It’s one thing to load cars with safety systems such as ADAS (advanced driver assistance systems), but as they get more sophisticated, it becomes more complicated to ensure they always do what’s expected.
So manufacturers are turning to the most advanced techniques to validate ADAS and make it as accurate as possible. Volvo is using virtual environments together with real-world testing for safety systems development, software training and validation.
Engineers can now scrutinise incident data collected from the advanced sensors of production cars more clearly using an imaging technique called ‘Gaussian splatting’. The new approach makes it possible to generate a large number of highly realistic 3D scenes from images taken in the real world.
These can then be viewed from different angles and the scene, quite literally, explored. Once it’s created, engineers can manipulate the virtual environment by modifying elements such as road users or obstacles and change the outcome of a scenario.
The approach makes it possible to subject safety software to a wide variety of traffic situations much faster and on a greater scale than before. It’s proving particularly useful for exposing safety systems to rare and sometimes dangerous ‘edge cases’ in days rather than months.
The virtual environments are developed through an in-house collaboration with Zenseact, an AI and software company founded by Volvo Cars. Porsche has also developed a new technique to improve its ADAS software.
Manual validation – involving a development engineer sitting in the passenger seat comparing what ADAS thinks a traffic sign reads with what it actually says – is becoming too labour-intensive.
If ADAS gets it wrong, the discrepancy is recorded on a data logger (a kind of hard drive) manually by the human co-driver, along with the actual vehicle speed data.
A vast amount of data is needed to weed out those rare instances that may trip ADAS up, but automated measuring equipment is expensive and using it across large-scale test fleets is unrealistic.
Porsche’s answer is the ComBox app, developed in-house, combined with an image-recognition app from start-up firm Peregrine.ai, both of which run on a smartphone.
If there’s a discrepancy between a road sign recognised by the Peregrine.ai app and what the vehicle’s ADAS ‘thinks’, the Peregrine.ai app photographs the sign and both the vehicle data and the picture are stored on the data logger.