Machine Learning Based Radar Heatmap Generation for Advanced Driver Assistance Systems

Radar sensors are commonly used in Advanced Driver Assistance Systems (ADAS) to detect surrounding objects. Radar signals pass a signal processing chain to compute the range, the velocity and the angular position of potential obstacles and other traffic participants. Major challenges of this procedure are the high complexity and latency of transferring a large amount of raw radar data to subsequent processing units. This paper presents a machine learning model that dramatically reduces the latency of the data transfer by compression of the raw radar data and that generates Range-Angle (RA) Heatmaps directly from the compressed data. RA Heatmaps present the essential visualization technique for the localization of surrounding objects in the radar horizon. In addition, they can be used to detect and identify objects. We use the autoencoder and the U-Net architecture for the compression and the RA Heatmap generation respectively. For the collection of raw radar data, a Texas Instruments (TI) A WR1243 FMCW radar sensor board is utilized. The data to train the U-Net architecture is obtained by applying analytical radar imaging algorithms including a 3-dimensional Fast Fourier Transform (3D FFT) and our proposed framework for Region of Interest (ROI) extraction on the collected data. Experimental results indicate that our final system can achieve a significant reduction of the original raw radar data while still maintaining all relevant features required for object recognition tasks. Excellent accuracy scores exceeding 99 % on input data that has been compressed by a factor of four validate the capabilities of the model and underline its potential for radar data processing in the context of ADAS.