Object Association using RADAR Data

Hi @nutonomy,
I m trying Kalman tracking of objects and want to associate the detections to the tracks using data from only RADARs.
Could you please clarify on the following,

  1. Does the information about rcs, dyn_prob, ambig_state, invalid state features suffice for the associating detection to the track? Are object dimensions given if available? Could road width be known ?
  2. Are IDs from same sensor from different sample time identifying the same object?
  3. Could mounting axis position of RADARs on vehicle be provided?

Thank you,


  1. I assume you are also including x/y/z of each radar return? I would start by computing the association purely geometric: check for each box and each radar return that (x/y/z) is in box. Then you can still refine that. You can infer the road width from the map (we are updating the map layers this week to include lanes, sidewalks, crosswalks, etc.).
  2. If you ping me at holger@nutonomy.com I can provide the radar manufacturer’s handbook. They do perform clustering (https://github.com/nutonomy/nuscenes-devkit/blob/master/python-sdk/nuscenes/utils/data_classes.py#L312), but I am not sure whether they track objects temporally. I don’t suppose so, as they don’t output a track id. The ID that you are seeing there is simply an integer that is incremented for every return.
  3. You can retrieve the extrinsics for each sensor from the calibrated_sensor table: https://github.com/nutonomy/nuscenes-devkit/blob/master/schema.md#calibrated_sensor