Small target detection of floating objects in river channels based on improved YOLOv7.

Area of Science:

• Computer Vision
• Machine Learning
• Environmental Monitoring

Background:

• High frame rate requirements in computer vision for moving streams necessitate fast detection algorithms.
• Variability in bounding boxes across frames leads to low precision in small object detection.
• Detecting small, floating objects in dynamic environments like rivers presents a significant challenge.

Purpose of the Study:

• To introduce a novel Region-Overlap Detection (ROD) method for enhanced small target detection.
• To address the limitations of existing methods in detecting small objects in dynamic, flowing streams.
• To improve object detection precision and convergence speed using a modified YOLOv7 architecture.

Main Methods:

• Implementation of a Region-Overlap Detection (ROD) method integrated with a Minimum Convoluted YOLOv7 (MCY) architecture.
• Utilizing a typical YOLO classifier to identify the largest overlap area among multiple regions.
• Employing a secondary method to extract the largest bounding box from minimal convolution in the neural network's final training layer.
• Modifying YOLO architecture by removing intersecting areas from convolutional layers to expedite convergence.

Main Results:

• The proposed MCY-ROD method accurately identifies small objects in flowing streams with high mean accuracy.
• Achieved a mean Average Precision (mAP) of 73.1% for small floating object detection.
• Attained a recall rate of 70.2% in dynamic river environments, demonstrating robust performance.

Conclusions:

• The MCY-ROD method offers a significant advancement in small target detection within dynamic environments.
• The modifications to the YOLO architecture effectively expedite convergence and enhance detection accuracy (mAP).
• This approach provides a reliable solution for monitoring small floating objects in real-world riverine settings.