Time-Sensitive and Distance-Tolerant Deep Learning-Based Vehicle Detection Using High-Resolution Radar Bird's-Eye-View Images

Advanceddriver assistance systems (ADASs) and autonomous vehicles rely on differenttypes of sensors, such as cameras, radar, ultrasonic, and LiDAR to sense thesurrounding environment. Compared with the other types of sensors,millimeter-wave automotive radar has advantages in terms of low hardware costand reliable object detection under poor weather conditions, such as snow,rain, or fog, and doesn't suffer from light condition variations, such asdarkness. High-resolution radar bird's-eye-view (BEV) obtained from radarrange-azimuth spectra through a polar-to-Cartesian coordinate transformcontains targets' geometric information that can be learned by deep neuralnetworks for object detection. Compared to radar point clouds, there is noinformation loss in radar BEV. Unlike RGB images, radar BEVs are single-channelgrayscale images with unique characteristics such as inconsistent resolutionand SNR. Therefore, directly implementing an image-based object detectionnetwork is not an optimal solution for object detection using radar BEV. Wepropose a Temporal-fusion, Distance tolerant single stage object detectionNetwork, termed as, TDRadarNet, to robustly detect vehicles up to 100 metersunder various driving scenarios. DRadarNet leverages historical radar frames toexploit temporal features and separates far and near fields to addressinconsistent resolution in radar frames. With qualitative and quantitativeresults, we show that TDRadarNet achieves 68.9% in precision and 66.8% inrecall, and 67.8% in F1-score, which outperforms the state-of-the-artimage-based object detection networks by 10.6%, 17.1%, and 14.1%.