SitwalaM/object_detection_football

Object Detection Project as part of a 46 hour challenge in the Africa Data Science Intensive (DSI) program # Object detection as a concept for automated football statistics ## Overview The project aims to determine a football player with possession of the ball using computer vision. Use cases for this include automatic possession statistics and potentially automatically obtaining xG (expected goals) statistics if we can capture when a short is taken by a player. The project uses a pretrained yolov5 model to identiy players and then draw a bounding box around the player with the ball. ## Data Source The data source for the project was obtained by scraping google image results in a browser console using a tutorial which can be found here. A script was then created to clean the url lists, remove duplicates then download the to a folder. [Link to data_download script] (github.com). The dataset is used merely to test the algorithm and not train it. Better results can be obtained by proving ground truth data which will help with calibrating out problematic scenarios like fans in the background and players in background who seem to be near to the ball in a 2-D perspective. Several annotating softwares are available that produce label formats that are accepted by common object detection algorithms ## Deployment and Use The demo pipeline can be fully run in the notebook provided which can run on Colab. To use the data download script ```bash pip install -r requirements.txt ``` Edit the file name inputs in the script as required. The algorithm is tested using gradio. It takes an input image and outputs an image with the bounding box. The following functions are used to input the image into the model: ```bash def predict(image): # main prediction function that gets image and returns pandas dataframe results = model(image, size=640) return results.pandas().xyxy[0] def processed(image): #inputs the image and draws the bounding box on closest player results = predict(image) results = calculate_cent …