How to modify values to improve accuracy of Tripwire and ROI?

Accelerated Computing Intelligent Video Analytics Metropolis Microservices for Jetson
1

Hi

We have setup VST and AI-NVR, but the Tripwire and ROI are not really robust in different scenes.
The startup command is
sudo docker compose -f compose_nx16.yaml up -d --force-recreate

From the demo video, obviously you can see they get bad detecting and counting results.
We expect that one person passes through the line, in or out will add 1.
Could you guide us how to set tripwire to get better results ?
How to modify values to improve accuracy of Tripwire and ROI ?

Thanks

3

I will check and feedback later.

4

Hi

I enable RTSP output.
It can accurately identify people, but the people counter of Tripwire are inaccurate.
Which configuration can improve the accuracy of detection?

Thanks

main
main1165×813 132 KB

5

Seems the video is 60fps. Can you have a try with 30fps video?

6

Hi

I try to reduce to 30fps. The detection seems to be a little better.
If our width and height of video source are 1280/720, Do I need to change these values (tracker-width and tracker-height) of sector nvtracker in ds-config-0_nx16.yaml ?

  • type: nvtracker
    name: tracker
    properties:
    ll-config-file: “/ds-config-files/pn26/config_tracker_NvDCF_PNv2.6_Interval_1_PVA.yml;/ds-config-files/pn26/config_tracker_NvDCF_PNv2.6_Interval_1_PVA.yml”
    ll-lib-file: “/opt/nvidia/deepstream/deepstream/lib/libnvds_nvmultiobjecttracker.so”
    sub-batches: “2:2”
    tracker-width: 960
    tracker-height: 544

Thanks

7

Hi

Does config_infer_primary_RN34_PN26_960x544_dla0_orin_unprune_nx.txt means that the engine only analyzes the input video streaming with 960x544 ?

[property]
gpu-id=0
net-scale-factor=0.0039215697906911373
uff-file=/pn26-files/peoplenet_V2.6_resnet34_hr_noreg_ph2_dla.uff
uff-input-dims=3;544;960;0
uff-input-blob-name=input_1
int8-calib-file=/pn26-files/peoplenet_V2.6_resnet34_hr_noReg_ph2_dla.cache
#model-engine-file=/opt/nvidia/deepstream/deepstream/rosie-perception-app/ds-rosie/models/pn2.6/peoplenet_V2.6_resnet34_hr_noreg_ph2_dla.uff_b4_dla0_int8.engine
model-engine-file=/pn26-files/dla0_pn26_jp6_halfmem_bs3.engine
labelfile-path=/ds-config-files/pn26/labels.txt
batch-size=3
process-mode=1
model-color-format=0

Thanks

8

Please don’t modify nvinfer/nvtracker configure value. Is it possible to configure your IP camera to 30fps? We found FPS is the root cause of inaccuracy of the analytics.

9

Yes.
I test with video files and have reduced to 30fps, but the missing rate is about 50% for my case.
We try to modify some configurations to improve it.
Can you give us which file can we try? or input video resolution ?

10

Can you share the 30fps video to have a check in my side? You can share with private message if don’t want to public it.

12

The analytics will use the center of the bottom edge of bbox(the center of people’s foot) to decide if people cross the line of tripwire. Please ensure the tripwire cover the center of bottom edge of all bboxes which you want to count.

image
image1380×772 50.5 KB