Deep Learning-Based Radar Detector for Complex Automotive Scenarios

Understanding the environment is essential for an autonomous driving system. Most learning-based perception models leverage data acquired with Lidar or camera sensors, leaving automotive radars relying on traditional algorithms. However, autonomous systems must be robust and reliable in all-weather and all-lights conditions. Radars have proven to be effective even in adverse weather conditions, whereas other sensors lose accuracy. Previous research in radar target detection investigated the benefits of a learning-based method as compared to traditional ones. Though, those methods have only been tested on single point targets collected in an anechoic chamber. In this work, we extend those studies to locate the targets in complex automotive scenarios. We propose a multi-task model based on Convolutional Neural Networks able to detect and locate targets in multi-dimensional space of range, velocity, azimuth, and elevation. Due to the lack of publicly available datasets, we simulated a dataset containing more than 20.000 frames of automotive scenarios and many road elements including (but not limited to) cars, pedestrians, cyclists, trees, and guardrails. The proposed method was trained exclusively on simulated data. The model demonstrated superior performance on unseen simulated scenarios as compared to conventional CFAR-based methods. Furthermore, we qualitatively evaluate our method on real-world radar recordings showing improved performance over prior methods.