In traffic accident analysis, accurately estimating vehicle speed is essential for understanding the circumstances of an accident, assessing the extent of damage, and determining the cause. In this study, we propose a novel framework for estimating the vehicle speed from video data, even when part of the vehicle path is completely occluded. The framework operates without prior camera calibration or spatial information and comprises four key modules: vehicle tracking, wheel center coordinate extraction, occluded coordinate prediction, and speed estimation. The design of each module corresponds to the characteristics of forensic accident analysis, incorporating deep learning models, Kalman filters, and geometric properties to increase the accuracy. The framework is validated through simulations, real vehicle tests, and actual accident case studies, demonstrating its reliability and applicability in forensic accident investigations. This framework enables the estimation of speed variations over time, providing critical insights into the driver's crash perception and braking response. Consequently, it enhances traffic accident analysis and contributes to accident prevention.