IWR6843: questions on people counting visualizer

Hi Justin

Thanks for your reply.

So in the visualizer, it only shows the non-static points associated with each target unless we enable the visualization of static points. In addition, does the visualizer always show the total targets and points in a couple of previous frames at each frame?

Thanks!

Kai

Hi Justin

Does the visualizer always show the total targets and associated dynamic points in a couple of previous frames for the current frame?

Thanks!

Kai

Hi Kai,

The target data is sent one frame after the point cloud, but the visualizer synchronizes these two data streams, so that you see points and target data from the same frame. You can set it to use one persistent frame so you only see data from one frame at a time.

Regards,

Justin

Hi Justin

Another question, in line 108-116 in graphUtilities.py, do you always use the box with the size of (0.25, 0.25, 0.5) for the target tracker?

Thanks!

Kai

Hi Justin

I am asking the bounding box of targets because I am trying to detect human fall. My initial idea is that the bounding box includes all the points associated with each target so that I can find the dimension of each target and use the dimensions to decide if a human is standing, sitting, or laying down.

Have you ever tried to plot the bounding box using the points from target indexes? It seems to me some of the points associated with a single target could spread pretty randomly, which makes the dimension estimate difficulty. What do you think?

Thanks!

Kai

Hi Justin

Yes, the averaging over frame is a good idea. In order to do the average, I need to know the point IDs for the same target over a few frames. I was wondering if the target ID represents the same target over a few frames?

Thanks!

Kai

Hi Kai,

Target ID will represent the same target - only exception occurs when a target is lost then re-allocated.

regards,

Justin

Hi Justin

I am working on the averaging method now. The problem I have is the number of points in point cloud is not the same as the size of indexes. Please note that I already considered the factor that point cloud is for frame N and index is for frame N-1. I save the number of points and indexes in the past 10 frames and compare them for the same frame. Do you see the same scenario?

Thanks!

Kai

Hi Kai,

I haven't had that issue. At what point are you grabbing the point cloud and indexes?

Regards,

Justin

Hi Justin

After the following piece of code, but I save the number of points without removing the static points and compare it the len(indexes) for the proper frame.

        if (numPoints): # only save the data in previous 10 frames
            self.previousCloud[:5,:numPoints,fNum] = pointCloud[:5,:numPoints]
            self.previousCloud[5,:len(indexes),fNum] = indexes
        self.previousPointCount[fNum]=numPoints

Thanks!

Kai

Hi Kai,

Are you also saving indexes without cutting them off? The indexes are also trimmed to remove indexes related to static points here - if you aren't fixing this you will have too few indexes.

Regards,

Justin

Hi Justin

I am not trimming the indexes as well. I save the number of points and indexes in the past 10 frames. Most of the times, the two numbers are the same but they are different from time to time. And the number of points is larger and sometimes it is smaller.

Thanks!

Kai

Hi Kai,

There may be an issue with the data transmit on device, or the parser could be making an error.  Which version of the Industrial Toolbox are you using?

Regards,

Justin

Hi Justin

I fixed this issue and it is a bug from my code. The following question is as follows:

Since a single target ID at different frames could represent different targets (i.e., one target doesn't create a tracker while another target takes the target ID), how to accurately calculate the target dimension using the averaging method?

Thanks!

Kai

Hi Kai,

For the most part, I think you can assume that a target ID will always be the same target. However, when making this assumption, you can also implement a check to try to determine if a target ID now represents a new target. For example, if you find the dimensions change very quickly, this target ID may be tracking a different person.

Tracker will do its best to maintain a single track for each person, so I think its best to operate with that assumption and put checks in place to see if tracker has made an error.

Regards,

Justin

Hi Justin

Thanks for your suggestion.

I did some tests using the averaging method and found that the dimension of target is very large compared to the human dimension if I keep the static points. The target dimension looks better when I only keep the dynamic points just like what you did for visualization. Do you think it is a proper way to detect human?

In addition, the position of target in target list is also defined with respect to the center of sensor, correct?

Thanks!

Kai

Hi Justin

Oh, I didn't realize the AoA algorithm is different for static and dynamic points. Is there any reference describes the different AoA algorithm?

Thanks!

Kai

Hi Justin

If you can point me to the code in SDK where it implements the AoA for static and dynamic points, I can check it by myself.

Thanks!

Kai