Trackable

Disclaimer: This project is meant for research purpose only.

Overview

Tracking humans in a hallway.

• The project deals with tracking humans in a narrow hallway under different lighting conditions.
• Unlike other MOT models, we aim to track people without any training, that means, all tracking is done online.
• There are many state-of-the-art models and architectures, that have been a large source of our inspiration (they are all listed under references below).
• The videos were obtained from this link.

Getting Started

It is suggested that you create a virtual environment. The requirements are given below, we've also included a requirements.txt

torchvision==0.6.0a0+82fd1c8
torch==1.5.0
dlib==19.19.0
numpy==1.18.2
scipy==1.4.1
opencv_contrib_python==4.2.0.34

Project Structure

The directory structure is as follows:

• Core: This contains the core modules for the project.
  • Detector:
    • This uses the YOLOv3 model to detect humans after every N frames.
    • Once the object is detected, the matching algorithm is used to see if the object is already being tracked, and thus avoiding re-identification.
    • If the object is a new object (or if re-ID fails), the next available ID is given to the new (or re-IDed) object.
  • Trackable:
    • This is the trackable object which contains a ID and other data regarding the object.
  • Tracker:
    • This implements the correlation tracker (from dlib) along with centroid tracking.
    • It updates the position of the trackable object in each frame.
  • Matcher:
    • This contains the matching function which returns a score of how similar two objects are.
    • The score is the weighted average of correlation score, Bhattacharyya score (obtained from their histograms) and the cosine similarity from the feature extractor.
  • Feature Extractor:
    • This contains the ResNet model (from PyTorch), which is used to perform feature extraction and return a cosine similarity score.
• Utils: This contains helper functions and constants used by the core module.

Demo

As you can see the two major problems are:

1. Misidentification: This can be solved by training on the objects beforehand. However the main aim for this project was to differentiate between objects without any training.

2. Mistracking: This could be helped with a better matching algorithm. DeepSort could be another alternative.

License

This project is licensed under the MIT License - see the LICENSE.md file for details.

References

• Correlation Tracker
• Multi-Object Tracking with dlib.
• DeepSort.
• Tracking without bells and whistles.
• Person re-ID.
• Kalman Filter.