YOLOv5 with ConvMixer Prediction Heads for Precise Object Detection in Drone Imagery.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Unmanned Aerial Vehicles (UAVs) offer unique mobility for object detection applications.
• Existing object detection methods face challenges with UAVs, including scale variation and motion blur.
• There is a need for improved UAV-based object detection techniques.

Purpose of the Study:

• To develop an enhanced object detection architecture for UAVs.
• To address challenges like varying object sizes and image blur in UAV imagery.
• To improve the accuracy and efficiency of object detection in real-world UAV applications.

Main Methods:

• A You Only Look Once v5 (YOLOv5)-like architecture was proposed.
• ConvMixers were integrated into the prediction heads.
• An additional prediction head was introduced to detect minutely-small objects.
• The model was trained and validated on the VisDrone 2021 dataset.

Main Results:

• The proposed architecture achieved performance comparable to state-of-the-art methods.
• The modifications effectively handled scale variation and potential blur in UAV images.
• The system demonstrated proficiency in detecting small objects from UAV perspectives.

Conclusions:

• The novel YOLOv5-like architecture with ConvMixers and an added head is effective for UAV object detection.
• This approach successfully tackles common challenges in aerial imagery analysis.
• The findings suggest a promising direction for advancing UAV-based surveillance and monitoring systems.