Video action detection method based on central point trajectory prediction

Abstract

The invention discloses a video action detection method based on central point trajectory prediction. In video spatio-temporal motion detection, each action instance is regarded as a movement track ofthe central point of the action initiator; wherein the trajectory is represented by motion vectors of the central point of the intermediate frame action and the central points of other frame actionsrelative to the central point of the intermediate frame; firstly, features are extracted from a video frame sequence for obtaining central point position prediction and action category prediction of an intermediate frame, then obtaining a motion trail from the central point of the intermediate frame to central points of other frames, and finally generating a detection box of each frame central point to obtain position positioning, thereby obtaining an action category and a positioning result, and completing a space-time detection task on a video clip. According to the anchor-frame-free video action detection method provided by the invention, time-space positioning and classification tasks of actions are completed in the video along a time sequence, and compared with video action detectionwith an anchor frame in the prior art, the anchor-frame-free video action detection method is simpler and more efficient, embodies robustness and high efficiency, and has very strong expansibility andportability at the same time.

Description

technical field [0001] The invention belongs to the technical field of computer software, relates to a spatio-temporal motion detection technology, in particular to a video motion detection method based on center point track prediction. Background technique [0002] Spatio-temporal action detection is an important research task in the field of computer vision. Its purpose is to classify action instances in videos and locate them in space and time. Spatio-temporal action detection has broad application prospects in real-world scenarios, such as video surveillance and group action detection. The current commonly used method is to use the motion detector independently on each frame to complete frame-by-frame detection, and then use dynamic programming or target tracking to connect the detection results of single frames in time sequence. These methods cannot effectively utilize information in the temporal dimension when performing single-frame detection, so they do not perform ...

AI Technical Summary

Problems solved by technology

As the video time increases, the number of anchor boxes appearing in a video will increase sharply, which brings great challenges to the training and testing of neural networks

Images