Action detection model based on convolutional neural network

Abstract

The invention discloses an action detection model based on a convolutional neural network, and belongs to the field of computer visual studies. An efficient action detection model is constructed by using the convolutional neural network in deep learning, thereby recognizing an action from video and detecting and positioning the action. The action detection model is composed of a Faster RCNN (Regional Convolutional Neural Network), an SVM (Support Vector Machine) classifier and an action pipeline. Each part of the action detection model respectively completes the corresponding operation. The Faster RCNN acquires a plurality of regions of interest from each frame of picture, and extracts a feature from each region of interest. The detection model extracts the features by adopting a double-channel model, namely a Faster RCNN channel based on a frame picture and a Faster RCNN channel based on an optical flow picture, which are respectively used for extracting an appearance feature and an action feature. Then, the two features are fused into a time-space domain feature, the time-space domain feature is input to the SVM classifier, and an action type prediction value of the corresponding region is given by SVM classification. Finally, a final action detection result is given by the action pipeline from the perspective of video.

Description

technical field [0001] The invention belongs to the field of computer vision research, and constructs an efficient action detection model by using the method of convolutional neural network in deep learning, realizes action recognition from video and further detects and locates the action. Background technique [0002] Video recognition in the field of computer vision is divided into action classification and action detection. The problem to be solved by action classification is similar to "Is there an action or behavior like 'running' in this video?" Which frame sequence set the action appears in and where the action is in each frame". [0003] In recent years, benefiting from the great progress made in image recognition, video recognition has also made great progress. Most of the action recognition methods are proposed for action classification tasks. In fact, these methods can also be used for action detection tasks. J.Aggarwal, M.Ryoo, R.Poppe, etc. have summarized an...

AI Technical Summary

Problems solved by technology

In the past, many motion detection methods used the sliding window method (sliding window). However, motion positioning based on the sliding window method is extremely time-consuming and time-efficient. In order to speed up motion positioning, Oneatra et al. proposed a Fisher Vector strategy that is close to normalization. This strategy uses a more efficient method than sliding windows, the branch-and-bound search algorithm

Images