This repository contains my dissertation. This project exploit features of the detected players like their gaze-orientation, pose and outfit using state-of-the-art machine learning methods to build a HOI pipeline made out multiple components that included player and ball localization and a player interaction classification module was built. For each component multiple approaches were considered and evaluated, the best performing methods were integrated into the HOI pipeline. The functionality to track individual players and their shot attempt was also implemented successfully.
-
Develop an simple HOI detection pipeline that is capable of localizing the player and the ball and the interaction between them(analogous to player action) when in not-occluded scenarios.
-
Enable HOI pipeline to do the classification of action and localization of players and the ball that takes into account occlusion.
-
Improve the pipeline such that it can classify the action of all the players and have the ability to track specific player’s shot attempts.
- Install all the python librabies in the requirement text
- Download the models for Action Classification, Ball Detection, Team Detection, Vptrees
- Put the downloaded folders inside the models folder
- Install Node.js
- https://nodejs.org/en/download/
- After Node.js installation:
- Go inside "posenetSever" folder
- run the following command: npm install
- This will install all the javascript libraries used that is saved in the package.json file
- Run the TeamDectection.py file
- python TeamDectection
SpaceJam: a Dataset for Basketball Action Recognition is used for used for the project. A varition of this dataset which consist of player poses (L2-Normalized) was generated for action-classification, download here.