A UAV-Based Hybrid Human-AI Training for Wild-Animal Detection
Andrea Caruso · Frontiers in artificial intelligence and applications · 2025
Wild animal detection poses significant challenges in creating datasets suitable for training robust machine learning models, especially in dynamic environments. This paper describes the PhD research activity of the Author regarding a UAV-based Hybrid Human-AI System to address the limitations of AutoML in wild-animal detection by improving the variety and size of the dataset leveraging the human factor. The proposed approach combines hardware components, including collars worn by a subset of wild animals and drones equipped with onboard processing, with software frameworks that leverage Convolutional Neural Networks (CNNs) trained via both automated and human-involving pipelines. Collared animals contribute to generating large, targeted datasets, while human intervention expands data coverage to non-collared wildlife, improving diversity and generalization. To achieve an optimal balance between these methodologies, various configurations of collars, drones, and human inputs are evaluated using metrics such as dataset numerosity and diversity. Experimental validation, utilizing Arduino-like collars and DJI RTK350 drones, demonstrates system’s ability to generate enriched datasets for robust neural network training.