High-accuracy human body multi-position identification method based on convolutional neural network

Abstract

The invention provides a high-accuracy human body multi-position identification method based on a convolutional neural network. A deep convolutional neural network is adopted to extract image characteristics. The method fully utilizes the depth information of an image to greatly improve the identification accuracy rate of the image; secondly, on the basis of a RCNN (Region Convolutional Neural Network) method, a Selective Search algorithm is used to form a candidate border so as to better adapt to the accuracy of positioning information under the deep convolutional neural network than a sliding window method; furthermore, the last layer Softmax layer of the convolutional neural network is replaced with a SVM (Support Vector Machine) to finally obtain a classification-based score; in addition, after the SVM score, which is relative to each category, of each candidate boarder is obtained, a pixel-based position range constraint, a K neighbor constraint and a Gaussian mixture model are added to finally form a candidate border combination based on general comprehension, the accuracy of the human body multi-position identification is improved, and the high-accuracy human body multi-position identification method can carry out more accurate positioning than an original RCNN method.

Description

technical field [0001] The invention relates to the field of pattern recognition and deep learning, in particular to a high-accuracy multi-part human body recognition method based on a convolutional neural network. Background technique [0002] Human recognition is a hot topic in computer vision. In the past, human recognition was mostly based on low-level feature factors and high-level context. The SIFT operator and the HOG operator are commonly used operators. They are obtained based on the low-level orientation histogram, but the features of the image are layered and advanced layer by layer. Therefore, LeCun et al. Based on the research of Rumelhart et al., the training method of convolutional neural network (CNN) based on stochastic gradient descent algorithm based on backpropagation was proposed, which received considerable attention at that time and formed a new cognition in the field of computer vision. [0003] Although CNN was widely used in the 1990s, it was late...

AI Technical Summary

Problems solved by technology

How to establish an association model of the whole and each part that can maintain a high accuracy rate with changes in action posture and viewing angle is called the top priority. In the past methods, geometric constraint algorithms such as Poselets and DPM have achieved good results. However, most of these results are based on the given overall frame to locate the position of each part, so the above methods also need further improvement

Images