August 14, 2019
You may assume safety and innovation don’t go hand-in-hand. Safety often implies letting technology stand the test of time while making only small changes along the way. Yet as our roads and cities become increasingly congested and complex, only radical innovation can bring a new level of safety that’s needed to navigate these dense urban environments.
Many companies have already put autonomous vehicles on public roads. Waymo has completed more than 10 million driverless miles and last year drive.ai launched a public shuttle service, but the question of safety and public trust is still the most urgent one to be answered.
Continental AG has been manufacturing automotive equipment for nearly one hundred and fifty years and recently adopted a new goal, called Vision Zero, which strives for zero fatalities or accidents on roads. This is truly radical.
To achieve this ambitious goal Continental is working on a framework they describe as SensePlanAct to power the next generation of autonomous vehicles:
- Sense is accomplished by a variety of cameras, radars, and light sensors that are continuously gathering data 360 degrees around the vehicle.
- Plan takes the outputs of sensing step plus weather, traffic patterns, light conditions and other factors into account determining what action the vehicle should take next.
- Finally, Act implements the planned course of action, for example, changing lanes or applying emergency braking.
Sounds straightforward enough but there are many challenges to implementing this type of technology. Modern urban environments present a complex mix of buildings, infrastructure, busses, trains, cars and pedestrians - to name a few. Roads and traffic rules vary by country and signs are often in different languages or location specific. Weather conditions, such as rain or snow, and time of day, for example dusk or night time, further complicate the environment vehicles have to interpret and plan for.
As Robert Thiel, the Head of Machine Learning and Test Data Management at Continental, explains: “What we need to do is extract all this knowledge, all these situations out of data we collect all over the world and bring that in to our products using deep learning and machine learning technologies”.The more data you gather across all of these variables the better your autonomous driving software can become.
Continental recognized that it was no longer sufficient for its engineers to implement rule-based algorithms they knew to be accurate in the past. Also, it is no longer possible to test every scenario and combination of variables for safe driving in modern cities. So Continental took a new approach beginning to collect lots and lots of data and adopting deep learning to guide the development of the next generation of their sensors and software.
This is where MongoDB comes in. In order to collect, store and process massive volumes of complex data Continental needed an intelligent data platform that is flexible, developer friendly, and seamlessly scalable. Data generated by sensors, radars, and cameras is complex, multi-structured and changes rapidly based on new configurations or prototypes. Machine learning frameworks rely on iterative feature engineering based on this evolving data to train and tune new models.
Restricting the data to a rigid tabular schema, like those used in traditional relational databases, isn’t practical because data scientists do not know up front how each data element will be used in the next model iteration, and the one after that, and so on.
This type of approach was limiting the velocity with which Continental could innovate and so they moved from SQL-based tools to MongoDB to build their deep learning framework. Originally Continental only planned to use MongoDB to store and label image data, such as scenes from the road. But the team quickly found that they can use the same database for the analytical image data, the derived metadata and the results of their experiments, significantly increasing productivity.
Powered by Data and Collaboration
Continental also found that in order to collaborate effectively their teams needed to share common data sets and platform. In the past even if developers started with a common data, it soon diverged, making it impossible to calibrate results and merge individual workstreams. With MongoDB’s flexibility and parallelism, developers can build new models more rapidly, work together without sacrificing speed and accuracy, and quickly build and test new prototypes for the autonomous SensePlanAct framework. “In the end we were able to tame this deep learning beast with this flexible database”, says Martin Berchtold-Buschle, who is the subject matter expert for Big Data Infrastructure & Deep Learning at Continental.
Looking into the future, Continental is moving toward the cloud and plans to adopt MongoDB Atlas, the fully automated database as-a-service. The team believes that MongoDB will be a crucial component in helping them achieve Vision Zero which is being adopted as a new standard of safety across many cities and governments around the world.