A Lightweight Model for Real-time Detection of Small Objects in UAV Aerial Imagery on Embedded Platforms

Abstract

A lightweight small object detection model, LD-YOLO, is proposed for UAV aerial images under embedded-platform constraints. LD-YOLO is built around a unified two-level design: at the module level, an Adaptive Multi-scale GSConv (AMGSConv) is designed, which splits the input into an identity branch and 3×3/5×5 depthwise branches, generates channel-level weights for adaptive receptive-field selection, and uses channel concatenation with Channel Shuffle for cross-branch interaction; at the architecture level, VoV-GSCSP is reconstructed with AMGSConv to form VoV-AMGSCSP, which is embedded into a P2–P5 four-scale bidirectional feature fusion path to build the AMGS-Neck and the overall LD-YOLO architecture. Experiments on the VisDrone-DET validation set show that LD-YOLO achieves mAP@0.5 and mAP@0.5:0.95 of 34.7% and 20.6%, improving over YOLOv8n by 6.1 and 4.3 percentage points; the model has 2.35 M parameters and 6.3 GFLOPs, and reaches 37 FPS FP16 inference on Jetson Nano.