Introduction

This project intends to re-wrap the openvino's Multi Camera Multi Target Demo into a more accessible way. All dependents( pre-built models) are included.

Requirements

• a x86 Linux host
• docker runtime
• X Window System display server

Prepare

• pull the image

docker pull ghcr.io/changtimwu/retail_ovmc_demo:main

• On your Linux desktop, launch xterm and use xhost command to permit X clients

xhost +

Run the image

• fixed arguments

  • -e DISPLAY=:0 --net=host -v /tmp/.X11-unix:/tmp/.X11-unix:rw: make X server transparent between host and containers

• arguments

  • -v work_path: specify the path where the input video files and output results to be store
  • -e DET_MODEL: specify the detection model
  • -e REID_MODEL: specify the ReID model
  • -e video_file: the video files to be analyzed in work_path or video source in URL form

• optional arguments

  • -e no_show_id=1: show only the bounding box without the identification number

Examples

• analyze a mp4 file

docker run -itu root -e DISPLAY=:0 --net=host -v /tmp/.X11-unix:/tmp/.X11-unix:rw --name openvino_multicam_reid -v <your media/save path>:/opt/work -e video_file=abc.mp4 --rm retail_ovmc_demo

• analyze multiple RTSP streams

docker run -itu root -e DISPLAY=:0 --net=host -v /tmp/.X11-unix:/tmp/.X11-unix:rw --name openvino_multicam_reid -v <your media/save path>:/opt/work -e video_file="rtsp://user:password@192.168.1.12/stream2,rtsp://user:password@192.168.1.13/stream2" --rm retail_ovmc_demo

Custom configs

• Create a directory named configs in your work_path
• take configs/person.py in this repo as a template and copy it into the work_path/configs/
• now you can customize it. For example, you can enable the analyzer

analyzer = namespace(
    enable=True,
    show_distances=True,
    save_distances='',
    concatenate_imgs_with_distances=True,
    plot_timeline_freq=0,
    save_timeline='',
    crop_size=(32, 64)
)

the analyzer window would pop up

List of models

• Detection model -- specified with -e DET_MODEL=
  • person-detection-retail-0013
• ReID model -- specified with -e REID_MODEL=
  • person-reidentification-retail-0277
  • person-reidentification-retail-0286
  • person-reidentification-retail-0287
  • person-reidentification-retail-0288

Output

while input video is playing, you're supposed to see the people identifed with colored rectangle and number.

detections.json and history_results.json would be saved into the working directory

Build and Deploy

• build the image by the command

make

• Proper Github Actions have been setup to automatically build the whole demo as a docker image. By default, it will publish the result image to github's Container registry.

Reference

• MCMT demo source repo
• the Dockerfile of the base image openvino/ubuntu20_dev`