Deep learning approaches for radar signal processing and perception

Autonomous driving has recently gained significant attention due to its disruptive potential and impact on the global economy. However, these high expectations are hindered by strict safety requirements for redundant sensing modalities each able to independently perform complex tasks to ensure reliable operation. We explore a deep learning-centric approach for radar signal processing and tackle both perception and sensing research topics. Our focus is on advanced training methods for deep neural networks which alleviate the required resources for manual labeling as well as open new possibilities not possible using supervised learning.

Current radar systems are limited in their angular resolution causing a technological gap with the required needs. An industry and academic trend to improve angular resolution by increasing the number of physical channels, also increases system complexity, requires sensitive calibration processes, lowers robustness to hardware malfunctions and is associated with high costs. We offer an alternative approach using a self-supervised training methodology to create a coherent radar array with greatly improved angular resolution.

At the core of an autonomous driving algorithmic stack is road segmentation which is the basis for numerous planning and decision-making algorithms. Radar-based methods fail in many driving scenarios, mainly since various common road delimiters barely reflect radar signals, coupled with lack of analytical model of road delimiters. We introduce a weakly supervised training method and prove radar can be used to identify road delimiters in scenarios previously thought not possible.

We also follow a data-centric approach and examine the effects of angular resolution on lidar and radar DNN-based 3D vehicle detection. We focus on subsampling techniques and show that when applied on high-resolution lidar point clouds, improved performance and efficiency of DNN-based vehicle detection methods can be attained. In addition, we examine how DNNs performance is affected by the trade-off between resolution and grating lobes in sparse radar arrays. We show that improving the angular resolution by increasing the distance between antenna elements can improve the performance of DNN-based vehicle detection methods