PhD Thesis

Lidar-Based Scene Understanding for Autonomous Driving Using Deep Learning

Work default illustration

Information

  • Started: 01/07/2014
  • Finished: 25/02/2020

Description

With over 1.35 million fatalities related to traffic accidents worldwide, autonomous driving was foreseen at the beginning of this century as a feasible solution to improve security in our roads. Nevertheless, it is meant to disrupt our transportation paradigm, allowing to reduce congestion, pollution, and costs, while increasing the accessibility, efficiency, and reliability of the transportation for both people and goods. Although some advances have gradually been transferred into commercial vehicles in the way of Advanced Driving Assistance Systems (ADAS) such as adaptive cruise control, blind spot detection or automatic parking, however, the technology is far from mature. A full understanding of the scene is necessary in order to allow the vehicles to be aware of their surroundings and the existing elements on the scene, as well as their respective motions, intentions and interactions.

In this PhD dissertation, we explore new approaches for understanding driving scenes from 3D LiDAR point clouds by using Deep Learning methods. To this end, in Part I we analyze the scene from a static perspective using independent frames to detect the neighboring vehicles. Next, in Part II we develop new ways for understanding the dynamics of the scene. Finally, in Part III we apply all the developed methods to accomplish higher level challenges such as segmenting moving obstacles while obtaining their rigid motion vector over the ground.

More specifically, in Chapter 2 we develop a 3D vehicle detection pipeline based on a multi-branch deep-learning architecture and propose a Front (FR-V) and a Bird’s Eye view (BE-V) as 2D representations of the 3D point cloud to serve as input for training our models. Later on, in Chapter 3 we apply and further test this method on two real uses-cases, for pre-filtering moving obstacles while creating maps to better localize ourselves on subsequent days, as well as for vehicle tracking. From the dynamic perspective, in Chapter 4 we learn from the 3D point cloud a novel dynamic feature that resembles optical flow from RGB images. For that, we develop a new approach to leverage RGB optical flow as pseudo ground truth for training purposes but allowing the use of only 3D LiDAR data at inference time. Additionally, in Chapter 5 we explore the benefits of combining classification and regression learning problems to face the optical flow estimation task in a joint coarse-and-fine manner. Lastly, in Chapter 6 we gather the previous methods and demonstrate that with these independent tasks we can guide the learning of challenging higher-level problems such as segmentation and motion estimation of moving vehicles from our own moving perspective.

The work is under the scope of the following projects:

  • Cargo-ANTS: Cargo handling by Automated Next generation Transportation Systems for ports and terminals (web)
  • ColRobTransp: Colaboración robots-humanos para el transporte de productos en zonas urbanas (web)
  • ROBOCOM++: Rethinking robotics for the robot companion of the future (web)