ADAS and Autonomous Vehicles
Deep Learning enables ADAS and Autonomous Vehicles that heavily rely on Object Detection, Semantic Segmentation, or Single-Shot Object Detection, e.g. Pedestrian Detection, Lane Detection, Traffic Light Detection, Traffic Sign Recognition, Out-of-Domain Object Detection, Geo Fencing, or Mobile Mapping. Our product Avant[‘guard’] enables next generation ADAS and Autonomous Vehicles that no longer require continuous supervision by the driver (SAE level 3 and 4). Avant[‘guard’] directly implements the goals of SOTIF (ISO/PAS 21448), enabling safety and resilience of Deep Learning applications for ADAS and Autonomous Vehicles.
Use Cases – ADAS and Autonomous Vehicles
Avant[‘guard’] will detect misclassification of the lane boundary, e.g. in case of a special road surface or a missing curb.
Avant[‘guard’] will detect misclassification of pedestrians, e.g. in case of a special pose like kneeling on the ground.
Traffic Light Detection
Avant[‘guard’] will detect misclassification of traffic lights, e.g. in case of an inactive traffic light due to power outage.
Traffic Sign Recognition
Avant[‘guard’] will detect misclassification of traffic signs, e.g. in case a traffic sign has unknown annotations, the writing is not readable, or in case of a special sign on private ground.
Out-of-Domain Object Detection
Avant[‘guard’] will detect misclassification of unknown objects like boxes on the road.
With Avant[‘guard’], ADAS Systems and Autonomous Vehicles can easily detect when leaving their Operational Design Domain (ODD).
Case Study – Out-of-Domain Object Detection
To test our technology, we set up a road scene with boxes on the road, which our reference Neural Network has never seen during training. (see below)
Our reference Neural Network has been trained on the Cityscapes dataset for Semantic Segmentation. It has classified the boxes as road surface, indicated by the purple pixel color, with a few pixels classified as road sign, indicated by the yellow color. (see below)
Our technology Avant[‘guard’] identifies the boxes on the road as an anomaly. Besides the crates, our technology identifies other suspicious objects, e.g. the Mercedes star of the ego vehicle, the stadium floodlights, a van in unfamiliar pose with open door, and some lens artifacts. (see below)