Deep CNNs (YOLO / SSD / Faster-RCNN-style)
- Use: High-accuracy object detection in ADAS and automated driving.
- Example: Modern ADAS stacks (Tesla, Waymo, etc.) use deep object-detection networks for vehicles, pedestrians and cyclists; surveys of lane/object detection for autonomous vehicles describe this as de facto standard. MDPI+2ScienceDirect+2
GANs (Generative Adversarial Networks)
- Use: Data augmentation, adverse-weather enhancement for perception, simulation.
- Example: GAN-based LiDAR and image translation between sunny and adverse weather in autonomous-driving R&D; used to train perception models more robustly, mainly in research and test platforms, not yet clearly advertised as a shipping feature. MDPI+1
Deep Reinforcement Learning
- Use: Decision-making for autonomous vehicles (trajectory planning, interaction with other road users).
- Status: Strong presence in autonomous-driving research and some pilots (e.g., fleets like Waymo/Cruise); OEMs don’t explicitly label production features as “deep RL” in public docs.
Transfer Learning
- Use: Re-using perception models across markets/vehicles and adapting to new regions or sensors, reducing training cost; widely used in internal pipelines, rarely marketed.
Capsule Networks / Self-Organising Maps (SOMs)
- Status: Experimental architectures for perception and clustering; appear mainly in academic prototypes for autonomous driving and driver-behaviour analysis, with no clearly documented production deployment yet.