This repository demonstrates the concept of a "Follow-me" robot that can locate a person of interest (or the user) and keep following them from a safe distance. The person of interest is located with a bounding box and tracked even when obstructed with obstacles or other humans. The distance of the person from the camera is computed and displayed below the bounding box.
Wait for demo video to load.
Deep learning models utilized
- YOLOv8: A single-stage object detection network that was pre-trained to detect multiple objects including humans.
- MiDAS: A monocular depth estimation network that was pre-trained to compute relative depth of image pixels.
- DeepSORT: An object tracking algorithm that uses deep learning based appearance descriptors and Kalman Filter to track an object accurately even when obstructed by obstacles.
Quick Start
- Prerequisites: Python 3.8+, PyTorch and jupyterlab.
Running the notebook
- Start Jupyter and run the
follow_me_robot.ipynbnotebook. - An example video is provided in
datadirectory but you can use your custom video as well.
Notes & Tips
- MiDaS model weights are downloaded automatically by the notebook during the first run. If running offline, download the weights in advance and point the notebook to the local path.
- For best performance on larger images use a machine with a GPU and install a CUDA-enabled build of PyTorch.