Motion compensated prediction is one of the most important tools for achieving high compression rates in all video coding standards. However, the usually adopted translational motion model is not accurate for complex motion such as rotations or zooms. In this paper, a new video coding solution is proposed where geometric transform warped images of the last decoded reference frame are used as additional reference frames to provide better predictions for inter coded video frames in the context of a H.264/AVC standard video codec. For each warped block, a region growing algorithm is applied to segment the frame into regions with similar complex motion, represented by a specific set of warping parameters. The warping parameters of each region are used to create a warped image that will be inserted in the reference picture buffer of the codec. The technical novelty of this paper regards the higher order geometric transformation estimation method for any block in the frame and the region growing algorithm to select the most representative warped frames. Experimental results show PSNR gains of up to 0.42 dB for some video sequences characterized by complex motions.