README.md

SwimmerDetection

This project was done during an internship within the VAADER (Video Analysis and Architecture Design for Embedded Resources) team in IETR Lab. The internship is part of a project financed by DIGISPORT whose goal is the realization of a system for monitoring a swimming activity in a swimming pool, based on image and video processing algorithms and artificial intelligence.

🏊‍♂️ Description of the project

The main purpose of the project is to study methods of tracking a swimmer in a pool by embedded active vision in order to understand the swimming movement.

During my internship, I was focused on the detection of the swimmer body in the entire image with the use of different image processing algorithms in order to extract the area of the image containing the region of interest. This area is then transferred to the central processing system which performs a pose detection in order to analyse the swimmer’s movement.

I was in charge of the study and comparison in simulation of active vision algorithms allowing the localization of the swimmer in the image.

There 3 different algorithms :

• Color Segmentation
• Gaussian Mixture
• Deep Learning

The projects also contains a database of more than 2,600 images of swimmer with bounding box.

If you want to test by yourself, you have to fork the repository. After, you just have to execute the 2 commands to install all the packages needeed for the project and especially the package src :

pip install -r requirements.txt
pip install -e .

⚠️If you want Pytorch with cuda for GPU, you have to install it by your own.

📖 How to use the package src

If you want to test the model directly on the database, here are some guidelines :

• Firstly, import the model you want to use and the method IoU_video to see the results :

from src.color_segmentation import ColorBB
from src.gaussian_mixture import GaussianMixtureBB
from src.deep_learning import Swimnet
from src.metrics.model_performance import IoU_video

• Then, load the model :

#### Gaussian Mixture ####

filename = "../models/GMM_model_test_vid_23_full"
model = GaussianMixtureBB(filename, threshold=0.5, graph_cut=False, use_time=False)

#### Deep Learning ####

filename = "../models/mobilenet-V3-small"
model = Swimnet("mobilenet-v3-small")
model.load_state_dict(torch.load(filename))

#### Color Segmentation ####

model = ColorBB("hsv", use_time=False)

• Finally, call the method IoU_video and specify the location of the database

IoU_values, stat_values = IoU_video("../data/images/Testset", "../data/annotations", model, debug=True)

If you want more details about the parameters of each model, there are well defined in the documentation of each class.

Presentation of the different method

Color Segmentation

Gaussian Mixture

Swimnet

Organisation

• \data-- all the data of the project
  • \images all the images of the database separated in 3 directories Testset, Trainset and Valset
  • \annotations all the annotations of the database organized as in images
  • \dataframes all the results of validation tests to find the best parameters
  • \figures figures for the report
• \docs some documents made for the project
• \models all the saved models
• \notebooks different notebooks with example of code to illustrate the projects and to have some example of the syntax in order to use the package
• \reports the report made for the project
• \src the package developed for the project. Each method or class is documented
  • \annotations
    • read_annotations.py methods to read annotations of the database
  • \metrics
    • model_performance.py methods to compute metrics and to compare models
    • validation.py methods to make validation
  • \preporcessing
    • bb_tools.py some tools for bouding box
    • dataset.py all the class and methods in order to create the dataset
    • water_surface_detection.py methods to detect the surface
  • color_segmentation.py class for the Color segmentation
  • deep_learning.py class for the Deep Learning models
  • gaussian_mixture.py class for the Gaussian Mixture model
  • kalman.py implementation of the Kalman Filter
• \scripts scripts which used the packages
• requirements.txt all the package you need to install so that you can use the project on your own