AURIX™ TC4xx - Parallel Processing Unit (PPU) for Multi-Layer- Perceptron (MLP) based Direction of Arrival Estimation

Abstract:
Among the major megatrends in the automotive industry is the development of advanced driver assistance systems (ADAS). Extracting relevant information from radar sensor data is crucial and techniques based on artificial neural networks show great results for high resolution direction of arrival (DoA) estimation. At the same time, it is seen that artificial intelligence (AI) applications have different requirements regarding compute and memory than traditional algorithms. In this paper, Multi-Layer-Perceptron (MLP) based high resolution DoA estimation method is described which is capable to deliver superior results on synthetic radar data. The workflow to generate an optimized embedded implementation for a Vector DSP architecture is also demonstrated. The execution performance is measured on an cycle approximate model and the results are discussed in the radar application context.