SMILE II - Safety analysis and verification/validation of MachIne LEarning based systems

Purpose and goal

The goal of SMILE II is to develop and implement a safety cage concept to help automated vehicles´ perception systems to take reliable decisions. Perception systems are needed to make automation possible and they will to a large extent be based on deep neural networks (DNNs). These DNNs must be deemed safe enough to be deployed in the real world. That is why it is very important for vehicle OEMs to explore different methods for determining the safety of the networks. In SMILE II knowledge has been greatly increased in methods that can be used to verify and validate these deep learning systems.

Expected results and effects

The project have developed both algorithms, that implement safety cages, as well as two demonstrators that show the project results. Camera images are used as input for both demonstrators. An end-to-end driving algorithm was demonstrated in the VICTA Lab demonstrator and two different safety cage concepts was demonstrated in the Pro-SiVIC simulator. Further to the development of the demonstrator, the results were shared with potential customers and new partners for further development and new business opportunities. The methods was also tested in other domains such as healthcare.

Planned approach and implementation

In the beginning of the project a survey of NN-based architectures for object detection was made. The safety cage concept was explored and several approaches have been investigated. Two demonstrators were successfully setup on a PC with Nvidia GPU, however, further work is required to implement in an automotive hardware. During the project, the results from SMILE II has continuously been used as input to discussions on building pilots and products in autonomous driving. More work is needed but at e.g. Volvo it is considered a promising approach for a problem that at present is unsolved.