Multimodal Unmanned Aerial Vehicles Classification Using Vision and Acoustic Inputs

Unmanned Aerial Vehicles (UAVs) are started to use in military zones during wars, which creates fears regarding safety, privacy, and national security. In order to track their utilization effectively, we require effective mechanisms to identify and classify UAVs. Classical camera-based vision detection is poor in low light, fog, as well as under vision obstacles. It also subsists with vision since the UAVs emanate distinguishable sound patterns used for identification. A framework is proposed for the classification of UAVs that incorporates both vision as well as audio data. We worked on the dataset Roboflow, comprised of five classes: drone, helicopter, aircraft jet, bird, and plane. We worked on the YOLOv8 [1], ResNet18 CNN, as well as the Vision Transformer. ResNet18 CNN obtained the best visual accuracy, reaching 98.44%, followed by satisfactory results from YOLOv8 and ViT. We utilised the Drone AudioDataset for audio analysis. It comprised the CNN, the LSTM, as well as the Vision Transformer, where the ViT achieved the best precision, 95.81%. The results show that combining optical and acoustic techniques improves detection reliability, especially in challenging real-time conditions. Ongoing research will be directed towards building derivative fusion techniques to construct variable, accurate, and secure UAV surveillance platforms.