COGNITION FOR CARS
Civil Maps provides cognition for autonomous vehicles, enabling them to crowdsource continental scale, 3D semantic maps for safe driving. With a highly scalable approach, we are creating a new generation of maps that enable fully self-driving cars to traverse any road safely and smoothly without any human intervention.
(Eyes + Cognition + Decisions + Actuation) = Driving
As humans when we drive, we use our senses to generate a huge amount of raw perception sensor data. This includes auditory, hearing, touch, smells and more while trying to operate a car. The raw data from our senses although massive, is still insufficient to actually make decisions. Therefore the cognition part of our brains does some pattern recognition and classifies the input signal into different categories to create a notion of context. These patterns and context are then provided to the decision making part of the brain.
The abstract reasoning and the probabilistic scoring of different options allows us to weigh our different options and choose an appropriate state within a state estimation model. When we don’t have the ability to recall context from a previous trip, our probability of entering a non-ideal state within a state estimation model is higher. When we can remember and anticipate things, the ease of driving increases and we use less brain cycles to make a decision.
The ability to anticipate and remember has three modules.
- The ability to understand how to operate within a particular environment and continuous monitoring the intentions of other vehicles, pedestrians, bicyclists allows cars to anticipate and drive defensively. This category is mainly focused on moving objects.
- The ability to understand what’s changed from what you previously remember should trigger the car to reduce its operating envelope. Once the discrepancy is found, it should be reported to other vehicles so they can benefit from the knowledge and plan accordingly. This category is focused on finding what has changed since the last arrival.
- The ability to find objects in a scene, classify them, index them and package the information into a contextual layer is important for increasing the simplicity and robustness of the state estimation models in the car. This category is focused on indexing stateful information about a 3d scene so cars can recall the information to drive with more confidence.
Once the context is created for the moving objects, temporary fixtures, and permanent infrastructure, this information may be shared with the decision engine along with the position and orientation of the vehicle.