Abstract: In view of the problem that the global feature description is overly dependent on the precise positioning, background subtraction and tracking technology, and also to address the influence of the change of the angle of view, noise and occlusion, action recognition methods in video based on local feature description were studied. A human action recognition method based on discriminative regions was proposed. First, the video content through iterative training and filter process were analyzed, and the area of discrimination and distinction of regional representation in the video was automatically extracted. Then the model statistics and the extracted discrimination region were described by the bag of words. Finally, to determine the type of human motion were determined by the SVM (support vector machine). The methods proposed in this paper was demonstrated on the KTH and Youtube datasets. Results show that the method has a high recognition accuracy and is especially insensitive to the complex background interference.