DRNet: A Miniature and Resource- Efficient MAV Detector

This article focuses on the challenge of accurate microaerial vehicle (MAV) detection under memory-constrained and lower computational cost conditions. A miniature MAV detection framework, DRNet, is proposed to address this issue. DRNet incorporates a dual skip concatenation (DSC) network and squeezing excitation residual (SER) networks for feature extraction, which enhances the model’s representation capacity. Additionally, spatial attention computation improves object localization and representation capabilities. A lightweight network facilitates efficient feature fusion, further optimizing DRNet’s performance. DRNet’s superior performance over other methods is validated across four challenging datasets. DRNet achieves a comparable accuracy-matching heavyweight method while saving 99.97% parameters, and a compact model size of just 309 kB makes it the smallest high-accuracy, low-computational-requirement MAV detector to date. Furthermore, DRNet can reduce computational costs by 95.3% and GPU memory usage by 50% when processing high-resolution images with dimensions of $1280 \times 1280$ . Challenging real-world tests and experimental deployments on edge-computing devices further confirm DRNet’s feasibility and portability.

I. Introduction

Vision-based microaerial vehicle (MAV) detection and confidence measurement are crucial in various applications, such as flight safety, collision avoidance, autonomous navigation in air transportation, multi-MAV visual formation, and privacy protection [1]. The ability to detect small targets such as MAVs is a critical performance metric for assessing detection methods [2]. Existing advanced detectors require significant computational resources and memory footprint, challenging their implementation on memory-constrained edge-computing devices, while low-power consumption detectors typically suffer from insufficient accuracy. A compact MAV detection framework that ensures efficient and accurate detection with reduced computational overhead and memory usage can facilitate its application in edge-oriented microdeep learning systems.