Choosing The Best Vision System For Highly Automated Vehicles
Centralized or edge processing, which is the best choice?
By Jeff VanWashenova, Director, Automotive Segment Marketing, CEVA
One of the most important attributes of a highly-automated vehicle (HAV) is its ability to see. Multiple computer vision sensors, distributed along the periphery of the vehicle, are what make HAVs ‘aware’ of their surroundings. This awareness is what enables decision making, whether it’s changing the course of the vehicle, or notifying the human driver to avoid a potential hazard. You can see a nice explanation about how vehicles ‘see’ in this TED talk.
There are two distinct approaches to managing this multitude of sensors: edge processing and centralized processing. In edge processing each sensor is a ‘smart sensor’ meaning it converts the sensory data input to relevant information and triggers a reaction. In centralized processing, each sensor is ‘dumb’, meaning it collects raw data without any processing and sends it to a central unit, which does all the processing and decision making.
AI Super Computers for Autonomous Vehicles
Centralized processing is a great solution for fully autonomous vehicles. It requires a very intelligent ‘brain’ for the entire vehicle, which receives all the sensory input. It then compiles it together and processes the data using sensor fusion and deep learning algorithms to assess the current situation. It then performs the appropriate action (steering, accelerating, breaking, signaling, etc.). In addition to the input from the sensors, this ‘brain’ could also receive data from other vehicles or intelligent infrastructure and other vehicles via Car-to-Car communications to select the best course of action.
That said, it’s important to note that self-driving vehicles aren’t being mass produced yet. Most of the cars hitting the road include some level of automation but are not fully autonomous. As I described in a recent post about the challenges to fully autonomous vehicles, there are four categories of automation ranging from no automation at all (level 1) to fully autonomous (level 4). Currently, the market is dominated by level 2 and 3, so the centralized approach might not be fitting for most vehicles.
The main drawback of the centralized processing approach is that it requires a very powerful processor to receive all the data and churn out a decision, based on all the input. This type of intelligence on an embedded system poses some very serious challenges. Utilizing the cloud for processing is not a viable solution, due to extremely strict timing and safety constraints. All this makes the centralized solution very expensive and power hungry. This means that there probably won’t be many privately owned autonomous vehicles in the near future. When it does reach car owners, mostly high-end luxury and fleet/car sharing vehicles will benefit from its advantages, at least in the first few iterations. So, the centralized approach probably won’t be shipped in high volumes for quite a few years.
Smart Sensors for ADAS and Mass Market Automation Features
On the other hand, edge processing is already available, and it is fueling the progress in advanced driver assistance systems (ADAS) and HAVs. In edge processing, as I stated, each module has its own intelligence, meaning it can process the input data using sensor fusion, deep learning, and other algorithms. These smart sensors can ‘see’ the surroundings and are not dependent on a large power-consuming central processing unit. Smart sensors can be extremely power efficient, compared to a centralized computation unit that processes the input from all the vehicle’s sensors.
The main advantage of smart sensors is the scalability of the solution. Implementing each feature separately, in a standalone unit, significantly drives down cost for introducing a few ADAS features. That is why this is the current solution in almost all HAVs. This is also the reason that automation features, especially safety features, are available on economy vehicles, as well as luxury models.
The bottom line: Edge processing for the near future; Centralized for the distant future
Putting these pros and cons together, it looks like the more common solution is edge processing, and it will continue to be for at least the next few years. The centralized approach has significant advantages, but it will take time for the technology to meet mass market budget constraints. Either way, both options can accommodate some very interesting use cases and bring vehicle automation to new levels. It will be very interesting to see how the changes take effect on the roads in the next few years.
CEVA’s deep learning and intelligent vision platform, based on the CEVA-XM6 vision processor, is one solution for automotive computer vision. It is tailored to handle the most advanced deep learning algorithms while using extremely low power. Thanks to its flexible and scalable design, it can be suitable for both edge and centralized computing. To find out more, click here.
To learn more about intelligent solutions for highly automated vehicles and autonomous driving, log into our webinar: Challenges of Vision Based Autonomous Driving & Facilitation of An Embedded Neural Network Platform, click here.