AI Sensor Technology Makes Automotive Radar Systems More Robust

Modern driving assistance systems rely on several sensor types — cameras, LIDAR, ultrasound, radar — in order to function reliable in any situation that could arise. Radar, in particular, is an essential component, as it provides data on the locations and speeds of surrounding objects. However, problems arise when environmental noise reduces the quality of measurement, especially noise from extreme weather and other radar equipment.

Researchers at TU Graz, in collaboration with Infineon, have modeled an AI system for automotive radar sensors that filters out interfering signal, greatly improving object detection. Based on convolutional neural networks, or CNNs, which are already successfully used in image and signal processing applications. The structures, modelled on the layered hierarchy of our visual cortex, consume less memory than other neural network while still exceeding the capacity of currently available radar sensors

In experiments, the automotive driving neural networks were trained with noisy data and desired output values, allowing the team to identify particularly small and fast model architectures by noting the memory space and the number of computing operations required per denoising process. The most efficient models were compressed again, resulting in an AI model with both high filter performance and low energy consumption. Their denoising results are almost equivalent to the object detection rate of undisturbed radar signals. In other words, a model with a bit with of 8 bits achieves the same performance as comparable models with bit widths of 32 bits, but requires only 218 kilobytes of memory — a storage space reduction of 75%.

The team now wants to continue to improve the system so it works even when the input deviates from learned patterns. The system will have to be able to cope with the unexpected, and be able to notice when its own predictions are uncertain, in order to be safe and responsive in emergencies. In future work, the researchers want to determine which influencing factors are decisive in the system’s prediction-making process; in order to prepare the AI technology for real-world use, its behavior must first be a bit more explainable. Still, it is expected to be ready to equip the first radar sensors within the next few years.