How to simulate automotive radar?

I need to simulate automotive radars similar to Delphi ESR (scanning radar). There is a radar generation function as part of the Automated Driving System Toolbox. Unfortunately, that function appears to be tied to Matlab's own simulation... which is not complex enough for my purposes. Real radar returns include ground clutter and other stationary objects. Are there existing functions for generating radar functions from point cloud like representations, maybe with an added relative velocity layer?

Answer 1

[Honglei Chen]

Have you looked at the following example? Would that be at the level you are interested?

https://www.mathworks.com/help/phased/examples/automotive-adaptive-cruise-control-using-fmcw-technology.html

HTH

[Eric]

Thanks for the answer. That seems close, but I'm unsure of how to take it to the next step. This answer tells me how to estimate the returns to a single arbitrary target (e.g. a car). But given a static scene (e.g. a point cloud with no other moving objects), how can I determine which objects will generate a valid return? From experience with real sensing hardware, I know that some objects will not generate any signifcant return (e.g. a trash can), but others will generate lots of returns. What objects generate a return? And from what point on the object?

[Honglei Chen]

For radar, the strength of a target's return is determined by its radar cross section. You probably need to include more targets in the simulation, define teh grid, and the process the data to get the point cloud.

[Eric]

Right. For small objects that makes sense, but what about very long objects? I actually have point clouds generated by LIDAR and/or stereo vision and need to estimate radar returns. I know from practical experience that the long conrete barriers to one side of the vehicle generate lots of returns from corner-mounted radars that are hard to filter out using range-rate alone. I'm not sure how to estimate what a scanning radar might do when given such surroundings.

[Honglei Chen]

For simulation purpose, a common approach is to decompose large objects to small pont targets, one in each resolution cell. It certainly can never replace the measured data, but can be a good first order approximation for algorithm development.