This paper presents a feature-based early-termination approach for multiple reference frame motion estimation in H.264/AVC. Early-termination on frame level is adopted to simplify original full search algorithm. Meanwhile, according to an analysis on the reasons that some special video contents require more reference frames than others, three feature-based detectors are designed to detect those special contents. Moreover, an adaptive SAD criterion is set to guarantee that the prediction error is below a certain level. Simulation results show that the proposed algorithm is effective in reducing the computational complexity of motion estimation in H.264/AVC with an acceptable coding efficiency loss.