Research on Pedestrian Detection Model and Compression Technology for UAV Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Detecting small pedestrians in Unmanned Aerial Vehicle (UAV) images is challenging due to wide view angles and complex backgrounds, leading to missed or incorrect detections.
• Deep learning-based object detection models often require significant computational resources, limiting their practical application in real-time scenarios.

Purpose of the Study:

• To improve the accuracy of small pedestrian detection in UAV imagery.
• To reduce the computational resource consumption of object detection models for enhanced deployment.

Main Methods:

• An improved YOLOv5 object detection model was developed, incorporating a novel small object feature detection layer within the feature fusion layer.
• Channel pruning techniques were applied to compress the model, reducing memory and computational demands during inference.

Main Results:

• The improved YOLOv5 method achieved a 4.4% increase in average precision for pedestrian detection.
• Model compression via channel pruning resulted in a 11.9% reduction in GFLOPs (Giga Floating-point Operations Per Second) and a file size of 11.2 MB, maintaining inference accuracy.

Conclusions:

• The proposed enhanced YOLOv5 model significantly improves small pedestrian detection in UAV images.
• Model compression techniques enable efficient deployment of accurate object detection systems on resource-constrained platforms.