We introduce an innovative approach for the precise velocity estimation of an ego vehicle from dashcam video, leveraging optical flow analysis, feature engineering techniques, and a two-stage model architecture. Data comprising video and ego velocity information were used. In the initial stage, a velocity interval label classification model was constructed, while in the latter stage, a regression model was used for ego velocity estimation. The models were tested on the Comma.ai 2 K19 Dataset, and a Root Mean Square Error of 0.994 m/s was obtained from 300 test videos, and 2.204 m/s from 4 external test videos selected from KITTI Dataset. These results highlight the substantial reduction in estimation. The models can be applied to car accident analysis and GPS-based ego velocity estimation methods to enhance road safety and traffic accident investigation.