Sensor technology for autonomous driving filters out interfering signals

In order for the assistance and safety systems in modern cars to perceive the environment and function reliably in all possible situations, they must rely on sensors such as cameras, leader, ultrasound and radar. The latter, in particular, are indispensable components. Radar sensors provide vehicles with information about the location and speed of surrounding objects. However, they have to face many destructive and environmental impacts in traffic. Interference from other (radar) equipment and extreme weather conditions create noise that negatively affects the quality of the radar measurement, the University of Graz writes in this press release.

“The better the interference signal cleaning, the more reliable the position and speed of the objects can be determined,” explains Franz Pernkopf of the Institute for Signal Processing and Speech Communication. Together with his team and partners from Infineon, he developed a neural network-based artificial intelligence system that mitigates mutual interference in radar signals, far exceeding the current state of the art. Now they want to optimize this model so that it also works outside the studied patterns and recognizes objects even more reliably.

Resource efficient and intelligent signal processing

To this end, researchers have for the first time developed model architectures for automatic noise reduction based on so-called convolutional neural networks (CNN). “These architectures are modeled based on the hierarchy of layers of our visual cortex and are already used successfully in image and signal processing,” says Pernkopf. CNN filters visual information, recognizes connections, and completes an image using familiar templates. Due to their structure, they consume much less memory than other neural networks, but still exceed the available capabilities of radar sensors for autonomous control.

Compressed AI in chip format

The goal was to become even more efficient. To this end, the TU Graz team has prepared various of these neural networks with noise data and desired output values. In the experiments, they identified particularly small and fast model architectures by analyzing memory space and the number of computational operations required for the noise reduction process. The most efficient models were then compressed again by reducing the bit width, i.e. the number of bits used to store model parameters. The result is an AI model with high filter performance and low power consumption at the same time. Excellent F1 noise attenuation results (a measure of test accuracy) of 89 percent are almost equivalent to the detection rates of objects in intact radar signals. Thus, the interference signals are almost completely removed from the measuring signal.

Subscribe to our newsletter!

Your Weekly Innovation Review Every Sunday the best articles of the week are in your inbox.

Expressed in numbers: with a bit width of 8 bits, the model achieves the same performance as comparable models with a bit width of 32 bits, but requires only 218 kilobytes of memory. This corresponds to a 75 percent reduction in storage space, which means that the model is far superior to the current state of the art.

Focus on strength and clarity

In the FFG REPAIR project (a reliable and explanatory AI for radar sensors), Pernkopf and his team are now working with Infineon for the next three years to optimize their development. Pernkopf says, “For our successful tests, we used data (note: interference signals) similar to the ones we used for training. Now we want to improve the model so that it still works when the input signal deviates significantly from the studied patterns. This will make radar sensors many times more reliable in relation to environmental interference. After all, the sensor also faces various, sometimes unknown situations in reality. “So far, even the slightest change in the measurement data has been enough to ensure that the result is curtailed and objects are not detected or detected incorrectly, which would be devastating if autonomous control were used.”

The light shines into the black box

The system must deal with such problems and notice if its own predictions are uncertain. Then, for example, he can respond with a safe emergency procedure. To this end, researchers want to find out how the system determines forecasts and what influencing factors are crucial. This complex process in the network was previously understood only to a limited extent. To do this, the complex architecture of the model is transferred to a linear model and simplified. According to Pernkopf: “We want to make CNN’s behavior a little more explanatory. We are interested not only in the final result, but also in its range of variation. The smaller the variance, the more confident the network. ”

Either way, much remains to be done for real-world use. Pernkopf expects that the technology will be developed to such an extent that in the next few years it may be equipped with the first radar sensors.

Also interesting: A smartphone on wheels or vice versa?

Leave a Comment