Hi,
I would like to know how to calculate the proper value for the "noise sum divisor" and the "threshold scale" in cfarCfg based on the parameters:
numRxAnts
numTxAnts
numRangBins
numDopplerBIns
Another question is, how will those two parameters affect the detection results. According to my experiments, I will have more objects detected when I reduce the value of threshold scale but have more fake detections. Is my understanding correct? Is there a way to reduce the rate of fake detections?
Hi Jerry,
The mmwave SDK user guide has sections (pg 24 - shown below) on what these parameters mean. Essentially, the combination of 'divShift' and 'threshold scale' combine to form the CFAR (Constant FAlse Alarm Rate) threshold. In order to understand how to set thresholds, you need to understand the CFAR algorithm and its variants. Attached is a link to a book that you may find useful.
That being said, your understanding is correct: Lower the threshold, more are the detections. If you set the threshold too low, some of the objects detected can be false. Hence, the best advice to avoid false detections is to not set a low threshold.
Detecting false objects is an interesting research project in itself. Higher layer algorithms like tracking or clustering, can help in this regard. These algorithms can look at certain features of the object and make a decision. For e.g. if the object detected is not persistent (it doesn't appear in more than one frame), then it can be declared as false. You can try the 'traffic monitoring' TI Design which does both clustering and tracking. http://www.ti.com/tool/tidep-0090.
