Traditional feature extraction techniques like the KLT, Harris and Wavelet work only in the uncompressed domain. Hence an additional step of decompression is required before any of them could be applied. We propose a two-level technique for extracting high-level feature points directly from JPEG compressed images. At the first level, the Discrete Cosine Transform (DCT) blocks having high activity content are filtered using a variance measure. At the next level, a DCT block centered at every pixel present in the filtered block is constructed from the neighboring DCT blocks. Feature points are then selected by analyzing the AC coefficients of the DCT block centered about it. The proposed method is simple and efficient. The extracted feature points were found to be rich in information content, which could be used for image registration. The results of this technique showed almost the same amount of repeatability between two images with 60% to 70% overlap, when compared with techniques available in the uncompressed domain. The features thus extracted can directly be used to calculate the motion parameters between two images in the compressed domain.