This system is intended to ingest sensor data and produce control inputs for the vehicle such that it is able to navigate unmarked roads without triggering a protective safety condition at a target velocity (typically the speed limit).
There are few, if any, unmarked roads in America that have a speed limit over 40 mph (at least in the more developed areas). If this project is expanded to consider other countries where high speed unmarked roads are present (like the UK), this constraint will have to be adjusted. 60 mph would likely be a reasonable upper bound for the system.
The perception time of an attentive human driver sits around 500 ms and includes making a decision based on the stimulus. Therefore, the total time through the system should be about 250ms.
The initial version of this system is targeting steady-state operation; i.e. no decision making due to traffic flow controls like stop/yield signs. Similarly, the initial version is not intending to handle cross-traffic activity. It is also not intending to detect that vehicles are stopped not due to traffic; e.g. with hazards on; and will not navigate around them. It is also not intended to maneuver around slow moving vehicles such as industrial & agricultural equipment.
The system consists of three modules:
The safety system is intended to be compliant with all existing Automated Driving System regulation. It is designed to use a model of the vehicle to create a swept volume that the vehicle would cover if it maintained its current heading and applied maximum braking force (after maximum processing time). If there is a sufficiently sized object within this swept volume, the vehicle will engage a protective stop. This needs to maintan a certain distance above the driving surface so that it is not triggered by bumps and other imperfections in the surface. The model then needs to account for vehicle motion and dynamically adjust the volume to prevent it from intersecting the driving surface erroneously. The system also needs to detect when the surface is too far/absent from below the swept volume to indicate a drop-off which may be indicitve of a road edge, a hole sufficiently large to be obstructive or dangerous to cross, or the top of a hill that cannot be safely detected over. This needs to be deterministic to be compliant as a safety system and may require mapping the surface topology (which would also require effective map localization) to better model the expected swept volume.
The perception system creates contextual understanding of the scene. This requires the sensor data, but also coarse mapping info such as number of travel lanes and parking lanes to accurately estimate lane boundaries. A simple road perception model demonstrated that this can be determined fairly easily, but crudely, as consitent color and limited range of surface slope change (gradient). Learned methods, like AI, produce a finer, more adaptive model and as this is not part of the safety system, there is no restriction on doing so.
The behavior takes the representation of the scene from the perception system and generates a path that it believes is unlikely to trigger the safety system. The training of this system generates many paths that are virtually run through to take into account the potential actions of other road users or lack of visibility around blind corners and over hills. This is also where the various traffic laws are enforced; e.g. some juristictions require 3'/1m of space between the vehicle and a cyclist or pedestrian. This model will likely need to be specific for regions with similar traffic laws or the laws will need to be codified into some sort of input into the model, which aproach it will be is undetermined at this time.