3D TRAFFIC RECONSTRUCTION FOR AUTONOMOUS VEHICLES WITH GAUSSIAN PROCESS LATENT VARIABLE MODEL (GPLVM)

• Main Author: Amirul Husna, Bryan
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/82361
• Institution: Institut Teknologi Bandung

Most of traffic accidents are caused by human error so autonomous vehicles will potentially reduce the number of accidents. Autonomous vehicles need to understand the surrounding traffic and that understanding can be expressed as a 3D visualization which makes the passengers feel safe. In this final project, traffic understanding and reconstruction are carried out by processing video recording from monocular camera installed on vehicles. Some processing steps to do the task are surrounding vehicles understanding, surrounding road understanding, and final 3D reconstruction. Vehicles detection is carried out by using pretrained YOLOv9 model which can detect various types of objects, but only objects with car type are processed further. Detected vehicles are associated and tracked further with SORT method which is based on Kalman Filter to create smooth movements. Road detection is implemented using classical image processing to detect road lane markings’ edges and Hough transform to generate line endpoints which represent the lanes. These endpoints are then associated and tracked between frames using SORT method. Vehicles’ shapes are reconstructed using GPLVM, based on the vehicles’ silhouette that has been segmented using YOLOv9-seg. These various steps and methods are then integrated to create a program that can reconstruct traffic in 3D. The program can estimate the states and shapes of the vehicles and road so that they can be placed correctly in a three-dimensional space. Several improvements that can be made are reducing the processing time, increasing the accuracy of various components, and integrating with other sensors/data sources.