VUT FIT POVa 2018 Pedestrian Tracking

⚠ This project is deprecated and isn't maintained.

Computer Vision (POVa)
Faculty of Information Technology (FIT)
Brno University of Technology (BUT)

We have implemented a computer vision system intended for tracking pedestrians in observed scene. Our system is capable of detecting people in images from two cameras. Detected bodies are then matched together based on similarities in their histograms to make a pair of images of the same person. Detected person is then located in 3D space using triangulation, which uses depth planes in 3D space and their intersection. Located frames of people are then tracked to form path in space over time.

Team in alphabetical order: Lukáš Petrovič @flaxh, Filip Šťastný @xstast24, Martin Vondráček @mvondracek,

Path visualisation of the tracked person (green) in an observed scene

Scene observed by two cameras (left) and the result of pedestrian tracking (right)

Person detection

Installation

1. Python 3.7 is required.
2. Please create a virtual environment for this project.
3. With activated virtual environment, run: pip install -r requirements.txt in the project folder.
4. Prepare selected person detection backend — one from the following:
   • Download OpenPose GPU binaries. and extract them to openpose. Run openpose/openpose-1.4.0-win64-gpu-binaries/models/getModels.bat.
   • or download OpenPose CPU binaries and extract them to openpose. Run openpose/openpose-1.4.0-win64-cpu-binaries/models/getModels.bat.
   • or download just OpenPose model and save it to openpose/pose/coco.
5. Select detector used in main.py according to your detection backend from previous step.
   • OpenPoseBinaryDetector if you use GPU or CPU binaries.
   • OpenPoseDetector if you have downloaded just OpenPose model.
6. Please download testing data from https://github.com/mvondracek/VUT-FIT-POVa-2018-Pedestrian-Tracking

Run

./main.py

Documentation

• project report
• presentation slides (pptx with animations, pdf)
• supplemental materials