Driver Assistance and Self-Driving

Most new cars already contain machine learning that watches the road. A camera behind the windshield feeds a dedicated chip that runs vision models in real time (edge AI):

• Object detection: cars, pedestrians, cyclists, traffic signs.
• Segmentation and lane detection: where is the road, where are the lines?
• Depth and motion: how far away is the car ahead, and how fast is it closing?

From these, driver-assistance features act: automatic emergency braking, adaptive cruise control, lane keeping.

Levels of automation (SAE J3016) help separate assistance from autonomy:

• Levels 1 to 2: the car helps with steering and/or speed, but the human must watch the road at all times. Nearly all consumer systems are here, whatever their marketing names.
• Level 3: the car drives itself in limited conditions and asks the human to take over when needed.
• Level 4: no human driver needed within a defined area, as with robotaxi services in some cities.
• Level 5: drives anywhere a human could. Not available.

The hard part is the long tail: unusual situations that rarely appear in training data. Companies combine learned perception with rules, detailed maps, simulation and reinforcement learning research, and debate camera-only versus camera-plus-lidar designs. The most dangerous failure mode is human: drivers trusting a level 2 system as if it were level 4.