Template-Based Feature Map Creation Using GPS and 2D LIDAR Scan Intensity and Height Extrema
Open Access
- Author:
- Kazandjian, Vahan
- Graduate Program:
- Mechanical Engineering (MS)
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- March 25, 2022
- Committee Members:
- Daniel Haworth, Professor in Charge/Director of Graduate Studies
Sean N Brennan, Thesis Advisor/Co-Advisor
Satadru Dey, Committee Member - Keywords:
- Feature Templates
LIDAR
Autonomous Vehicles
Mapping
GPS
Lane Detection
Vehicle
Road Detection
Data Clustering
Sensor Fusion - Abstract:
- Many existing methods for detecting lane geometries require the use of image data or LIDAR data features. Of the two sensors, LIDAR is the primary means of obtaining direct geometric information as well as reflectivity information. The challenge with using LIDAR data is understanding the relationships between features in each scan and creating a lane map that a vehicle can use to localize itself. The purpose of this study is to develop a robust method for aligning features in LIDAR scans, particularly aligning intensity and geometry both with each other and with repeated scans, even in scenarios where scan features, such as lane lines, are worn, occluded, or missing. This study also develops methods of templating features to produce scan-level maps of road features. To create a map of these complex lane topologies for use with a vehicle localization algorithm, optimal extrema filtering of LIDAR data is used to independently identify the position of lane and road boundary features in multiple lanes at once. Then, a filtering and averaging process is applied to these features to align LIDAR scans and estimate centerline positions for pairs of lane markings. By using the intersection of lane centerline predictions, the areas where lane features – lines, geometric extrema, etc. – cluster are stored to produce a template associated with a specific road location. When applied to data from a closed-loop test track, the algorithm presented in this thesis was found to identify lane markers with 9.8 mm 2-σ precision and height features with 35.6 mm 2-σ precision. Creating accurate maps lane features enables autonomous vehicles to operate in more complex situations with less need for human intervention. Path planning algorithms can use the a priori information about when and where lane topology changes to optimize maneuvers. In addition, maps of individual lane lines, road boundaries, and centerlines can be used to improve real-time vehicle localization in scenarios where road features are occluded, worn, or missing.