A novel optimized tiny YOLOv3 algorithm for the identification of objects in the lawn environment.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection algorithms like YOLOv3 face accuracy limitations in complex environments such as lawns.
• The original tiny YOLOv3 algorithm exhibits insufficient feature extraction and struggles with small object detection due to its backbone architecture.

Purpose of the Study:

• To propose an Optimized tiny YOLOv3 algorithm that enhances accuracy and reduces computational load for lawn object detection.
• To address the limitations of the original tiny YOLOv3 algorithm in feature extraction and small object detection.

Main Methods:

• An enhancement module was developed to improve feature extraction in the shallow layers of the network.
• A multi-resolution fusion module was introduced to improve information interaction between deep and shallow layers and decrease computation.
• The Optimized tiny YOLOv3 algorithm was evaluated on a dataset including tree trunks, spherical trees, and people.

Main Results:

• The proposed algorithm demonstrated improved detection accuracy compared to the original tiny YOLOv3.
• The Optimized tiny YOLOv3 algorithm achieved a reduction in computational requirements.
• The enhanced model showed superior performance in detecting objects like tree trunks, spherical trees, and people in lawn environments.

Conclusions:

• The Optimized tiny YOLOv3 algorithm effectively addresses the accuracy and computational challenges of object detection in lawn environments.
• The integration of enhancement and multi-resolution fusion modules significantly boosts the algorithm's performance for real-world applications.