Deep multi-object tracking for ground truth trajectory estimation
|Funding from Vinnova||SEK 4 999 000|
|Project duration||July 2018 - December 2022|
|Venture||Traffic safety and automated vehicles -FFI|
Purpose and goal
The aim of this project is to develop algorithms that will provide high-precision estimates of the trajectories of all dynamic objects in the vicinity of the host vehicle. The purpose is to obtain an efficient technique to extract estimates that can be viewed as ground truth, which is of utmost importance for the development and verification of both perception and control modules. The intended focus is on off-line techniques and on investigation of combinations of deep learning and sensor fusion, to enable extraction of as much information as possible from the data.
Expected results and effects
The project is expected to have an effect on the development and verification of environmental perception algorithms for self-driving cars. This will be done by designing tools and strategies to automatically obtain accurate estimates of vehicle properties and trajectories. From an academic perspective, the project will establish a theoretical framework for estimation of multiple dynamic objects, reach conclusions on how to best combine deep learning techniques with statistical inference and prior information about, e.g., vehicle dynamics, and to a Ph.D. dissertation.
Planned approach and implementation
The project will be coordinated by Zenuity AB, by the coordinator Daniel Svensson, who works as a product owner within target tracking and object interaction. The majority of the work will be carried out by Chalmers University of Technology, primarily by a Ph.D. student. Supervision of the student will be done by postdoc Karl Granström and Associate Professor Lennart Svensson. Zenuity will support the project with technical expertise, participate in discussions of problem formulations and solutions, and take part in joint data collection activities.