Austria develops AI-enhanced radar system to improve target detection capabilities of autonomous vehicles

Artificial intelligence (AI) is at the core of building future autonomous driving, but there are still many safety issues that need to be addressed. According to foreign media reports, researchers at Graz University of Technology (TU Graz) in Austria have created a powerful AI-enhanced radar that can help improve existing radar technology and also help solve safety issues.

Graz University of Technology is working with Infineon Technologies to develop new radar sensors for autonomous driving (Image source: Infineon Technologies)

Autonomous driving and driver assistance systems rely on a variety of advanced sensor systems such as lidar, ultrasonic and radar, which can feedback key information about the surrounding environment to help the vehicle operate safely. In particular, radar sensors can transmit the position and speed information of surrounding vehicles and objects to the moving vehicle. However, such systems are susceptible to some interference in traffic and environmental influences, such as interference from other (radar) equipment. In addition, radar is severely affected by extreme weather that can cause noise, resulting in poor measurement quality. Therefore, the development of radar systems that can filter noise and interference information and are enhanced by AI technology means that autonomous driving applications are shifting towards safer object detection systems.

Convolutional Neural Networks

The TU Graz team developed an artificial intelligence model based on a neural network that eliminates interference from radar signals and that goes far beyond existing technologies.

The team now aims to optimize the model so that it can work beyond fixed parameters and inherent learning modes, and develop capabilities that allow it to recognize objects more reliably.

Initially, the researchers developed an automatic noise suppression model architecture based on a convolutional neural network (CNN), which filters all visual input information and determines the specific connections to generate a complete image.

The way the architecture is built means that CNNs will consume less memory when processing information than current radar systems, but will achieve greater and more advanced capabilities. The main goal of the Graz University of Technology team was to significantly improve the efficiency of existing systems while introducing a new technology that has greater capabilities in real-world environments.

Advanced and powerful technology

During the experiments, the TU Graz team trained various neural networks with noisy data and the desired output values. Afterwards, by analyzing the memory space and the number of computing operations required for each denoising process, a particularly small and fast model architecture was determined. The most efficient model was then compressed again by reducing the bit width (i.e. the number of bits used to store the model parameters), ultimately resulting in an AI model with both high filtering performance and low energy consumption. The excellent denoising results, combined with an F1 score of 89% (a measure of test accuracy), give it an object detection rate that is almost equivalent to an undisturbed radar signal. The interfering signal has therefore been completely removed from the measurement signal.

Expressed numerically, a model with a bit width of 8 has the same performance as a similar model with a bit width of 32, but only requires 218KB of memory, which is equivalent to a 75% reduction in storage space, which also means that this model far exceeds the existing state-of-the-art level.

As a next step, the TU Graz team will focus on the REPAIR project (Robust and Explainable AI for Radar Sensors) to optimize the AI-enhanced radar system. The team is currently working with Infineon Technologies.

This also means that the system will be able to tolerate high levels of interference and demonstrate greater robustness; the team will develop a radar system that will not only improve efficiency but also help prevent any catastrophic accidents in autonomous driving applications.