Improving long-tailed object detection on UAV platforms through adaptive sampling and lightweight model optimization

Abstract

Object detection models for unmanned aerial vehicle (UAV) surveillance are critical in emergency response and wildfire monitoring. However, these models frequently encounter two key challenges: class imbalance in training datasets and computational constraints on edge devices. This thesis addresses both problems by proposing and evaluating an adaptive sampling method, Exponentially Weighted Instance-Aware Repeat Factor Sampling (E-IRFS), and optimizing the trained models for real-time edge deployment. E-IRFS improves on previous rebalancing strategies by applying exponential scaling to the geometric mean of image and instance frequencies, thereby enhancing rare-class detection without overfitting. The method is evaluated using a dataset derived from the Fireman-UAV-RGBT collection and several public datasets containing fire, smoke, people, and water-related classes in aerial imagery. Results demonstrate improved mean average precision, especially for underrepresented classes, when compared to baseline sampling methods. In addition to model training, this thesis explores model quantization techniques for edge deployment on UAV hardware. The impact of quantization on detection performance, inference time, and model size is analyzed. Experimental results show that quantized models retain competitive accuracy while significantly reducing computational demands, making them suitable for real-time UAV operations. This work contributes a combined methodology for improving detection accuracy under class imbalance and enabling efficient deployment on resource-constrained UAV platforms.