computer-vision-project

Project for Computer Vision Course from La Sapienza Univesrity of Rome.

This project aims to carry out two main tasks.

1. The first task concerns the issue of Bike Fitting, i.e. the initial bike configuration procedure, which is very important for biomechanical reasons, since an incorrect configuration can lead to problems with the cyclist's joints in the long term. However, this procedure is very difficult to perform alone, so my goal is to use a short video, in which to trace the cyclist's posture while pedaling on his bike, in order to calculate the distances and angles between the various joints to understand if the configuration is correct or make suggestions, such as raising or lowering the saddle, or moving it forward or backward.

2. The second Task has the objective of correcting the posture of MTB cyclists during descents on rough Trails, in fact given an input side image of the cyclist traveling downhill, the program identifies the wheels of the bicycle and the Key Points of the cyclist to calculate if the center of gravity of the latter is located inside the two wheels. This is because if the center of gravity is too far back, grip on the front will be lost, while if it is too far forward, the cyclist could fall forward if he encounters an obstacle such as a stone or a root.

Links

Video Demo

You can find a video demo of the website and its features at the following link Computer-Vision-Project-Demo.

Web Site

You can also visit the site at the following link cv-website.

Slides

It is possible to find the slides of the presentation of the project at the following link Computer Vision Slides.

OpenPose vs MediaPipe

For the project I didn't use OpenPose, but MediaPipe mainly for two reasons:

• the first concerns the fact that OpenPose is not optimized for CPU usage, and not having a device with a dedicated GPU, image processing required about ten seconds, while videos could not be analysed.
• the second reason is that while deploying on Google Cloud using OpenPose gave me problems, as I was unable to use GPU Cuda due to the fact that my test credit for Google Cloud is almost empty.

I decided to use mediapipe because in addition to the Pose Estimation it is also possible to obtain a mask of the human body useful for the second task of the project, when calculating the center of mass. Furthermore, through various searches I have found that it is more reliable than Openpose, as is written in the following article Detection of human body landmarks - MediaPipe and OpenPose comparison.

Folder Structure

.
├── backend                             # Computer Vision Tasks
│   ├── bike_fitting        
│   │   └── capture_video.py            # Bike Fitting Task (pipeline() is the main function)
│   ├── mtb_down_hill       
│   │   ├── yolo_object_detection
│   │   │   ├── yolov3.cgf              # Config File for YOLO
│   │   │   ├── yolov3.txt              # Classes List for YOLO
│   │   │   └── yolov3.weights          # Weights for YOLO (download it!)
│   │   ├── bike_detector.py            # MTB Downhill Task (pipeline() is the main function)
│   │   └── yolo_object_detection.py    # Bike Detection Task
│   ├── math_utils.py                   # Set of function for math problems
│   ├── mediapipe_pose_estimation.py    # Class for Mediapipe operation
│   └── print_utils.py                  # Set of functions for displaying plots
├── frontend                            
│   └── frontend.py                     # Process outputs of backend
├── doc                                 # Documentation files
│   └── etc.
├── static                              # Web Site static resources
│   └── etc.
├── templates                           # Web Site HMTL Code
│   ├── bike_fitting.html
│   ├── downhill.html
│   └── index.html
├── app.py                              # Flask root file
├── config.py                           # Configuration file for Flask and Enum
├── Dockerfile
├── requirements.txt
└── README.md

Requirements for local execution

Download yolov3 weight at pjreddie.com:

wget https://pjreddie.com/media/files/yolov3.weights

Install Python requirements:

pyhton3 -m pip install -r requirements.txt

Install ffmpeg for video encoding:

apt-get install -y ffmpeg

Steps to deploy on Google Cloud Platform

1. Create a bucket if it not exists.

   a. Creating a GCS Bucket for Your Files

    gsutil mb gs://cv-project-bucket
   

   b. Adding Permissions

    gsutil defacl set public-read gs://cv-project-bucket
   

2. Download repository

       git clone https://github.com/AndreaBe99/computer-vision-project.git 
   

3. Download the .weights file.

   cd computer-vision-project/backend/mtb_downhill/yolo_object_detection/
   wget https://pjreddie.com/media/files/yolov3.weights
   

4. Download and then rename Google Account Credential (It is private):

   mv source_file GAC.json
   

5. Export the Project ID

   export GOOGLE_CLOUD_PROJECT=computer-vision-project-372315
   

6. Set the project

   gcloud config set project computer-vision-project-372315
   

7. Build Docker Image, (i is the version number of the image)

   docker build --tag cv-website:v[i] .
   

8. Run locally

   docker run --rm -p 8080:8080 cv-website:v[i]
   

9. Deploy on Google Cloud Platform

   gcloud builds submit --tag gcr.io/${GOOGLE_CLOUD_PROJECT}/computer-vision-website
   
   gcloud run deploy cv-website --image gcr.io/${GOOGLE_CLOUD_PROJECT}/computer-vision-website
   

References

• Bike Fitting:

  • The site from which I was inspired for the graphics: MyVeloFit

  • The site I used as a reference for the body joint angle value: Bikedynamics

• MTB Descent Position:

  • The site I used to get all the attack position tips: mtb-dynamics

• General:

  • Site Template: free-css