Multistage video action detecting method

Abstract

The invention discloses a multistage video action detecting method. The multistage video action detecting method comprises the steps that rough action fragments fusing multiscale sampling and single-scale training are generated by an input video which is not clipped through a dichotomy and voting fusing strategy based on a deep residual network; the rough action fragments are subjected to action classification and action boundary joint distinguishing through a statistical fusion strategy based on a frame-level action recognition result, and primary action detecting fragments are obtained; in combination of IoU between the primary action detecting fragments, an improved non-maximum suppression algorithm is utilized for processing the primary action detecting fragments, and finally the action detecting result, namely, the action classification and beginning and ending time position of each video action detecting fragment, of the video which is not clipped is obtained. According to the method, the action classification accuracy and action locating precision can be improved.

Description

technical field [0001] The invention relates to the technical field of video motion detection, in particular to a multi-stage video motion detection method. Background technique [0002] With the rapid development of network and multimedia technology, video has become an important carrier for people to obtain information, and the number of videos is growing explosively, so the analysis and understanding technology of video content is very important. Uncropped videos usually contain multiple action instance segments and a large number of irrelevant background segments, where the action occurs, the time interval, and the action category label are unknown. The video action detection task is to be able to identify the category labels of multiple action instances in the uncut video and locate the start time and end time of the action instance. As one of the important research directions of current video processing technology, this task is widely used in intelligent Surveillance,...

AI Technical Summary

Problems solved by technology

However, since the candidate classification network requires the same time length of the video clips input by the network, this method obtains video clips of different time lengths by controlling the frequency of downsampling. However, using the same network structure to train video clips obtained with different sampling frequencies will lead to The difference within the class increases, and the continuity within the action is destroyed, so that the 3D CNN cannot learn good motion features and increases the difficulty of network training.

Moreover, this method only classifies the action categories of the video clips, and does not realize the fine-tuning of the boundaries of the video clips, so that the improvement of the accuracy of action positioning is limited.

Images