IWR6843: EKF Tracking on DSS

Hi Attiya,

Currently, this demo does not exist. However, the tracker can be compiled for the DSP.  You could then run the tracker on the DSP, and simply send all of the ouput (point cloud and tracker out) to the Cortex R4F.

What are you trying to achieve with this set-up?

Regards,

Justin

Hi Justin,

Thank you for the quick reply. 

We were thinking that DSP is sitting idle when HWA performs repetitive tasks (FFTs) so why not to do tracking on it in parallel and Cortex R4F is used solely for the communication purpose. I was thinking this may cause some deadline issues on DSP that's why asked if there exists any demo using all three (HWA, DSP, uC) together?

You said "tracker can be compiled for DSP" meaning I need to include gTrack library and the task_app.c (for EKF ) after when PointCloud data is generated on DSP and then send the combined output (PointCloud and TrackerList) to the Cortex R4F.

What we are are trying to see is if HWA along with DSP provides (time-wise) fast detection and tracking capability than the existing setup (DSP for detection in the current frame, and Cortex R4F for tracking in the previous frame) or not.

Regards,

Attiya

Hi Attiya,

To use the tracker in the DSP, you would need to:

1. Ensure the proper library is compiled. It will be called libgtrack.ae674 (ae674 is the file extension for DSP libraries)
2. Include this library with the DSP lab
3. Copy/Rewrite task_app.c to work with the data flow on the DSP.

Currently, the DSP is memory limited in the People Counting demo, so I am not sure if you will be able to fit the tracker into this core.  You may have more success trying this with the OOB demo first.

Regarding time to detection:

• Currently, the people counting config has an active frame time close to 25 ms (rough estimation). Lets assume that on average, it requires another 15 ms to generate the point cloud. 
• Tracker should run in 1 ms or less on DSP (compared to 5 ms or less on Cortex R4F)
• So best case on the DSP is 41 ms (25 ms + 15 ms + 1 ms)
• Otherwise, point cloud is sent to Cortex R4F, then runs through tracker.
• Point Cloud is output and tracker is run (tracker run time is 5 ms or less).
• Assume best case data transfer for DSP to MSS is 2 ms, then time to detection is 47 ms (25 ms active frame + 15 ms point cloud + 2 ms transfer to MSS + 5 ms run tracker)
• Detection is not sent until next frame, roughly ~50 ms later

My opinion is that the greatest savings you can get is to transmit target detection as soon as they happen.  You may be able to rework the MSS to send detections as soon as they happen to minimize detection time. This will be easier than adding the tracker to the DSP.

Regards,

Justin

Hi Justin,

Thank you for providing a very detailed response. I will work on either of this and post a new thread if I face any issue. 

Is there any demo where HWA and DSP are used together. Just want to see how they share workload in C-Code and then I will close this thread.

Thank you,

Regards,

Attiya