Slicing-Tracking-Detection: Simultaneous Multi-Cylinder Detection From Large-Scale and Complex Point Clouds.

Area of Science:

• Computer Vision
• Geometric Modeling
• Robotics

Background:

• Detecting multiple cylinders in large-scale, complex point clouds is a persistent challenge in 3D data processing.
• Existing methods often struggle with efficiency and accuracy, especially in industrial settings.

Purpose of the Study:

• To propose a novel framework, slicing-tracking-detection (STD), for accurate and simultaneous detection of multiple cylinders.
• To address the limitations of current approaches in handling large-scale and complex point cloud data.

Main Methods:

• Reformulated 3D cylinder detection as a multi-object tracking (MOT) task.
• Generated slice sequences from point clouds and modeled cylinder cycles using Markov Decision Processes (MDP).
• Employed reinforcement learning for tracking cylinder components and handling missed detections.

Main Results:

• The proposed STD framework demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant improvements in efficiency, accuracy, and robustness for multiple cylinder detection.
• Successfully detected multiple cylinders simultaneously in large-scale and complex process plant point clouds.

Conclusions:

• The STD framework offers a robust and efficient solution for multiple cylinder detection in challenging 3D point cloud datasets.
• The novel approach of combining slicing, tracking, and reinforcement learning provides a significant advancement in the field.