Approach to denoising of interfered 4-channel FMCW radar data using Convolutional Neural Network

Radar sensors are essential for the further development of advanced driver assistance systems(ADAS) and autonomous driving (AD) vehicles. Due to a rapidly increasing number of vehicles with built-in radar sensors (whether as an assistance system or a stand-alone system for autonomous driving), overlaps in the radar frequency band may occur. Any overlaps can cause the radar waves to interfere with each other. Therefore, minimizing the effects of interference when processing radar data is the goal to aim for. For this purpose, various model architectures of convolutional neural networks (CNN networks) are investigated and examined for their effectiveness. Compared to other interference cancellation methods, this approach does not examine the absolute values of the Range Doppler Matrices (RDM). Instead, the RDMs are interpreted and processed as complex values. By interpreting the RDMs as complex values, the correlation between real and imaginary parts can be optimally exploited. In this paper, as an extension of the existing methods the relation between the different channels of a FMCW radar is also considered by the CNN architecture. This has improved the performance of existing approaches by exploiting the relationships among the complex RDMs. To learn different model architectures, interfered RDMs are simulated. Using this simulated data, the different models are compared with each other. Subsequently, the most performant model is tested for its effectiveness with real-world data.