An improved YOLOv8s-based UAV target detection algorithm.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Aerospace Engineering

Background:

• The rapid growth of the low-altitude economy necessitates advanced Unmanned Aerial Vehicle (UAV) operations.
• UAVs require robust environmental perception and security measures for safe navigation in complex airspace.
• Existing target detection algorithms like YOLOv8s face limitations in multi-scale processing and small target detection for UAV applications.

Purpose of the Study:

• To develop an improved deep learning-based target detection algorithm for UAVs.
• To enhance detection accuracy and speed for autonomous UAV perception in the low-altitude economy.
• To address the limitations of YOLOv8s in multi-scale feature extraction and small target identification.

Main Methods:

• Introduced AKConv into the C2F module for adaptive convolution operations and efficient feature extraction.
• Integrated the LSKA mechanism into the SPPF module to improve small target feature extraction and long-range dependency capture.
• Proposed a novel Bi-SCDown-FPN feature pyramid network for accelerated and enriched feature fusion in the model's neck.

Main Results:

• The improved algorithm achieved a 5.9% increase in detection precision, 4.5% in detection recall, and 6.1% in mean average precision on the VisDrone2019 dataset.
• Reduced parameter count by 13.41% and weight file size by 13.33%, indicating model lightweighting.
• Demonstrated superior performance compared to other mainstream target detection algorithms.

Conclusions:

• The proposed algorithm offers a dual improvement in model lightweighting and detection accuracy for UAVs.
• The enhancements enable more efficient and accurate autonomous perception for drones in complex environments.
• This advancement supports the safe and orderly operation of UAVs within the burgeoning low-altitude economy.