3D Perception and Prediction of Pedestrians for Improved Decision-Taking in Autonomous Driving and Active Safety

Driverless vehicles and advanced safety systems are reliant on being able to perceive, interpret and understand the 3D scene surrounding the vehicle. In this project, we will develop algorithms for perception, prediction, and control. By joining our competences, new algorithms targeting the joined performance on system level, instead of optimizing each algorithm individually. More specifically, algorithms for improving the state-of-the-art in 3D perception and prediction of human behaviour from cameras mounted on a moving vehicle developed in combination with the decision algorithms controlling the vehicle so their combined performance is optimized.

In recent years, the performance of computer vision applications, like image classification, object recognition and detection has dramatically improved due to the Deep Learning paradigm. The breakthrough has been enabled by theoretical developments, increased computational power and big annotated datasets. Still, the challenges posed by a fully autonomous vehicle due to the dynamic environment is enormous and yet to be solved. In this project, we will leverage on recent developments and consider a key problem in traffic safety: 3D modelling of human behaviour in traffic scenes.

The overall purpose of the project is to develop algorithms combining perception, prediction, and control to optimize system level performance. Deep learning algorithms in the vision system will be developed in connection to dynamic models of the pedestrians, where the uncertainty of the predictions will be crucial for the control performance. A large part of the research problem will be to integrate the algorithms so that the over-all performance can be well described. More specific research questions will be to evaluate how the pedestrians should be modelled so that the uncertainty relevant for control algorithms is obtained.

Our objective is to be able to better perform combined predict and decision taking to improve performance of autonomous driving without compromising safety. For active safety systems the algorithms will make earlier intervention possible, which improves the performance of these systems as well.

Start date 01/09/2018
End date The project is closed: 31/08/2019

Page manager Published: Thu 25 Oct 2018.