Fine-grained time sequence action detection method

The invention belongs to the field of computer vision, and provides a fine-grained time sequence action detection method, which comprises the following steps of: 1, carrying out preprocessing and feature extraction on an original video in a data set; step 2, constructing a feature pyramid for the video features obtained in the step 1; 3, encoding the feature sequence of each level in the feature pyramid obtained in the step 2 by using a global time sequence feature encoder, and generating a rough prediction result; and 4, performing boundary correction on the rough prediction generated in the step 3 by using the corresponding lower-layer features of the feature pyramid obtained in the step 2 to obtain a final action instance detection result. The invention provides an anchor-free time sequence action detection method based on global feature perception and local boundary correction, the accuracy of fine-grained time sequence action detection is effectively improved, and the method has great significance in the