DFTD-YOLO: Lightweight Multi-Target Detection From Unmanned Aerial Vehicle Viewpoints

Due to the low detection accuracy of small and dense target objects in multi-target detection tasks from the unmanned aerial vehicle (UAV) perspective and the deployment of deep learning models for UAVs as embedded devices, these models must be lightweight. In this study, we propose an improved algorithm, DFTD-YOLO, based on YOLOv8n. We designed a new neck feature fusion network. The network better balances information transfer between shallow and deep layers through a detailed information extraction module and an abstract feature information aggregation module, effectively reducing the loss of detail information with gradient flow and improving detection performance. In addition, we designed a new detection head called the TDD-Head. This module enhances the feature interaction between the classification and regression tasks through the task alignment mechanism and shared convolution, which reduces model parameters and computation and improves model performance. To validate the model, we conducted validation experiments on the VisDrone2021 dataset. The experimental results showed a 33.67% reduction in the number of parameters, 17.3% reduction in the amount of computation, 10.74% improvement in mAP@0.5, and 13.2% improvement in mAP@0.5:0.95 compared with the existing YOLOv8n. The results demonstrate the considerable potential of the model for multitarget detection tasks from the UAV perspective.