Publicação
Next-generation 3D object detection and tracking for self-driving vehicles using object velocity
| Resumo: | Currently, a wide range of research and development activities are taking place in the field of autonomous vehicles, and many companies and organizations are working on developing and testing vehicles with various levels of automation. Autonomous vehicles require multiple sensors and advanced control systems to perceive the surrounding environment and make decisions without resorting to a human driver. LiDAR (Light Detection and Ranging), highly utilized in the development of autonomous vehicles, stands out for its potential to provide long-range highresolution 3D vision. Utilizing lasers and photodetectors to measure the time it takes for light pulses to bounce back, LiDAR produces precise and accurate 3D point clouds of a vehicle’s surrounding environment. The development of a secondgeneration LiDAR, coherent LiDAR, promises better reliability and the capability to directly measure relative radial velocity alongside distance for each point. This direct velocity measurement has the potential to revolutionize perception tasks in autonomous vehicles, including Multiple Object Tracking (MOT), paving the way for significant advancements in 3D perception. The fundamental motivation of this thesis is to investigate whether LiDAR-based 3D MOT can be improved by taking advantage of the radial velocity in each point of the point cloud provided by a coherent LiDAR. Due to the novelty of this sensor and the lack of a publicly-available MOT dataset with velocity information, a synthetic dataset was generated with a modified version of PreSIL, displaying 19 training and 20 testing sequences, with over 16k frames and 80k annotations. By adapting a state-of-the-art MOT model, it is was possible to conclude that, with velocity information, this model yields more accurate object proposals, at the cost of worse trajectory associations. These results were influenced by limitations of the data generation code, such as the lack of inertial data, problematic car models and poor performance in high inclination driving scenarios. When only taking into account sequences where these problems are absent, the proposed solution obtains improved performance when compared to the original implementation, showing that the additional velocity information has a positive effect on MOT tasks. This work thus points out that a second-generation LiDAR may indeed improve the perception systems, that are critical to the development of safe autonomous vehicles. |
|---|---|
| Autores principais: | Mendonça, Diogo Emanuel Nascimento |
| Assunto: | Autonomous driving Object tracking Coherent LiDAR Radial velocity |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Aveiro |
| Idioma: | inglês |
| Origem: | RIA - Repositório Institucional da Universidade de Aveiro |
| Resumo: | Currently, a wide range of research and development activities are taking place in the field of autonomous vehicles, and many companies and organizations are working on developing and testing vehicles with various levels of automation. Autonomous vehicles require multiple sensors and advanced control systems to perceive the surrounding environment and make decisions without resorting to a human driver. LiDAR (Light Detection and Ranging), highly utilized in the development of autonomous vehicles, stands out for its potential to provide long-range highresolution 3D vision. Utilizing lasers and photodetectors to measure the time it takes for light pulses to bounce back, LiDAR produces precise and accurate 3D point clouds of a vehicle’s surrounding environment. The development of a secondgeneration LiDAR, coherent LiDAR, promises better reliability and the capability to directly measure relative radial velocity alongside distance for each point. This direct velocity measurement has the potential to revolutionize perception tasks in autonomous vehicles, including Multiple Object Tracking (MOT), paving the way for significant advancements in 3D perception. The fundamental motivation of this thesis is to investigate whether LiDAR-based 3D MOT can be improved by taking advantage of the radial velocity in each point of the point cloud provided by a coherent LiDAR. Due to the novelty of this sensor and the lack of a publicly-available MOT dataset with velocity information, a synthetic dataset was generated with a modified version of PreSIL, displaying 19 training and 20 testing sequences, with over 16k frames and 80k annotations. By adapting a state-of-the-art MOT model, it is was possible to conclude that, with velocity information, this model yields more accurate object proposals, at the cost of worse trajectory associations. These results were influenced by limitations of the data generation code, such as the lack of inertial data, problematic car models and poor performance in high inclination driving scenarios. When only taking into account sequences where these problems are absent, the proposed solution obtains improved performance when compared to the original implementation, showing that the additional velocity information has a positive effect on MOT tasks. This work thus points out that a second-generation LiDAR may indeed improve the perception systems, that are critical to the development of safe autonomous vehicles. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.