A method and system for generating cascaded action candidate frames based on action boundary probability

Abstract

The invention discloses a method for generating a cascaded action candidate frame based on the action boundary probability. An input video to be detected is obtained by using a feature extraction network to obtain features, and a cascade model is used to predict the probability of each position belonging to an action boundary point. The cascade model is divided into two levels: the first-level model performs rough boundary prediction on the input features, and respectively predicts the probability that each position belongs to the start point, end point, and inside of the action; then the output of the first-level model is connected in series with the original features As a new feature input to the second-level model, let the second-level model predict the probability of each position belonging to the start point, end point, and inside of the action; The combination of the position of the threshold value and then get the action candidate box. The method is simple and easy to implement, has strong promotion ability, produces high-quality candidate frames, and can effectively overcome the influence of unfavorable factors such as video background clutter and target motion speed changing range.

Description

Technical field[0001]The present invention belongs to the technical field of computer vision, and more specifically, relates to a method and system for generating cascaded action candidate frames based on action boundary probability.Background technique[0002]Human motion detection has many applications in people's daily lives, such as smart security, human-computer interaction, sports analysis, autonomous driving, and smart home. Human motion detection studies how to find out the time period when an action actually occurs from a whole video, and the motion candidate frame generation method described in this article is part of the research content of human motion detection.[0003]The current method of generating action candidate frames is usually to first perform dense sliding windows in time sequence, then perform context feature fusion on each sliding window, and finally perform candidate segment prediction. This solution is more sensitive to local feature changes, but it is difficu...

AI Technical Summary

Problems solved by technology

This scheme is more sensitive to local feature changes, but it is difficult to capture feature changes in a longer time range, and it is easy to get multiple redundant candidate segments in the local area, while missing some candidate action segments with a large time span.

Images