Skip to content

Latest commit

 

History

History
110 lines (93 loc) · 2.49 KB

README.md

File metadata and controls

110 lines (93 loc) · 2.49 KB

Between us API

API that receives an image as input and responds with a JSON that contains:

  • bounding boxes of detected human poses
  • for each human pose, a boolean indicating whether or not they are respecting the social distancing norm

Setup

Requirements

  • Anaconda
  • GPU + CUDA 10.0

Create an env with all required dependencies

cd betweenus-api
conda env create -f betweenus.yml

Install specific version of Monoloco

git clone -b social-distance --single-branch https://github.com/vita-epfl/monoloco.git
cd monoloco
pip install -e .

Usage

Start the server on port 8000

python server.py

Example output

curl --request POST \
  --url http://localhost:8000/predict \
  --header 'Content-Type: multipart/form-data' \
  --form 'image=images/fun/famous_beatles.jpg'

{
  "boxes": [
    [
      307.69431072732675,
      515.5085627395914,
      457.87624591329825,
      839.7450383346273,
      0.7588235294117648
    ],
    [
      82.49378531319755,
      517.3484599427406,
      270.4304853166853,
      851.9474995299156,
      0.7117647058823531
    ],
    [
      487.85962255252815,
      532.6579805930949,
      658.3666958068469,
      853.0940946022176,
      0.6470588235294118
    ],
    [
      686.9696666646448,
      525.6854031959676,
      865.8030994486364,
      865.5501925071574,
      0.6764705882352942
    ],
    [
      745.0856167811045,
      473.97913734957297,
      761.524368570458,
      547.1124764199583,
      0.6294117647058824
    ]
  ],
  "too_close": [
    true,
    true,
    true,
    true,
    false
  ]
}

Approach

A multi step approach to detecting people and determining the distance between them:

  1. Detect the poses (keypoints & bboxes) using openpifpaf
  2. Transpose to birds-eye view using Monoloco
  3. Calculate distances between people
  4. Calculate future positions based on orientation

NOTE The poses alone are not enough to accurately calculate distances, we need info about the depth.
Estimating depth from bounding box sizes yields very poor results, since they vary in shape very much.
Monoloco solves this by transposing in 3D space from monocular RGB images.
This enables a much more accurate distance computation.

Credits