A lightweight UAV target detection algorithm based on improved YOLOv8s model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned Aerial Vehicles (UAVs) require efficient and lightweight models for target recognition due to limited computational resources.
• Existing complex models often fail to meet the stability and performance demands of UAV systems.
• There is a need for optimized object detection algorithms tailored for UAV deployment.

Purpose of the Study:

• To propose LW-YOLOv8, a novel lightweight object detection algorithm for UAVs.
• To enhance model efficiency and reduce computational cost without compromising detection accuracy.
• To enable practical and effective target recognition in UAV applications.

Main Methods:

• Developed a Cross Stage Partial Convolutional Neural Network (CNN) Transformer Fusion Net (CSP-CTFN) integrating CNNs and multi-head self-attention (MHSA) for global feature extraction.
• Introduced a Parameter Shared Convolution Head (PSC-Head) to improve detection efficiency and minimize model size.
• Replaced the original loss function with SIoU (Scalable Intersection over Union) to boost detection accuracy.

Main Results:

• LW-YOLOv8 achieved a 37.9% reduction in parameters, a 22.8% decrease in computational cost, and a 36.9% smaller model size compared to the baseline.
• The model demonstrated improvements in average precision (AP), AP50, and AP75 by 0.2%, 0.2%, and 0.4%, respectively.
• Experiments were conducted on the VisDrone2019 dataset, validating the model's effectiveness.

Conclusions:

• LW-YOLOv8 offers a significant improvement in lightweight design and efficiency for UAV target recognition.
• The proposed model effectively balances reduced resource consumption with enhanced detection performance.
• LW-YOLOv8 is well-suited for practical deployment on UAVs, addressing key challenges in real-world applications.