Simulating object lists using neural networks in automotive radar

Simulation of sensor readings is important within the field of advanced driver-assistance systems, specifically with respect to feasibility studies of high-risk scenarios, where carrying out such test drives in the real world is expensive. Automotive radar signals are multi-dimensional and their data sizes are large, due to multiple measurements being taken within a short time frame in order to resolve ambiguities. Analyzing such signals and generating models from them is therefore no easy task. This paper presents an approach to generate models from data using neural networks. Conditional variational auto-encoders are used for their ability to learn complex distributions and can easily be trained via gradient descent-type algorithms. Our method is able to handle big data sizes and generate synthetic data of high accuracy.