Classification and Role Estimation of Objects in Distorted Video

Abstract

Object tracking plays a vital role in computer vision, especially when handling distorted video frames. The challenge escalates when confronted with frames that are not in their original state. Accurately determining the positions of objects within these distorted frames is essential for effective video analysis. This research introduces a thorough framework to improve analysis, utilizing soccer match videos as a case study. It focuses on addressing two crucial elements: correcting distortions and estimating player positions. We employ a Convolutional Neural Network (CNN) to correct the distortions in the video. Subsequently, a Single Shot Detection (SSD) model, augmented by a Kalman Filter (KF), is applied to accurately extract and track the positions of individual players throughout the match. The proposed distortion correction approach outperforms existing methods, showcasing a notable improvement of 0.1% in Structural Similarity and 4.1% in Mean Square Error metrics. This indicates a superior accuracy in restoring the visual fidelity of the soccer match video compared to other distortion correction techniques. In the player position analysis phase, the integrated SSD model with KF is employed to precisely locate and track each player’s movements. The extracted player positions are then subjected to a clustering algorithm to assign specific roles to players based on their interactions and movements on the field. Experimental results reveal a commendable 78.85% accuracy in player role assignment when compared to ground truth data. This study presents a robust and effective framework for video analysis, using soccer match footage as a specific application. It integrates sophisticated techniques for both distortion correction and player position estimation

Date of Award	2023
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	David Hunter (Supervisor) & Jungong Han (Supervisor)