.png){kind=icon}
.png){kind=icon}
Return
DS-YOLOv7: Dense Small Object Detection Algorithm for UAV
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2024.3400613.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Modern unmanned aerial vehicle (UAV) reconnaissance is widely used. UAV equipped with FPV camera can realize the reconnaissance of dense small objects with the help of object detection technology. However, it is difficult to detect people or vehicle objects in UAV overhead images because the pixels are small and densely distributed. In this paper, we propose an improved YOLOv7 UAV dense small object detection algorithm (DS-YOLOv7) to improve the UAV detection performance of dense small objects. SFN technology optimizes the fusion network to further improve the semantic information and location information of dense small objects. LDSPP is an optimized module for feature extraction of small objects. It pays more attention to edge information of small objects and reduces the rate of missing detection. Dimensionality reduction techniques focus on reducing model parameters to facilitate the deployment of algorithms on lightweight devices. The experimental verification on the public data set VisDrone2019 shows that mAP50 and mAP50-95 have increased by 4.3% and 3.3% respectively, F1 scores have increased by 3.81%, model volume has decreased by 23.3MB, parameters have decreased by 13 million, and the improved algorithm is more conducive to UAV deployment and dense small object detection.
Keywords:
Feature extraction
Autonomous aerial vehicles
Convolutional neural networks
Prediction algorithms
Training
Head
Deep learning
YOLO
Dimensionality reduction
Reconnaissance
YOLOv7
SFN
LDSPP
dimensional reduction
dense small object