Multipath identification and multipath-assisted multitarget tracking in automotive radar for autonomous vehicles

Abstract

Automotive radars play a unique role in autonomous vehicles due to their wide coverage and ability to operate in adverse weather conditions. However, in complex driving environments, radar signals can reflect and scatter in the nearby surfaces, reaching the receiver through multiple paths (multipath). If these signals are ignored in the measurement process, they can lead to false tracks and reduce tracking accuracy. Conversely, these multipath signals can also originate from actual targets, providing valuable information about them.

This thesis focuses on multipath-assisted multitarget tracking by identifying multipath signals and utilizing them to improve tracking performance. The presented approaches demonstrate promising results with both simulated and real-world data.

The identification of multipath signals during data association often neglects the uncertainties present during track initialization. This oversight can lead to false tracks when multipaths are involved. To address this issue, a multiframe assignment technique is proposed, which jointly considers track initialization and data association. Additionally, for trackers that do not account for propagation uncertainty, a cumulative hypothesis-based probability calculation is introduced at the track level, along with a method for estimating potential reflection surfaces.

The reflection surfaces in the coverage are estimated by accumulating sparse radar detections over multiple frames. To improve multipath-assisted tracking and focus on regions with ghost detections, a multitasking transformer network is proposed in this work to identify ghost detections and their source of origin in uncertain environments. Although the detection level multipath identification is instantaneous, it may struggle to differentiate between multipath aligning with the direct path. Thus, a Long Short-Term Memory network is introduced for ghost track identification based on sequential data and known surfaces.