A Model for Urban Environment Instance Segmentation with Data Fusion.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Fine-grained instance segmentation of urban environments is crucial for autonomous vehicle perception.
• Existing methods often struggle with differentiating similar objects and require significant computational resources.

Purpose of the Study:

• To develop a robust and efficient model for 3D urban environment instance segmentation using fused LiDAR and camera data.
• To enhance the reliability and accuracy of object classification and segmentation in complex urban scenes.

Main Methods:

• A dual fusion approach combining data-level and feature-level fusion of LiDAR point clouds and camera images.
• Integration of Markov Random Field (MRF) for optimizing Support Vector Machine (SVM) classification based on spatial context.
• Utilizing Mean Shift algorithm for object classification and instance segmentation.

Main Results:

• Achieved a classification accuracy of 99.3% for environmental objects.
• Successfully segmented individual objects of the same type without requiring instance labels.
• Demonstrated differentiation between similar but distinct objects through spatial contextual linkage.

Conclusions:

• The proposed dual fusion model offers a lightweight, efficient, and accurate solution for 3D urban environment instance segmentation.
• The method enhances autonomous vehicle perception by improving object classification and segmentation reliability.
• The model's efficiency makes it suitable for real-time applications in autonomous driving systems.