Abstract: The SURF (Speed Up Robust Features) feature extraction algorithm and FLANN (Fast Library or Approximate Nearest Neighbors) feature matching algorithm in current video stitching technology have the problems of feature point extraction errors and low feature point matching accuracy in harsh environments of fully-mechanized working face. An improved SURF−FLANN feature extraction and matching algorithm for video stitching of fully-mechanized working face is proposed. To improve the accuracy of feature point extraction, the improved algorithm extracts SURF key feature points of video images by changing conventional Gaussian filter to advanced bilateral filter, and improves the descriptor operator by adding feature point 4−domain feature point descriptor information to the feature vector. This improvement further improves the description of feature points. To improve the speed of feature point matching, The R−FLANN (Random sample consensus-Fast library or approximate nearest neighbors) feature matching algorithm is proposed. The R−FLANN algorithm uses RANSAC algorithm to get the matching prior information of feature points and eliminate the unmatched and mismatched feature points to improve matching speed of feature points. In order to verify the improvement effect, ablation experiments are conducted to verify that the improved SURF−FLANN feature extraction and matching algorithm effectively enhances the feature extraction and matching accuracy of video images of fully-mechanized working faces. Comparison experiments of feature point extraction and matching are carried out by the proposed method with SIFT+FLANN, Hairrs and SURF+FLANN feature extraction and matching algorithms, and the results indicate that the proposed method has the highest feature extraction and matching accuracy and matching speed, with the accuracy and matching speed reaching 81.47% and 51.47fps.Video image stitching comparison experiments are carried out by the proposed method with SIFT+FLANN, Hairrs and SURF+FLANN feature extraction and matching algorithms, the results indicate the proposed method has the best stitching performance in terms of clarity, contrast, entropy, and stitching speed.The best results are obtained.