Methods:
For Person Detection from image, YOLOv5 object detection algorithm is used.
Each cropped person image is fed into a Deep Learning model that uses Triplet loss and Resnet-50 model as backbone. A size of 1x512 embedding vector is generated for each cropped image by this model.
Cosine Similarity is measured between all the generated embeddings of the current image frame and all the previous frames and using a novel algorithm each person is assigned a unique Identification number.
Tracking is performed using the unique identification number associated with each person.
Homography Transform is used to map each person's position from a camera image to a 2D Floor plan.
Gaussian Kernel and Kernel Density Estimator(KDE) are used to generate Heatmap of person's movement
Fig: Homography Transform from image to Floorplan
Fig : HeatMap Generation using Gaussian Kernel