In our latest study at VSI, we looked at what an autonomous vehicle needs to make its way safely and comfortably through changing lane conditions.
In our applied research at VSI, we’re continually testing and observing real-world use cases for self-driving technologies to help drive research and development in the automated vehicle industry.
In a previous article, we discussed the benefits of providing a high definition roadmap to a self-driving car. In that study on Adaptive Cruise Control, we looked at how an autonomous car handles the variable speeds needed to negotiate a complex turn…and does so without resulting in an uncomfortable experience for the passenger.
Comfort (along with safety and efficiency) for passengers is the goal of Advanced Driver-Assistance (ADAS) features. That comfort can come in the form of gradually adjusting speed as a car approaches, moves through, and exits the apex of a turn. It can also come in the form of not panicking the passengers as the vehicle encounters trouble with lanes.
Putting lane keeping to the test
The real-world is messy, and the road environment is a dynamic place. Lane markings can fade or be paved over. Roads get wet, muddy and snowy, resulting in conditions that even human drivers can find challenging.
Autonomous features must cope with these challenges – and how they can cope is the focus of our latest report. As with previous studies, our approach addresses the application of a high definition map to improve the safety and performance of the automated vehicle in the context of lane-keeping and trajectories.
We employed a single vehicle to perform a series of challenges focused on lane keeping. In the first test, the vehicle was controlled by computer vision only – meaning the cameras, sensors, and onboard software alone had to negotiate a set of lanes. In the second test, the same vehicle was controlled by a combination of the sensor input from the vehicle, and the HERE HD Live Map.
What we found, in brief, was that vehicles with computer-vision-only navigation systems have trouble operating when lane markings either change or disappear.
As an example, imagine driving in the left lane of a two-lane road approaching an intersection. As you near the intersection, your two lanes widen to three, as the left lane expands into a turn lane, while the center and right lanes continue forward. Your intention is to move straight through.
When we tested this exact scenario, the camera-only lane keeping setup continued to align itself with the center lane marker, bringing the car into the left-turn lane. While not the desired result for the test, this is a common occurrence in such vehicles, when they have difficulty determining where to be when lanes expand into a new configuration.
That change in configuration is part of the data provided by the HERE HD Live Map. With that data, the vehicle knows the location and the configuration of the oncoming intersection. It also has what is called a localization model. That model uses lane-markers, signs and roadside objects to understand its position on the road within a measure of centimeters – which is ultimately how it was able to succeed in this scenario.
This approach produced similar successes when we conducted tests under turning conditions, in lanes that expanded, and in areas where lane markers disappeared altogether. The ability for the vehicle to understand its position on the road by utilizing lane-markers as one factor among many to understand its position resulted in a better, more comfortable experience for our passengers.
More tools, better success
In the arena of self-driving cars, it’s common for the question of whether cars need maps to drive themselves. Ideally, should they not be equipped with the same sensorial and historical knowledge that a human driver has, and be sufficient enough to drive?
To date, all the testing we’ve conducted indicates that it’s certainly possible – it is not very practical. Adding a high definition map to the equation decreases the demand for processing, the needs of sensor inputs, and in cases like lane-keeping, the precise data needed to continue on when lane markers change unexpectedly.
If you’re interested in reading our full report, you can download it here.