Solving the traffic light problem for autonomous vehicles is critical for all vehicle safety, not just autonomous vehicles. And unlike human-driven cars, AVs rely solely on their computer vision system and the data used to train them to navigate the world around us.
Currently, the best self-driving assistance systems incorrectly perceive something in their environment once every tens of thousands of hours. If that object is a traffic light, and the car gets it wrong, passengers, pedestrians, cyclists, etc., are all at risk. Here’s a look at the traffic light problem for autonomous cars from three perspectives.
1. Traffic lights tell road users when to stop and go, but they only work when everyone follows the rules.
Traffic lights are an engineered system, timed around traffic patterns. There are rules in place that tell cars when to stop and go, but the inherent human behavior of drivers, pedestrians and cyclists sometimes means these rules are loosely interpreted.
There are no physical barriers forcing road users to abide by traffic signals. They work because drivers follow the rules. AVs must also get the rules right, and that means countless hours of real-world exposure to the unspoken rules (or lack thereof) of the road.
The quality and accuracy of the data received from sensor packages must be precise, leaving no room for interpretation or inconsistencies. Also, license plates, faces or other personal identifying information (PII) may need to be anonymized to protect the privacy of people who might appear in the raw footage.
Ensuring AVs learn the right rules requires unbiased, appropriately labeled and high-quality training data based on a range of driving scenarios.
2. Traffic lights challenge both the vision system and the team developing the algorithms.
Because traffic lights are not a distance detection problem, AVs cannot use lidar or radar to navigate traffic signals. They must rely solely on their computer vision system to understand when to stop and go.
This can be difficult for both the vision system as well as the team developing the algorithm because the visibility of traffic lights may vary based on weather conditions like bright sunlight, rain, snow or fog. Similarly, not all intersections have traffic lights, so if the AV doesn’t detect one, that could be correct.
The treatment of out of service traffic signals can also vary. One city might use a plastic bag to indicate an out-of-service light while another city, or even county, might use masking tape or some other method to cover broken traffic signals.
Context clues like head nods or hand signals help human drivers manage low visibility or the lack of a traffic light, but since AVs cannot register this supplemental visual information, machine learning and computer vision engineers must train the AV on such scenarios as they arise.
3. Stock datasets aren’t enough to help AVs navigate traffic lights safely.
The volume and diversity of data required for autonomous vehicles is vast, and stock datasets cannot cover all use cases.
Duncan Curtis, Sama VP of Product shares approximately 1,000 to 10,000 images are needed to deploy a solution onto a vehicle, and in order to launch a product publicly, approximately 10,000 to 100,000 images are needed per ADAS feature. That’s a lot of data, and the amount of data needed to account for edge cases like traffic lights in unique road conditions is unknown.
Since AVs often need data from a specific camera or sensor package (because that’s what the car will use in production), generic stock data isn’t enough to help AVs navigate traffic lights safely. In addition to this, models need to be refreshed with ongoing training data as road rules, and the world around us changes.
Our team of data labeling experts annotate over 4M tasks per month, and for AV use cases, we’ve achieved 99.5 percent quality SLAs with more than 100 tags per image. We’ve also successfully achieved full productivity within our labeling teams in as little as two to five days.
If you need ground truth data for your ML model, connect with our team for a virtual consultation and demo of our industry-leading data annotation platform.