Human Recognition Based On Multi-Posture Model

The purpose of human recognition is identification,with the development and popularization of human recognition,identity has great applications in various fields.At present,the research on human recognition mainly focuses on the face region,and the recognition rate is very low in the absence of clear face.Therefore,this paper deeply analyzes the research status of the current human recognition algorithm,from multi-posture aspect,extracts the multi-posture samples and posture feature,and uses the multi-posture feature to carry on the human recognition.During the extraction of multi-posture samples,this paper proposes two different methods to extract multi-posture samples based on the poselets algorithm.First method is to combine the maximum weight bipartite graph matching algorithm(P-BG),and the second method is to use CNN features and natural neighbor manifold ranking algorithm(P-CMR).The difference between the two methods is that by screening the figure box detected by the filtering model of the poselets algorithm,and seting the filtering model according to the specific location of the human head calibration frame,P-BG algorithm matches the filter results to find the target person of the specific location by combining the sorting score with the maximum weight bipartite graph matching algorithm;P-CMR algorithm extracts the CNN feature from the image,and uses the natural neighbor manifold ranking algorithm to sort the image features,and find out that the image corresponding to the optimal feature is the target person in the specific location.Finally,the two algorithms extract the corresponding posture.Experiments show that the two methods in this paper can effectively detect the target person,and extract the corresponding posture,and the accuracy of the second method is slightly better than the first method when the target person is occluded by the background.In the feature extraction process,this paper adopts different feature extraction methods for different positions.Compared with the traditional feature extraction algorithm,the use of CNN(convolution neural network)algorithm to extract image features can better express the correlation between features.We use the Alexnet network architecture for model training of the non head region posture samples,and fine tune the CNN model trained on the ImageNet dataset;while the posture samples belonging to the head region is modeled by the latest VGGnet network architecture,due to insufficient sample size,so fine-tuning the pre-trained VGG-face model directly.Finally,the CNN model of training set one corresponding position is obtained,and the fc7 layer features of training set 2 and test set are extracted on this model for later classification training.In the process of feature classification training,the combination of multi-dimensional features easily leads to too large feature space and slow data processing.Therefore,we proposes a method of using the weights to combine the posture classification results for human recognition.First,a binary classification SVM is trained to obtain the weight value corresponding to each posture.Then,a corresponding multi-classification probability SVM(Support Vector Machine)model is trained on the features of each posture of the test set,and the corresponding probability value is obtained.Combining the probability value of each posture with its corresponding weight value,the total probability value is the final joint identity prediction value.Experiments show that the classification accuracy of the algorithm is better than other traditional methods in the PIPA database.