Smart-CFAR, a machine learning approach to floating level detection in radar

Machine Learning (ML) allows us to craft a system's behavior based on examples instead of being explicitly programmed or designed. In our research, we show that ML can be applied to mimic some conventional radar signal processing algorithms and remedy known underperformance of these algorithms. In particular, by taking the CA-CFAR as use case, we train a feedforward artificial neural network (ANN) on a dataset that contains the inputs and outputs of the CFAR sliding window. In addition, the same ANN is fed with desired input/output combinations where the CA-CFAR is known to give undesired results, for example in case of large clutter discontinuities. Through computer simulations we show that the still computationally friendly ML-based CFAR outperforms the classic CA-CFAR algorithm, which leads the way for further research on applying ML to detection in the radar processing chain.