A robot state estimation algorithm based on the vision feedback is proposed in the paper. The algorithm consists of an image feature detector and an extended Kalman filter (EKF) based estimator. The detected image features are scale-invariant and provide a robust representation of moving objects and static landmarks in the environment. The recursive EKF-based estimator is utilized to determine the pose and velocity of moving robots. Experiments are carried out on a hand-held binocular camera to verify the performances of the proposed state estimation algorithm. The results show that the integration of the image feature detector and the state estimator is efficient in highly dynamic environments.