detect people for each frame using yolo v3. compare each person coordinates (center of the binding box) with a bucket where each bucket contains the history of each person tracked during the previous frames. a score is evaluated by considering the distance from the last coord of the bucket and the next coord expected from that bucket according to its trajectory in time.
full idea can be found here: https://drive.google.com/open?id=1cwVWWydWiCwkSIiFqZewdeJUecBgM4L9
download YOLO weights
cd content
wget https://pjreddie.com/media/files/yolov3.weights
cd ..
install requirements:
python3.7 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
run the network
python predict.py
more sample videos available at: https://www.dropbox.com/s/6egotfb5n5a1lkp/Archivio_video.zip
if you want to calibrate your camera, you first need to estimate the camera matrix (use https://docs.opencv.org/2.4/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html as reference). for a quick estimator of the distortion coefficient (k1) we implemented a nice algorithm, check it out https://github.com/potpov/camera-correction
