נושא הפרוייקט
מספר פרוייקט
מחלקה
שמות סטודנטים
אימייל
שמות מנחים
ניתוח שחיית חתירה באמצעות כלים של ראייה ממוחשבת ולמידת מכונה
Freestyle Swimming Analysis Using Computer Vision and Machine Learning
תקציר בעיברית
תקציר באנגלית
In the past, performance analysis of athletes, including swimmers, took time and relied solely on professionals’ vision. Recently, the integration of computer vision applications in sports has risen. With the advancement of deep learning, techniques of Human Pose Estimation (HPE) for identifying human joints in images and videos have become significantly more reliable and efficient. In HPE, we identify the person in the picture and represent their posture in space by marking joint locations and connecting them to a “skeleton”. Most existing tools for HPE were trained on data captured in non-aquatic environments, with varied body positions and filming angles that do not correspond to swimming. Moreover, the aquatic environment makes it more difficult to identify the swimmers’ joints. The objective of this project is to detect the joints of swimmers in freestyle swimming using side view videos with the aim of later providing movement analysis. This project is a part of research aimed at understanding how to use this type of technology to analyze swimmers’ techniques and provide them with correction feedback. My project consists of two parts. The first part is data collection and preparation for model training. We chose to focus on data of men swimming in freestyle with a body side view inside the water. Data was collected by filming the swimmers in a swimming pool and videos from external sources. Next, each video was split into frames tagged with Coco Annotator software according to a joint model containing joints relevant to the analysis. The dataset contains seven videos of different swimmers. Four training videos (776 frames), two validation (515 frames), and one test (107 frames). The second part is data analysis which includes training different models with a deep network based on YOLO (You Only Look Once) algorithm for joint detection and a model with pre-trained data for comparison. The output is a rectangular field identifying the swimmer (Bounding Box) and the joints’ location (Keypoints). Model evaluation was done using standard metrics such as mAP (mean average precision) and by observing the visual results of model predictions. After comparing several models, we concluded that the model that achieved the most accurate results underwent prior training and additional training on swimming data. The initial products show that using Pose Estimation algorithms yields good results in identifying the joints of swimmers. Our next goals in this research are to improve the accuracy of prediction, extend training to include women's videos, a deep analysis of the movement profiles of swimmers, and characterize swimmers based on various parameters