Probabilistic Semantic Mapping for Autonomous Driving in Urban Environments.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• High-definition (HD) maps are essential for self-driving car technology.
• The high cost of creating and maintaining HD maps is a significant challenge.
• Current methods rely heavily on expensive, manually updated map data.

Purpose of the Study:

• To develop an automated method for generating semantic HD maps.
• To reduce the cost and effort associated with HD map creation and maintenance.
• To fuse pre-built point cloud data with 2D images for landmark identification.

Main Methods:

• Utilizing semantic segmentation of 2D images to identify features like roads, sidewalks, and crosswalks.
• Associating semantic labels from images with points in pre-built point cloud maps.
• Employing a confusion matrix formulation to generate a probabilistic bird's-eye view semantic map.
• Implementing the pipeline within the Robot Operating System (ROS).

Main Results:

• Successfully generated a semantic map with road features in an urban environment.
• The approach accurately identifies static landmarks, providing rich environmental context.
• Demonstrated the fusion of image-based semantic segmentation and point cloud data.

Conclusions:

• The proposed method offers a cost-effective solution for generating and maintaining semantic HD maps.
• The generated maps enhance downstream tasks like trajectory generation and intent prediction.
• The approach has the potential for automatic generation of semantic features in HD maps.