Quantization effects on drones and unmanned aerial vehicals detection

Abstract

This work investigates the impact of neural network quantization on the detection of tiny unmanned aerial vehicles (UAVs), including drones. A YOLO-based object detection model is employed to compare the performance of FP32 and INT8 weight representations using the «Drone vs Bird dataset». The evaluation focuses on the detection of small objects, which represent a challenging scenario for modern computer vision systems. The Intersection over Union (IoU) and mean Average Precision (mAP) metrics are used to assess detection accuracy. Experimental evaluation is conducted on an RTX 5070 Ti GPU to analyze the trade-off between model precision and computational efficiency. The results of this study aim to provide insights into the applicability of quantized neural networks for UAV detection tasks, particularly in scenarios involving resource-constrained or edge computing environments