataset was constructed and used in this thesis. Different versions of YOLO were compared and tested on the real-world dataset. Modifications on the detection heads of the models were conducted to examine their effects on small object detection. The method of tiling on datasets was also adopted to help with the detection of small drones. Images from different sources were trained and added to compare with the model trained with only one source. Bird images were added to the training dataset in different ways for reducing the false positives when birds were included in the test set. In conclusion, YOLOv5n and YOLOv5m overall yielded the best results in precision, recall, and inference speed. The additional detection head on shallow layers at small scales can improve the precision at the cost of recall and inference time. However, it was not effective when the objects were extremely small. The tiling approach yielded the most effective improvement in increasing the recall value of the model. Adding bird images into the training dataset as background had better precision and recall value than adding annotated birds as a separate class in training. Training without extremely small drones and birds helped improved the precision metric, while the recall value would decrease. Further study is required to examine the generality of the results in all types of drones.