ADAS – Step by Step Towards Fully Automated Driving

Automotive developers are now working on the next generation of ADAS. The first cars with Level 3 highly automated functions are already on the road. The main difference between Level 2 and Level 3 is that driver supervision is no longer required at all times (eyes-off). This means the OEM is liable for everything that happens while the vehicle is driving in Level 3 mode.

To achieve this, systems must “understand” their environment and a wide range of situations ever more effectively, equipping vehicles with human-like perception. Powerful electronics, real-time communication, and AI are increasingly converging to make this possible. This enables fully automated vehicles to “decide” within milliseconds. Early sensor data fusion and seamless V2X integration (vehicle-to-everything) make ADAS more precise, robust, and safe. These technologies deliver more than just an extended field of vision – they allow for deep, context-dependent analysis and interpretation of the surroundings.

Processing large volumes of sensor data streams requires correspondingly powerful hardware and efficient memory solutions. Specialised hardware accelerators provide the necessary computing power and memory partitioning. The growing integration of V2X communication also demands additional hardware and cybersecurity measures to ensure secure cloud connectivity.

ADAS focus not only on the vehicle’s external environment. Driver monitoring systems (DMS) are increasingly being integrated – not least due to stricter Euro NCAP requirements. Especially in semi-autonomous systems, facial recognition and eye-tracking are being implemented to ensure driver attentiveness.

This approach ensures redundancy and higher accuracy, even under challenging conditions such as fog or heavy rain.

Lidar and radar are the two most important technologies for perceiving the more distant environment. Lidar is currently the more precise of the two, enabling high-resolution, three-dimensional mapping – but costs three to five times more than radar. Radar, in contrast, is compact, cheaper, and works reliably in different weather conditions and darkness, though traditional radar systems have much lower resolution compared to lidar.

One of the key technology trends is the introduction of solid-state lidar sensors. These use few or no moving parts – improving durability and reducing maintenance and unit costs. Solid-state lidar also simplifies integration into car body designs, such as bumpers, headlights, and windscreens, meeting both aesthetic and safety requirements of OEMs. Combined with AI-driven point cloud compression and edge analytics, solid-state lidar enables scalability without performance loss.

By using multiple input/multiple output (MIMO) antenna arrays, they can generate point cloud images of the vehicle’s surroundings. This makes it possible to detect not only distance, relative speed, and azimuth (with an accuracy of 0.11 degrees), but also object height above road level. A major trend in automotive radar technology is the integration of this 4D imaging into radar-on-chip solutions. Multiple radar functions are integrated into a single chip – reducing the bill of materials (BOM), simplifying thermal considerations, and lowering the risk of electromagnetic emissions.

That is why AI is increasingly being integrated into ADAS. Through deep learning, systems can classify objects by movement patterns, shape, and size, while reinforcement learning models improve decision-making by simulating real-world driving scenarios. As AI models continue to learn from millions of kilometres driven, their predictive accuracy will increase, enabling autonomous vehicles to handle complex edge cases, such as pedestrians crossing in poor visibility or merging traffic.

Today, AI is used along the entire software technology chain: from object detection and multi-sensor fusion (cameras, radar, etc.) through to decision-making and safe automated control of powertrain, steering, and braking. Based on modern computing platforms, vehicles can now run even complex AI models with billions of parameters directly on-board (edge AI). For example, multimodal AI models combining image, speech, and sensor data can interact with driver assistance systems and personalised agents in real time.