Modern Cars and Modern Design Challenges
For higher levels of autonomy data processing speed is a key requirement that needs to be met in any AI solution. Advanced Driver Assistance Systems (ADAS) need such speeds to operate effectively. These systems have to conduct multiple types of processes such as Vulnerable Road Users (VRU) protection, lane support, collision avoidance, Automatic Emergency Steering (AES) and automated parking! To make things even more complex some of these systems require 360-degree sensor feeds to be assessed in real-time. Apart from the obvious challenges of conducting all these calculation in real-time these systems need to have a low energy footprint.
AI Processors to the Rescue
Luckily the ADAS can now be designed to integrate a specific high performance low power and low foot print AI deep learning chip. These chips are small enough to be directly mounted on a main board, or provided in an M.2 or mini-PCI board offering. While small these AI processors can do all the heavy lifting as they are specifically optimized for AI and work with the mainboard CPU to obtain sensor data.
A neural net is a decision-making weighted matrix that has different layers with pre-learned weightings for each decision that is required. This allows a neural net to work out if there is a person in front of the car or just a pothole. AI processors contain core structures that are similar to that of the neural nets that they need to do calculations for; making them highly efficient. They need to do calculations that are numerous but low complexity similar to a GPU in a graphics card yet are more efficient for AI solutions because of the removal of unnecessary features that are useful in GPU processing.
Scalability in Your ADAS ECU
One AI Processor chip can process multiple video or sensor streams at once however you may decide that you want to create more complex workflows to optimize your design. Thankfully these chips are fully scalable; multiple chips can work in together in tandem or in a cascade arrangement. This can allow you to use two separate systems to check for disparity in tandem or just spread the workload for each application. For example, you may have an AI processor for VRU, one for land detection and so on! A cascade approach can help with filtering operations; the neural network for decision making is split into different layers and these each could utilize a processor specifically for each filtering process although in most applications you just need one.
Developer Tools
No matter whether you decide to use one chip or many in different configurations there are great tools that you can use. Typically, solutions these days provide the software developer with SDK for easy integration of existing stacks. For example, most modern AI solutions utilize a container for work to be conducted by a team efficiently and additionally allow for easy integration into such solutions.
If the developer needs added help to get the product to market then some AI processors also have out of the box solutions based on common decision-making processes. This means that you can have a solution up and running in minutes with little modification required.
The Future of ADAS ECUs
ADAS control units are benefiting from AI processors not just from the ease to install the hardware, ease of programming but also the fact that a constant connection to the internet is not required. Once a system has been trained then all the decision making is handled by the device. The only time an internet connection may be needed is if any software updates are required. ADAS ECU’s now have the ability to be let lose on the roads and function effectively, processing data streams in real time and leveraging the best sensors that the system requires.
