Low fps in back to back detector

Accelerated Computing Intelligent Video Analytics DeepStream SDK
,
1

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU):jetson(orin-nano)
• DeepStream Version:7.1
• JetPack Version (valid for Jetson only):6.1
• TensorRT Version:10.3.0
• Issue Type( questions, new requirements, bugs):quesion
I am implementing 3 back to back detector , primary detector detect vehicles , secondary detector detects lp region and tertiary detector detect lp text (secondary mode). This is my pipeline :
h264parser → decoder → streammux → pgie → tracker → nvvidconv_postpgie → caps_filter → queue1 → sgie1 → queue2 → nvvidconv_postsgie1 → caps_filter2 → queue3 → sgie2 → nvvidconv → nvosd → sink

After running the code i am getting 6-7 fps . I have used different properties of nvinfer to increase the processing speed like batch size and interval still not improved the fps.

My question is If I only run primary detector continuously and activate secondary and tertiary detector only when vehicle present in the zone , will it lead to increase the fps if yes?How to do it?

3
  1. how many sources? RTSP or local file? could you share the media information like fps, resolution? please refer to this link for performance improvement. especially setting “clocks high”, solution 5 and Solution 12.
  2. if still can’t work, please the log of “sudo tegrastats”, wondering the decoding and GPU usage.
4

single source , local h264 file , fps:20 , resolution:3840 × 2160 but downscaling it in 1920 x1080 in streammuxer.
I have implemented all the suggestion you provided but not significant improvement in processing speed.
What I want to achieve is primary detector will running continuously but secondary and tertiary detector only activate whenever vehicle is inside the specific zone else rest two detector should be deactivated.How to do that?

5

please refer to DeepStream ready-made sample deepstream_lpr_app. the three models are car detection, LP detection, LP recognition respectively.
2. the current workflow is already the same with your description. if pgie finds vehicles, objects will be generated, then two sgie will work. if no objects, two sgie will do nothing.
3. to narrow down this issue, can you remove two sgie temporarily to check if the fps can reach to 20. if no, can you share the configuration of nvinfer and the log of sudo tegrastats” when running the app?

6

1.I am getting 20 fps when i removed sgi1 and sgi2.
2.I am attaching my code ,I am using tee and valve element to contral the buffer flow sgi1 and sgi2 remain deactivate until vehicle is not reached roi once it present in roi sgi1 and sgi2 wil activate but whenver vehicle is inside the roi, i am getting below error.

anpr.txt (17.0 KB)

Error: gst-stream-error-quark: Internal data stream error. (1): /dvs/git/dirty/git-master_linux/deepstream/sdk/src/gst-plugins/gst-nvinfer/gstnvinfer.cpp(2423): gst_nvinfer_output_loop (): /GstPipeline:pipeline0/GstNvInfer:secondary2-nvinference-engine:
streaming stopped, reason not-linked (-1)

7
  1. please use nv3dsink on Jetson instead of nveglglessink. if always setting “drop” of both valve1 and valve12 to false, will the app run well?
  2. nvspreprocess plugin support ROI feature. please refer to DS sample deepstream-preprocess-test.
  3. if using three models, could you share the configurations files of two sgie and the log of sudo tegrastats” when running the app?
8

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

9

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