Rekognition tracking boxes

This project uses Amazon Rekognition to get bounding boxes around detected objects. The Rekognition response is then used to play the video and overlay the boxes. In addition, in this sample the bounding box become red if its too close to another box.

Prerequisites

• python 3.x
• AWS account with correct rights and cli configuration

Set up

• clone the repository
• run pip install -r requirements.txt inside the repository

Usage

Run the video detection process with:
python detect_video.py -v ./VIDEO_PATH/YUR_VIDEO.mp4 -b S3_BUCKET_NAME

Play the video with:
python play.py -v ./VIDEO_PATH/YUR_VIDEO.mp4 -d ./labels/YUR_VIDEO.mp4.json
You can pass an optional parameter to increase the frame rate as:
python play.py -v ./VIDEO_PATH/YUR_VIDEO.mp4 -d ./labels/YUR_VIDEO.mp4.json -f 25

To make the distance feature precise in case the video is not taken from a perpendicular top view you need to map the coordinate of the video with the coordinate of a plane.

Mapping coordinates

You can use the map.py file which creates a JSON file with mapping coordinates.
Usage: python map.py -v YOR_VIDEO_FILE.mp4
Click with your mouse to create points, press s to save to a coordinates.json file or press q to quit from the window. The result is a json file as following:

{
    "camera" : [[105,360], [0,1252], [1221, 465], [2133,1200]],
    "plane":[[0,0], [700,0], [700, 700], [0,700]]
}

Finally you can run the play script as:
python play.py -d labels/VIDEO.mp4.json -v video/VIDEO.mp4 -c coordinates.json -t 200
Where -t identifies the threshold distance between objects.

In the below image you can see how the 3D spacial distance is correctly respected.

Process explained

The detect_video.py file upload your video file on S3 and calls Amazon Rekognition, then save the respomnse to a json file inside the labels folder.

The play.py file parse the response creating the bounding boxes and overlay the results over the video.

Limitations

The Rekognition call doesn't support response pagination. Can be easily implemented with further calls using the NextToken response parameter (see official documentation)