As one of the key techniques in inverse synthetic radar (ISAR), autofocus is to eliminate the phase errors caused by the target radial motion. The maximum likelihood (ML) estimation of phase errors was derived form the ISAR signal model, and an autofocus algorithm based on the ML estimation was proposed. In the algorithm, the phase gradient algorithm (PGA) was performed, and the signals of multiple scatters on a few range bins were utilized. Hence, the scatters might not be well isolated, and the troublesome phase unwrapping was avoided. Moreover, the algorithm also eliminated the accumulation errors in the phase error estimation. The imaging results based on the measured ISAR data demonstrate a good performance of the autofocus algorithm.