This thesis proposes an automatic approach to multi-perspective image stitching. Differently from previous stitching methods using global transformation such as homography, we frame non-homogeneous warping into Lucas-Kanade approach and optimize for multi-perspective images. Our algorithm consists of two stages: alignment stage and aggregation stage. The former uses non-rigid alignment after applying feature-based registration, the latter iteratively solves for multi-perspective results. Experiments on the numbers of images taken from different view-points showed that our algorithm improves the image stitching without using layering or blending techniques.