When I am adding Nvdspreprocess with tracker three is a excepted FPS drop, How can we solve this!

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Please provide complete information as applicable to your setup.

**• Hardware Platform ---------------> GPU
**• DeepStream Version ------------> 7.0
**• TensorRT Version ----------------> 8.6
**• NVIDIA GPU Driver Version ------------> 545

I am testing with deepstream-python-apps deepstream-preprocess-test and I added tracker after nvinfer, how queue is adding I have done the same way ! When I am adding two cameras for testing 25 FPS goes to 12-13 FPS for both camera !
Could you share me why it’s happening and How I can re-solve it ???

3

please apply the following patch. 40000=1000000/max_fps = 1000000/25 = 40000(ms).

-MUXER_BATCH_TIMEOUT_USEC = 33000
+MUXER_BATCH_TIMEOUT_USEC = 40000
-    streammux.set_property("batch-size", 1)
+    streammux.set_property("batch-size", number_sources)
4

@fanzh
I have done the same things I checked for 5 cameras batch size is 5 (streammux) , pgie batch size is 16

FPS drops to 18 FPS and one more thing is stream delay is there for 10-15 sec !!!

How I can enhance it ?

5
  1. if testing 2 rtsp sources, is the fps still 12~25?
  2. noticing you are testing 5 rtsp source, did you set MUXER_BATCH_TIMEOUT_USEC correctly? the new value should be 1000000/max_fps.
  3. please add this code and enable NVDS_ENABLE_LATENCY_MEASUREMENT=1 , NVDS_ENABLE_COMPONENT_LATENCY_MEASUREMENT=1 to add Latency measurement. then you can check which plugin consume too much time.
6

I set MUXER_BATCH_TIMEOUT_USEC is 40,000, and streammux batch size is num_sources, 25 FPS drops to 18 FPS after sometime it’s increase to 25 FPS.

7
  1. do you mean there is no fps drop issue again? are both testing with 2 rtsp sources or 5 rtsp sources are fine?
  2. about “10-15 sec delay” issue, did you add other code modifications? do you mean using player plays the output rtsp has 10-15 seconds delay? if replacing udpsink with nveglglessink, is there still delay issue? wondering if it is related to rtspserver.
9

Hi @fanzh

When I’m running NGC trained model “models/Primary_Detector/resnet18_trafficcamnet.etlt”, It’s working fine and with batch push timeout 40000, allmost FPS is 25.
But when I am using custom yolo model yolov4-tiny (darknet) with the help of Deepstream-yolo repo
with Nvdspreprocess I am getting “CUDA failure: an illegal memory access was encountered in file yoloPlugins.cpp at line 261”

This yoloPlugins.cpp is in Deepstream-yolo

Can you help me out to solve the problem ??

10

About “an illegal memory access”, it seems that this issue is not related to the original issue. could you open a new topic to focus on the new issue? Thanks!

11

Okey ! I will do this. Thank you.

12

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.