Graph neural networks for clutter and target classification in automotive radar

Abstract

Treball fi de màster de: Erasmus Mundus joint Master in Artificial Intelligence (EMAI)

Supervisor: Lejla Batina

Current driver assistance systems employ exteroceptive sensors such as radar, Li-DAR, and cameras to collect information about the road and the status of surrounding traffic participants and other obstacles. Due to its relatively low complexity and high versatility in all lighting and weather conditions, radar sensors are an essential component for various functions such as cruise control, emergency braking, and blind spot detection. The large wavelength electromagnetic pulses that radar sensors use are the reason for its advantages, but also the principal reason for its downsides. Most flat surfaces act as perfect mirrors for the radar pulses and reflect the signal specularly rather than diffusely. Consequently, outgoing radar pulses are likely to bounce off multiple objects before returning to the radar sensor, resulting in multipath reflections that are known as ghost objects. These detections are especially problematic since they comprise a large majority of the clutter detected by the radar, surpassing most traditional signal filtering techniques due to their appearance in groups of detections with similar velocities, which offers even less distinguishability from real objects. As a result, detecting clutter early in the automotive radar processing pipeline is important to ensuring the smooth and safe execution of later components of the pipeline and ultimately of the driver assistance systems. Owing to the limited resources of automotive chips, automatic clutter detection methods must have a low memory footprint and a fast enough execution time in order to be deployed in real-time scenarios. This thesis proposes a series of lightweight models based on graph neural networks for detecting clutter in radar point clouds, particularly in resource-constrained environments.