ADAS and Autonomy:
the value of behaviour predictions
Within the next 10 years it’s a certainty that driverless vehicles will be commonplace on city streets and on the highway; a mixed fleet of driven and autonomous vehicles (AVs) will have to negotiate the roads together in increasingly challenging environments.
This will be the reality looking out towards 2040 at least and probably further ahead in many countries; it’s simply a fact that large numbers of humans will be driving themselves for the foreseeable future.
However, until all vehicles are autonomous and communicating with each other through V2X and centralised fleet management systems, there will always be a need for autonomous vehicles to understand how the human-driven vehicles around will behave, in order to calculate the path and movements of the AV. Indeed, it’s often said that predicting the “next 5 seconds” is the Holy Grail of autonomous systems that will lead to wide-scale proliferation of AVs.
Currently, AV systems use a variety of onboard sensors such as cameras, radar, lidar and ultrasonic to observe other elements – vehicles, pedestrians, cyclists, motorbikes, animals, street-signs etc – to learn from; these are brought together in a ‘sensor-fusion’ and then machine learning tools allow the AV to ‘predict’ what the likely next event (trajectory, speed, etc) of the element will be. This allows the vehicle to then plot its course through the world.
This is a highly complex and compute-intensive activity; the result being that AV systems will remain high-cost and slow to expand beyond small geo-fenced zones.
A more sophisticated approach, which reduces the ambiguity solving on the AV side, would be to provide destination, trajectory and behaviour predictions from ‘human-driven’ vehicles to the system, freeing it up to concentrate on the AV trajectory. This involves the driven-vehicle making predictions in real-time for imminent actions and feeding this to a cloud-based open API, or working as an edge client providing V2V straight to the AV.
The prediction of driver behaviour and vehicle trajectory involves learning from car data and building a profile of the driver, then using this as a real-time ‘suggestion’ for other services to consume, typically as an API or part of an embedded solution. CloudMade has been doing exactly this for many years with OEM data and using machine learning to uncover personalised behaviour and journey predictions.
CloudMade’s Adaptive Framework can predict:
- Trajectory prediction. Predict next N km of driving and get context info along this path: traffic info, road signs, road geometry;
- Predict the next manoeuvres (e.g. lane changing and braking when user need to turn right of stop in a traffic light);
- Predict the user behaviour for the next N km to understand the potential risks for each specific driver and adjust the ADAS system according to it (e.g. earlier warnings when needed);
- Analyse data from other vehicles to make ADAS adjustment more precise (e.g. slippery road segment ahead);
- Support vehicle-to-vehicle communication;
- Integration with Cognitive Capacity detection software. Predict the cognitive load for the next N km or for the next manoeuvre.
How it works:
These predictions are highly valuable in the context of AV “next 5 seconds calculations’, offloading the need for the AV to track multiple elements.
However, it’s not just this benefit that CloudMade behaviour predictions can help with.
The benefits of using CloudMade’s Adaptive Framework for advanced vehicle assistance include:
- Consuming highly likely driver behaviour and vehicle trajectory as a service;
- Adjusting the sensitivity of ADAS system (e.g. collision warning) and make it more sensitive in advance of potential danger manoeuvre;
- Notifying stability control systems (e.g. ESP) for a potential danger situation upfront;
- Recommendation on activation and settings for ACC, Drive Modes and other optional control systems;
- Providing a more sophisticated notification interface (communication channel & notification type) based upon the criticality of the context.
If you are an OEM developing new ADAS technologies or looking to improve your existing ones with greater intelligence, we would be excited to discuss your issue and talk more about our Behaviour Prediction modules, which may be able to reduce your compute requirements or augment your decision-making algorithms.