Hardware-Friendly Variational Bayesian Method for DoA Estimation in Automotive MIMO Radar

Sparse Bayesian algorithms have attracted a lot of attention in various application areas for solving sparse recovery problems. One of these is the direction-of-arrival estimation in automotive radar due to the super-resolution capability. However, the computational complexity makes real-time capable implementations on state-of-the-art embedded platforms difficult. To tackle this challenge, we combine three techniques in this work resulting in a hardware-friendly sparse variational Bayesian algorithm that can handle high accuracy and throughputs with reasonable hardware costs. Firstly, we apply intra-iteration speed-up via angular decoupling of the calculations. Secondly, a highly efficient convergence acceleration technique based on exponential weighting is developed, which features minimal additional memory demand. Lastly, we derive a division-free algorithm by interlacing the algorithm with Newton's method. This reduces the demands on the utilized hardware platform and enables the implementation of the algorithm on embedded, power- and cost-optimized FPGAs and ASICs. The proposed algorithm is implemented on a novel application specific AI processor featuring a massive parallel vertical vector architecture as well as on a PC for benchmarking purposes. The results are compared to state-of-the-art algorithms.