This paper proposes an extended Kalman filter (EKF)-based and training symbol aided carrier frequency off-set (CFO) estimation scheme for the uplink OFDM systems. Typically, in EKF-based estimation scheme, the measurement equation is a function of the CFO and channel coefficients. Therefore, a vector EKF has to be employed to estimate all the unknowns, which may sometime result in convergence problems. To avoid these problems, the proposed scheme uses a scalar EKF. The user signals are first separated by using multiple-access interference cancellation. Then, the unknown channel coefficients in the measurement equation are replaced with a non-linear function of the CFO so that the scalar EKF can be employed. The observation noise power is analyzed and its approximation is used in the EKF algorithm. Several numerical examples are presented to validate the efficacy of the proposed scheme. It is shown that the proposed scheme can achieve the Cramer-Rao lower bound (CRLB) when the number of users is small, whereas it degrades when the number of users increases. We also compare its computational complexity with several existing schemes.