Drone detection methods and comparison metrics for audio and vision modalities

Drones are unmanned aerial vehicles with a significantly growing importance in science, technology, and defense systems. However, their existence and detection a complex issues in terms of risky situations and compromise security. In this thesis, a multi-modal drone detection system has been designed and evaluated using a real experimental sample. The drone detection procedure has two modality approaches: audio and optical modalities. While Yolov8, Long-short-term memory (LSTM), and Vision transformers are applied and evaluated for optic modality, Convolutional neural network (CNN) and LSTM are evaluated and applied for audio modality. After that, a fusion approach for the two modalities is evaluated and interpreted with the affection of different levels of darkness conditions (20, 40, 60, and 80 % of darkness respectively.) Experimental results indicate that precision-recall curves and ROC curves of different models for both modalities. The finding of results reflects insights for further research approaches in terms of the importance of the study with regard to using another modality or thermal infrared imaging. It also reflects the total prediction increase after fusion and darkness novelties. The challenge becomes when adding darkness to vision samples classifying the final detection results of modalities for each frame or sound snippet. The final results demonstrate that using logical methods different levels of samples from different darkness levels might affect the overall performance of probability detection for vision and audio modalities.