Improved GBS-YOLOv5 algorithm based on YOLOv5 applied to UAV intelligent traffic.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Traffic Management Systems

Background:

• Increasingly complex road traffic necessitates advanced management solutions.
• Unmanned Aerial Vehicles (UAVs) offer potential for traffic monitoring but struggle with small target detection accuracy.
• Existing YOLOv5 models face challenges with information loss in deep feature extraction and shallow feature utilization.

Purpose of the Study:

• To enhance the detection accuracy of Unmanned Aerial Vehicles (UAVs) for small targets in traffic management.
• To address limitations in the YOLOv5 model regarding small object detection and feature utilization.
• To introduce a novel algorithm, GBS-YOLOv5, optimized for UAV-based traffic surveillance.

Main Methods:

• Designed an efficient spatio-temporal interaction module to deepen feature extraction.
• Integrated a spatial pyramid convolution module to enhance small target information mining.
• Proposed a shallow bottleneck with recursive gated convolution for improved feature fusion and detail preservation.

Main Results:

• The GBS-YOLOv5 algorithm achieved an mAP@0.5 of 35.3% and mAP@0.5:0.95 of 20.0%.
• Demonstrated performance improvements of 4.0% and 3.5% in mAP@0.5 and mAP@0.5:0.95, respectively, over the standard YOLOv5.
• Validated the effectiveness of the proposed modules in improving small target detection for UAVs.

Conclusions:

• GBS-YOLOv5 significantly improves small target detection accuracy for UAVs in complex traffic scenarios.
• The novel architectural modifications effectively address information loss and enhance feature utilization in deep learning models.
• This algorithm presents a promising advancement for intelligent traffic management systems utilizing aerial surveillance.