This paper introduces a video representation based on dense trajectories and motion boundary descriptors for action recognition. Trajectories capture local motion information, while motion boundary descriptors (MBH) are robust to camera motion. The method uses dense optical flow to extract trajectories and computes descriptors such as HOG, HOF, and MBH. The MBH descriptor outperforms others, especially on real-world videos with camera motion. The approach is evaluated on nine datasets (KTH, YouTube, Hollywood2, UCF sports, IXMAS, UIUC, Olympic Sports, UCF50, HMDB51), where it outperforms state-of-the-art results. Dense trajectories provide better coverage of motion and are more robust to irregular motion than sparse methods like KLT or SIFT. The MBH descriptor is particularly effective in suppressing camera motion, improving performance on real-world datasets. The method uses a bag-of-features approach with SVM classification and spatio-temporal pyramids to combine different descriptors. Experimental results show that dense trajectories and MBH descriptors significantly improve action recognition accuracy across various datasets.This paper introduces a video representation based on dense trajectories and motion boundary descriptors for action recognition. Trajectories capture local motion information, while motion boundary descriptors (MBH) are robust to camera motion. The method uses dense optical flow to extract trajectories and computes descriptors such as HOG, HOF, and MBH. The MBH descriptor outperforms others, especially on real-world videos with camera motion. The approach is evaluated on nine datasets (KTH, YouTube, Hollywood2, UCF sports, IXMAS, UIUC, Olympic Sports, UCF50, HMDB51), where it outperforms state-of-the-art results. Dense trajectories provide better coverage of motion and are more robust to irregular motion than sparse methods like KLT or SIFT. The MBH descriptor is particularly effective in suppressing camera motion, improving performance on real-world datasets. The method uses a bag-of-features approach with SVM classification and spatio-temporal pyramids to combine different descriptors. Experimental results show that dense trajectories and MBH descriptors significantly improve action recognition accuracy across various datasets.