The container includes a web server that is listening after an incoming HTTP POST message to the /lookforperson path. The incoming message must include a video that is to be analyzed on a parameter named video. Based on the environment settings it will look for a person in the frames of the video, and if found it will do an HTTP POST request to a web server address of choice and attach the detected frame, with the highest confidence level, and the details of the detection.

• latest: With OpenCV compiled to run on a Synology NAS
• jetson: Using nvcr.io/nvidia/l4t-ml:r32.5.0-py3 as base and tested on Jetson Nano 2GB Developer Kit

I've created a very simple docker-compose file that can be used to test the service. The demo is using Node-RED to present the results from the video file analysis. Use your own mp4-file and post it to the service endpoint that will analyze the frames and send the result - if a person is found - to display in the Node-RED UI.

1. Start the setup with the file in the github repository

docker-compose -f docker-compose-demo.yml up

2. Open the Node-RED local webpage http://127.0.0.1:1880

3. Post a video file using curl to the exposed http port. E.g. use a sample file such as the this one found from https://videos.cctvcamerapros.com/surveillance-video-samples/15-frames-per-second.html

curl -v -F [email protected] http://127.0.0.1:8080/lookforperson

HTTP post to /lookforperson

• video: (POST parameter) File to be analyzed

HTTP post to the address specified

• file: JPG file object attached with the post including the detection

The HTTP POST response with be a JSON document and if successful (found a person) it will have the format as per below. params are the incoming variables used in the analysis.

{
   "success":true,
   "confidence":"0.71319485",
   "width":14,
   "height":31,
   "ratio_diff":10,
   "comment":"Person was found",
   "params":"{\"urlpath\": \"http://192.168.2.244:1881/reportpersoninvideo\", \"frames\": 5, \"conf\": 0.2, \"good_enough_conf\": 0.7, \"width_person\": 40, \"height_person\": 80, \"width\": 640, \"height\": 480, \"ratio\": 20}"
}

I use the container to get rid of any CCTV notification noise - i.e. I'm alerted only if a person is detected around my house and not only by the amount of changed pixels in a frame.

1. In MotionEye OS I have 6 video streams that is triggering recordings based on changes in the number of pixels in a frame
2. If a motion is detected MotionEye will run a command when the file has been saved

curl -F video=@%f http://192.168.2.244:8080/lookforperson

3. curl does a post of the mp4 file to the docker container
4. The container will post the result to a Node-RED instance that listens to a HTTP POST with a video and it will send the image further to my Discord server

version: "2"
services:
  people_recognition:
    container_name: opencv_findperson
    build:
      context: .
      dockerfile: Dockerfile
    image: "nidayand/video-people-detection-with-http-post"
    ports:
      - "8080:8080"
    environment: 
      # URL to post image result to. Path will be appended with a field called "file" with type image/jpg
      - URLPATH=http://192.168.2.104:1880/videonotify

      # Required minimun level of confidence when detected a person on the analysis of a frame. Default 0.2
      - CONFIDENCE=0.5

      # Good enough confidence level. Skip additional frame analysis when this level has been reached. Default 0.8
      - GOOD_ENOUGH_CONFIDENCE=0.7

      # Every x frame will be checked for a person
      - FRAMES=5

      # Resizing frame for analysis to improve performance. Default values 640/480
      - WIDTH=640
      - HEIGHT=480

      # MAX % of image that can be a person
      # Set to 0 to disable or remove entry
      - WIDTH_PERSON=40
      - HEIGHT_PERSON=140

      # Validate that height/width of a person is not above difference compared to max HEIGHT_PERSON/WIDTH_PERSON
      # Avoids strange dimensions e.g. 50x3 when more likely 50x25
      # Set to 0 to disable or remove entry
      - WIDTH_HEIGHT_RATIO_COMPARE_DIFF=32

      # Max size for the video to be analyzed in MB. If set to larger files make sure the set GOOD_ENOUGH_CONFIDENCE
      # to limit the amount of frames required to be processed
      - MAX_SIZE=20

docker run -p 8080:8080 \
  -e "URLPATH=http://192.168.2.104:1880/videonotify" \
  -e "CONFIDENCE=0.5" -e "GOOD_ENOUGH_CONFIDENCE=0.7" \
  -e "FRAMES=5" \
  -e "WIDTH=640" \
  -e "HEIGHT=480" \
  -e "WIDTH_PERSON=40" \
  -e "HEIGHT_PERSON=140" \
  -e "WIDTH_HEIGHT_RATIO_COMPARE_DIFF=32" \
  -e "MAX_SIZE=20" \
  --name opencv_findperson --restart always nidayand/video-people-detection-with-http-post:latest

docker-compose build

• Container includes a web server that listens for a POST request on the "/lookforperson" path
• The POST request must include a video file to be analysed. Parameter "video"
• Container is searching for a person in the video frame by frame (every [FRAME env.] frame)
• If a person is detected with a level of confidence above [CONFIDENCE env.] an URL-post will happen to `[URLPATH env.]``
• If [GOOD_ENOUGH_CONFIDENCE env.] is set the frame scanning will stop when a value is found above otherwise it will assess the entire video
• If the width or height is larger that [WIDTH_PERSON] or [HEIGHT_PERSON] the frame will be skipped
• If the dimensions of a person detected has a larger ratio that [WIDTH_PERSON]/[HEIGHT_PERSON] the frame will be skipped
• The result will be returned in a JSON string

{"success": true, "confidence": "0.7131952", "width": 14, "height": 31, "ratio_diff": 10, "comment": "Person was found"}

By mounting a local directory to /debug you will get frame by frame analysis that can be used to tune your parameters.

  ...
    volumes: 
      # USED FOR DEBUG WHEN TUNING PARAMETERS. If set will save each frame analyzed in a subfolder
      - "./debug:/debug"

Each frame will be stored as a jpg with comments on why it was not the final frame. Note that this will affect performance and use only when tuning your parameters.

Examples

• Conf<: the confidence level of the person detected is lower than previously found. The box inidicates the position of the frame where the higest confidence was detected
• Ratio 30>20(30x47): The ratio of the box is higher than maximum allowed and in brackets are the dimensions of the box identified. 20 is in this case the WIDTH_HEIGHT_RATIO_COMPARE_DIFF value
• Width>: The width is larger than WIDTH_PERSON
• Height>: The height is larger than HEIGHT_PERSON