In H.264/AVC, the adoption of multiple reference frames (MRF) helps to improve the video coding quality because smaller residue can be generated by precise motion estimation (ME). However, the procedure of finding the suitable reference frame is a computation intensive task. In fact, if macro blocks that have homogeneous characteristic can be detected in advance, the computation burden can be released greatly. This paper gives an edge detection based pre-process MRF-ME algorithm and proposes a threshold decision criterion based on both PSNR and bit-rate analysis. Through experimental results, we find that maximum ME time reduction can be achieved by configuring threshold linearly with quantization parameter (QP). With this adaptive edge detection pre-process MRF-ME algorithm, average 37.48% ME time can be reduced with negligible video quality degradation.