Feature: Automotive
To be eff ective, a perception system must take raw sensor
measurements and turn them into a coherent understanding of the world – all within a fraction of a second. More information can make this harder, since more data consumes more memory, processing bandwidth and power, and takes longer to synchronise and interpret. With AI now reshaping automotive perception, dealing with
information overload becomes ever more important. T e fi rst Advanced Driver Assistance Systems (ADAS) architectures were built around deterministic processing pipelines, with sensors generating measurements, signal processing extracting features, and tracking algorithms identifying objects. Meanwhile, planning soſt ware made driving decisions, and each stage in the process curbed and reduced the amount of information passed to the next. Modern perception systems actually turn that on its head.
Instead of simplifying sensor outputs into object lists, modern AI architectures extract relationships that older algorithms might discard. Systems like occupancy networks, world models and end- to-end perception don’t ask sensors to decide what an object is, but instead need access to rich physical observations to do their own interpretation.
Different realms Cameras, radar and LiDAR are diff ering sensing methods, measuring diff erent properties of the environment around the car. A camera records refl ected visible light, allowing perception
soſt ware to easily recognise objects around the car. However, cameras depend on lighting conditions; for example, they function poorly in fog or darkness. Radar operates in an entirely diff erent physical realm. It
transmits radio waves and measures the signal that bounces back off an object; see Figure 1. It can measure range and relative velocity very accurately, and works well in poor weather conditions. Newer, so-called “4D” radar systems also off er elevation information, i.e., how tall or high up objects are, making it even easier to capture a three-dimensional representation of nearby objects and identify them. A key advantage of having radar within a car’s perception stack
is its ability to directly measure relative velocity using the Doppler eff ect (changing frequencies when objects move closer or further away). Camera-based systems compare successive image frames to infer motion, while radar measures it in each sensing cycle. In complex traffi c situations where cars, cyclists and pedestrians are constantly changing speed, this is extremely valuable, enabling perception soſt ware to identify crucial changes in behaviour earlier than vision-based systems. LiDAR off ers yet another perspective. It measures the time-
of-fl ight of laser pulses, thus creating very detailed geometric information of nearby objects. But, like cameras, LiDAR performs less well in certain weather conditions and the price of the sensors can be high. Sensor fusion brings all these inputs together to remove
uncertainty through an eff ective perception stack. T at’s why the specifi cations of sensors can never be considered in isolation,
www.electronicsworld.co.uk October 2026 35
with each specifi cation involving engineering compromises. For example, extending detection range oſt en means longer integration times and more sophisticated processing techniques. T e trade-off s are oſt en just as important as the main specifi cations. Trade-off s and metrics are calculated together to create a perception system for real roads. A perception system for motorway driving needs long-range detection, but also fast update speed, and urban environments require spatial resolution to diff erentiate pedestrians from cars, but also low latency for a highly dynamic environment. In addition, the development of these sensors has increasingly
become interdisciplinary; decisions in one domain infl uence performance through the whole perception stack. T e overall goal is to optimise the entire sensing system within the practical constraints of real-world operation.
The big questions T e progress toward autonomous driving has always been measured through technological advancement. T is has brought increasingly capable sensors, bigger datasets and more powerful processors, as well as more effi cient perception architectures. T us, the next generation of automotive sensors is not just about performance. Radar will no longer evolve to just fulfi l diff erent ADAS functions. It will be used in diff erent roles as AI-driven perception systems pose diff erent questions. Developers want direct access to richer radar information, not highly processed outputs. By providing direct access to synchronised radar streams, dense point clouds and higher-fi delity environmental representations, perception soſt ware can easier interpret complex data with AI. Signal processing and object detection remain important. Here,
the sensor becomes the provider of physical information to an AI system; see Figure 2. Today, engineers are focusing on how the whole stack performs.
Central to this is the question is how much the stack – the whole stack, not the individual sensors within it – can reduce uncertainty on the road before driving decisions are made. We are fi rmly moving toward sensors that provide high-quality information to an AI-enhanced perception stack, which will be at the heart of autonomous driving.
Figure 2: Sensor data feeding an AI-driven vehicle perception stack
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44
