Road user detection on radar data cubes with deep learning

Abstract

Automotive radar sensor information plays a vital role in various Advanced Driver Assistance Systems applications like target detection, adaptive cruise control, collision detection, and target velocity estimation. The radar sensor is the only source of information that can provide distance and velocity estimation in all environments and weather conditions. Current methods on object detection with radar sensor data depend only on sparse radar point cloud information. However, converting raw radar data into point cloud information using signal processing techniques like Constant False Alarm Rate (CFAR) algorithms is not optimal. Because the efficiency of CFAR algorithms depends on various hyper-parameters like the number of guard cells, the number of train cells, and noise threshold. It is generally challenging to estimate optimal parameter values due to high variation in the road scenes. Hence, there is a high chance of missing key targets leading to sub-optimal detection results. Recent research showed that object detection using low-level radar data tensor with the help of deep learning networks had provided better results. However, they have either incorporated additional sensor information or used only the projection of the radar cube, which is difficult to associate with the spatial Range-Angle (RA) pixels. This work deals with developing a deep learning technique that utilizes the complete radar tensor with added attention to the essential doppler values. The research concentrates on the analysis of diverse strategies in building object detection pipelines using radar tensor information. Moreover, this method has also employed stitching of radar data cubes from three different radar sensors to get a higher field of view. To the best of my knowledge, this is the first method to utilize tensors from multiple radar sensors along with essential doppler information for oriented bounding box detection using deep learning.