Publication
Impact of calibration in a LiDAR based on stereoscopic vision
| Summary: | Every year 1.3 million people die due to road accidents. Given that the main culprit is human error, autonomous driving is the path to avert and prevent these numbers. An autonomous vehicle must be able to perceive its surroundings, therefore requiring vision sensors. Of the many kinds of vision sensors available, the three main automotive vision sensors are cameras, RADAR and LiDAR. LiDARs have the unique capability of capturing a high-resolution point cloud, thus enabling 3D object detection. However, current LiDAR technology is still immature and expensive, which makes it unattractive to the automotive market. We propose an alternative LiDAR concept – the LiDART – that is able to generate a point cloud simply resorting to stereoscopic vision and dot projection. LiDART takes advantage of mass-produced components such as a dot pattern projector and a stereoscopic camera rig, thus inherently overcoming problems in cost and maturity. Nonetheless, LiDART has four key challenges: noise, correspondence, centroiding and calibration. This thesis focuses on the calibration aspects of LiDART and aims to investigate the systematic error introduced by standard calibration techniques. In this work, the quality of stereoscopic calibration was assessed both experimentally and numerically. The experimental validation consisted in assembling a prototype and calibrating it using standard calibration techniques for stereoscopic vision. Calibration quality was assessed by estimating the distance to a target. As for numerical assessment, a simulation tool was developed to cross-validate most experimental results. The obtained results show that standard calibration techniques result in a considerable systematic error, reaching 30% of the correct distance. Nonetheless, the estimated error depends monotonically on distance. Consequently, the systematic error can be significantly reduced if better calibration methods, specifically designed for the application at hand, are used in the future. |
|---|---|
| Main Authors: | Alegria, João Jesus de Sousa |
| Subject: | Autonomous vehicles Computer vision LiDAR Stereo imaging Calibration |
| Year: | 2022 |
| Country: | Portugal |
| Document type: | master thesis |
| Access type: | open access |
| Associated institution: | Universidade de Aveiro |
| Language: | English |
| Origin: | RIA - Repositório Institucional da Universidade de Aveiro |
| Summary: | Every year 1.3 million people die due to road accidents. Given that the main culprit is human error, autonomous driving is the path to avert and prevent these numbers. An autonomous vehicle must be able to perceive its surroundings, therefore requiring vision sensors. Of the many kinds of vision sensors available, the three main automotive vision sensors are cameras, RADAR and LiDAR. LiDARs have the unique capability of capturing a high-resolution point cloud, thus enabling 3D object detection. However, current LiDAR technology is still immature and expensive, which makes it unattractive to the automotive market. We propose an alternative LiDAR concept – the LiDART – that is able to generate a point cloud simply resorting to stereoscopic vision and dot projection. LiDART takes advantage of mass-produced components such as a dot pattern projector and a stereoscopic camera rig, thus inherently overcoming problems in cost and maturity. Nonetheless, LiDART has four key challenges: noise, correspondence, centroiding and calibration. This thesis focuses on the calibration aspects of LiDART and aims to investigate the systematic error introduced by standard calibration techniques. In this work, the quality of stereoscopic calibration was assessed both experimentally and numerically. The experimental validation consisted in assembling a prototype and calibrating it using standard calibration techniques for stereoscopic vision. Calibration quality was assessed by estimating the distance to a target. As for numerical assessment, a simulation tool was developed to cross-validate most experimental results. The obtained results show that standard calibration techniques result in a considerable systematic error, reaching 30% of the correct distance. Nonetheless, the estimated error depends monotonically on distance. Consequently, the systematic error can be significantly reduced if better calibration methods, specifically designed for the application at hand, are used in the future. |
|---|
Funded activities
No funded projects found.