Research On Human Detection And Motion Recognition In Indoor Environment

The goal of human body recognition and human motion recognition is to find the human body in the image and to further identify and judge the actions of the human body.Human body detection and human motion recognition have a large range of application,such as human-computer interaction,security monitoring,assisted driving and autonomous driving.As a kind of statistical learning method,deep learning improves the accuracy of classification quite a lot compared with the traditional classifier,but the traditional statistical learning method has a lower requirement for computation and device costs.Therefore,deep learning and traditional statistical methods,both of which are parts of machine learning,have their own advantages respectively.In this paper,human body recognition and motion recognition are studied by both traditional methods and deep learning.The main research work is as follows:A static human target detection method is studied for the task of static human body detection in indoor environment,which can adapt to complex environment.This method combines image segmentation with statistical learning.Firstly,the entire image is fed into Mean-shift segmentation,and then the feature,histogram of gradient(HOG),is extracted from the segmented tile,followed by the first support vector machine(SVM)with a loose threshold to do classification.In this case,a looser threshold can ensure that the human body parts will not be missed and that human body is fully detected in the final phase.The pre-classification of the blocks and the prior knowledge of the human body structure is further combined together respectively.Extract the HOG features again from the merged tiles and re-send them into the SVM2 with strict threshold.The human body detection using the coarse-precision two-level SVM can quickly localize the possible human body area,effectively reducing interference and improving the detection success rate.For the problem of human motion recognition,a method is designed based on the combination of HOG feature from image sequence and feature from 3D convolutional neural network features.Firstly,every 16-frame image sequence is together sent to 3D convolutional neural networks to obtain the first feature.Secondly,the HOG feature is extracted from each frame of this 16-frame sequence,and the feature subtraction value between two adjacent frames is considered to be the HOG feature of the sequence.Finally,XGboost,as a commonly used statistical learning method,can achieve better classification results than a single classifier by learning multiple classifiers.In detail,we input these two features into XGBoost to get the final recognition result.Among the two features,HOG feature is a global feature,which is sensitive to human contour changes.Extracting the HOG feature from the sequence picture and subtracting the adjacent frame feature values can effectively reduce the influence of the background of the picture sequence.Plus,3D convolutional neural network per se considers both the spatial and temporal dimensions,so combining two different features can perform better in the recognition of human motion through XGBoost.