First-time exhibitor SafeAD, a startup spin-off from the Karlsruhe Institute of Technology, was at ADAS & Autonomous Vehicle Technology Expo this week (13-15 June) to explain how it leverages AI for human-like scene understanding, using data from automotive-grade full surround cameras, lidar and radar to detect and identify 3D objects and the road layout. Its neural networks are also able to infer road elements, topology and more to create centimeter accurate high-definition maps, complete with continuous change detection to trigger automatic map updates on the cloud.
“We’re developing a 3D multisensor perception system for use in vehicles and consumer cars,” explained Dr Niels Ole Salscheider, CTO & co-founder, SafeAD. “We use a neural network to fuse the raw sensor data of cameras, lidar and radar to do object detection, tracking and prediction up to 300-400m. We can also infer local HD maps just from [camera] sensor data – in other words, you can do mapless driving in a Level 2+ system, or you can use our technology as a fallback or validation system for Level 3 or Level 4 vehicles,” he continues.
“If you want to train a neural network, you need 3D scenes as training material. For us, it was very important to be able to use only your fleet sensor data to create this training data as it is the fleet that finds the corner cases you need to make a robust system. That’s why we have also developed a mapping approach that works using only automotive cameras. We use these to create a 3D reconstruction of the scene, also including dynamic objects, then we label the scene with our AI model, and then we have human annotators fix the last remaining parts. We have had a lot of interest from simulation companies who buy the scenarios [available in open scenario format]or maps and integrate them into their simulation.”
The company has already had enquiries from customers in the off-highway sector, and has come to Stuttgart to meet OEMs and Tier1s in the automotive sector. “We have really had a lot of interest, especially given that we are such a small company,” said Salscheider. “All the major simulation companies have come to see our maps, and we have talked to three or four Tier1s and OEMs, so it has been perfect for us.”
