VISUAL LOCALIZATION OF AUTONOMOUS VEHICLES ON INTERSECTIONS BASED ON HD MAP

• Main Author: Shafwah Utsula, Bizza
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75395
• Institution: Institut Teknologi Bandung

The increasing number of autonomous vehicles will result in a safety factor that must be maintained. One of the crucial prerequisites for this is proper localization, especially at intersections where most road accidents occur each year. For this reason, this research aims to improve the localization system, especially at road intersections. A new approach is introduced to improve the localization process, which has drawbacks such as the inaccurate Global Positioning System (GPS) when satellite signals are blocked, Light Detection and Ranging (LiDAR), which is expensive and heavy to compute, and ordinary cameras, which are very sensitive to changes in visual conditions. This is done by implementing the High Definition (HD) mapping technology which determines the vehicle's position and depicts its surrounding environment. This HD Map relies on rough road-level images and poses estimates, which are used to retrieve intersection data. The proposed system processes images from a monocular camera into a semantic segmentation feature map at the pixel level. The LiDAR point cloud data is then used to reconstruct a 3D intersection model, which is then segmented semantically so that position data is obtained after merging using the occupancy grid and scoring function. This semantic segmentation-based approach localization produces a considerable error value between 30-80 meters compared to GPS data as ground truth and an average of 69.9% Mean Intersection over Union (MIoU) value with processing time less than 4 hours compared to other methods with more than 8 hours. Thus, the proposed approach can save computational costs.