DIGITAL.CSIC. Repositorio Institucional del CSIC
Publication
Tesis doctoral (DoctoralThesis)
GRID-BASED PERCEPTION FRAMEWORK USING LIDAR SENSORS: A MULTI-REPRESENTATION APPROACH
In recent decades, autonomous driving research has received considerable attention due to its potential to improve people's lives by reducing traffic accidents and increasing travel efficiency.
As automated vehicles progress towards more realistic driving scenarios, they have encountered scenes with increasing complexity and unforeseen situations. Current efforts are primarily focused on enabling autonomous driving in urban scenarios, which present exceptionally complex challenges due to their uncontrolled nature, intricate maneuvers, and high density and diversity of objects.
Within the field of perception of the environment, a great variety of approaches have been proposed showing promising results. However, these strategies need further refinement to meet the stringent requirements for ensuring safe driving. Additionally, many of them focus on addressing specific tasks individually, wasting potential beneficial interactions with other complementary strategies and lacking a comprehensive environment description.
This Ph.D. thesis introduces a novel LiDAR-based perception framework for road user detection and occupancy estimation. It leverages four different environment representations to provide a more comprehensive description: point cloud, grid map, object-level description and lane grid. Firstly, LiDAR data is used to detect and precisely locate all the surrounding obstacles. Because the amount of data in LiDAR point clouds is vast, and many laser beams generally hit on the background, an obstacle-ground classification algorithm is applied. Subsequently, the resulting classified point cloud, along with other potential input data, such as information from digital maps or Vehicle-to-Everything communications, is modeled and fused into an occupancy grid map. Since driving scenarios change dynamically, this grid is extended to also estimate the velocity and type of the objects. This environment representation is highly valuable as it offers a general description of the surroundings, modeling all obstacles despite their type and shape, while also characterizing the non-occupied areas as free or unknown. However, there are certain aspects that are challenging to achieve using only this approach. Consequently, to obtain a more complete estimation of the road users, an object-level tracking strategy based on model assumptions is employed, thereby, obtaining a type-based shape description and a consistent ID over time. Furthermore, to achieve a more understandable description of the driving scene, a lane grid is also implemented, re-representing the information of the occupancy grid at a higher level and focusing only on the road lanes.
This thesis has been conducted under the activities of the AUTOPIA research group. This group possesses various autonomous vehicle prototypes with which the proposed perception framework has been extensively validated in real-world scenarios., This thesis has been partially supported by the Spanish Ministry of Science and Innovation
through the National Project COGDRIVE (DPI2017-86915-C3-1-R) and the Project PRYSTINE
(PCI2018-092928). It has also been supported by the Community of Madrid with SEGVAUTO
4.0-CM Programme (S2018-EMT-4362). Lastly, it has been supported by the European Commission
and ECSEL Joint Undertaking through the Project PRYSTINE (783190)., Peer reviewed
Project:
EC / H2020 783190
View at: http://hdl.handle.net/10261/351667