This paper presents a key-frame based video fingerprinting method in which the fingerprint matching is modeled as a two hypothesis testing problem. The perceptual fingerprints that uniquely identify the video content are extracted by non-negative matrix factorization (NMF) via Gaussian weighting in order to assure compactness and robustness to global luminance distortions. the system performance is further improved to enhance its robustness to geometric attacks by integrating the transform invariant NMF (T-NMF) indicies into the matching scheme. The overall performance is evaluated on TRECVID video sequences. it is shown that the proposed video fingerprinting method is highly robust to global attacks described by TRECVID and it can also handle the geometric attacks for the transformations used at indexing phase.