SHARPEN - Scalable Highly Automated vehicles with Robust PErceptionN
| Reference number | |
| Coordinator | EMBEDL AB |
| Funding from Vinnova | SEK 8 887 082 |
| Project duration | April 2019 - June 2022 |
| Status | Completed |
| Venture | Traffic safety and automated vehicles -FFI |
| End-of-project report | 2018-05001eng.pdf (pdf, 4013 kB) |
Purpose and goal
The aim of the project was to improve today´s machine learning methods based on deep learning to become more robust in challenging environments, such as night, rain, snow and dirt on the sensors. To achieve this objective, both technologies to generate synthetic data, where one can control these variables, as well as the development of new methods, have been developed. The project has also focused on bringing these systems closer to production by compacting them to reduce their resource usage.
Expected results and effects
Our results show that you can improve today´s system with the use of synthetic data and also significantly reduce the cost of annotation of data. We have shown that 90% of the annotated data can be replaced with synthetic data. Furthermore, we have developed methods that can synthesize data from failed sensors in vehicles through machine learning and data from other sensors. We have also explored the best way of sensor fusion for object detection. We have also shown that with methods developed in the project we can reduce the latency of the same object detection system by over 60%.
Planned approach and implementation
We have designed an interface to the Carla Simulator with semi-automatic functionality to generate 3D worlds with challenging conditions. The vehicle sensor setup was based on the Kitty open database for the development of autonomous vehicles. Furthermore, we used Nvidias Jetson Xavier AGX as a hardware platform for accelerated deep learning. Organizationally, monthly meetings and additional technical meetings were held as needed. Coordinating partner at the start was Volvo GTT. This role was taken over by Embedl, who was added as a partner during the course of the project.