People Counting

This repository was created as part of a technical assessment for estimating the number of people in a video frame using computer vision.

Prediction Results:

• Average number of people detected per frame (0:18 to 0:19 time window): 172
• Estimated total number of people on the bridge: 2,111

All predicted results, including images with bounding boxes and labels, can be found in the results directory.

Installation

First, install the necessary dependencies:

sudo apt install ffmpeg
pip install -r requirements.txt

Data processing

To fetch the video from YouTube and extract frames between 0:14 and 0:32, run the following:

python data_processing.py \
  "https://www.youtube.com/watch?v=y2zyucfCyjM" \
  "data/video/" \
  "data/video/Drone Footage of Canberras HISTORIC Crowd.mp4" \
  "data/frames/output_%04d.png" \
  14 \
  32

Select the images

To select frames from the 0:18 to 0:19 interval (frames 100 to 125) for model inference:

mkdir data/selected_frames
for i in {0100..0125}; do cp  data/frames/output_$i.png data/selected_frames/; done

Model Setup

Prepare the YOLO-CROWD model and set up the environment:

mkdir model
cd model
git clone https://github.com/bytebarde/YOLO-CROWD.git
cd YOLO-CROWD
pip install -r requirements.txt

Inference

Run the model on the selected frames (note that the model weights are included in the repository for convenience):

python detect.py --weights yolo-crowd.pt --source /home/erwan2/Projects/people_counting/data/selected_frames

The predicted results will be saved at model/YOLO-CROWD/runs/detect/exp.