Real-time Semantic Segmentation and Grid-space Optimization for Sparse Point Cloud

다국어 초록 (Multilingual Abstract)

The identification of granular materials is a crucial ability for automating construction machinery. Current point clouds segmentation methods struggle to segment shapeless objects, such as sand piles, due to sparse features. A real-time semantic sparse point cloud segmentation framework based on multi-view method is developed using a multi-sensor fusion SLAM algorithm with visual odometry. This approach has inherent semantic mapping errors caused by external calibration, dynamic pose estimation, and image segmentation. To enhance segmentation accuracy, a grid-space optimization algorithm has been proposed. The first step involves finding the incorrect segmented points by checking the pixel-depth gradient in grid space. Secondly, depth density clustering are applied to these points for re-segmentation. Our algorithm was tested on gravel and sand piles in construction scenarios. The experimental results demonstrated that our segmentation strategy can effectively segment granular objects. Furthermore, our two-step scanning and re-segmentation methods can significantly improve the performance of point cloud semantic segmentation.

목차 (Table of Contents)

• Abstract
• 1. INTRODUCTION
• 2. METHODOLOGY
• 3. EXPERIMENTS
• 4. CONCLUSION AND FUTUREWORK
• REFERENCES