Video human body behavior recognition method based on motion saliency

Abstract

The present invention discloses a video human body behavior recognition method based on motion saliency. The method comprises: performing motion saliency detection on a behavior video frame by using a motion saliency detection algorithm, so as to obtain a motion saliency image; by adopting non-maximal suppression (NMS) sampling algorithm, calculating a motion saliency area candidate frame based on the motion saliency image; clipping the video frame around the motion saliency area candidate frame to acquire an image block containing a human body behavior; scaling the image block obtained through clipping into a size required for input data of a depth convolutional neural network; by using the depth convolutional neural network, extracting a depth feature of the human body behavior based on the scaled image block; and performing feature classification on the depth feature of the human body behavior to obtain a human body recognition category result. According to the method, the image block required by the convolutional network is constructed around the behavior saliency motion area, so that the areas of human body behavior changes are effectively captured, the depth convolutional features of human body behaviors with good recognition are effectively extracted, and the accuracy of human body behavior recognition is effectively improved.

Description

Technical field [0001] The invention belongs to the technical field of automatic video analysis, and relates to a video human body behavior recognition method based on motion saliency. Background technique [0002] Video human behavior recognition can meet the needs of automatic analysis and intelligence for tasks such as video surveillance, intelligent monitoring, and video content analysis, and promote social development and progress. Human behavior feature extraction plays a key role in the process of behavior recognition, and the quality of behavior features directly affects the final recognition effect. At present, the high-efficiency feature expression capability of the deep model makes it replace the traditional human behavior local feature extraction model, and has become a research hotspot of video human behavior recognition. [0003] Behavioral video sources are different, and the video frame resolution size varies. On the other hand, the deep convolutional neural netwo...

AI Technical Summary

Problems solved by technology

The image zoom sampling method will cause the visual deformation of human behavior. The latter two sampling methods will introduce background image blocks that do not contain behavior when the behavior area of ​​the video frame is small or biased. These factors seriously affect the accuracy of behavior recognition. Rate

Images