Research On Pedestrian Identification Based On Kinect Bone Information

In recent years,the research on biometric identification has developed rapidly.Face recognition,fingerprint recognition and other methods have been widely used in various fields of daily life,which has promoted the steady progress of social and public security protection.However,these methods also have the defect of being easily destroyed or camouflaged.As an internal attribute of the human body,skeletal features can reflect the differences between individuals and have strong stability.Compared with other biological features,they can be collected from a long distance.Therefore,Bone feature recognition technology has become a research hotspot in recent years.The Kinect equipment is used to obtain highly reliable human skeleton and joint data,and typical static skeleton features is extracted that can reflect individual differences for identification in order to apply to more work scenarios.A pose recognition method based on depth image is proposed to solve the problem that the pose is greatly affected by the skeleton feature extraction.Two classification and recognition algorithms,BP neural network and SVM,are constructed,and the two models are improved based on particle swarm algorithm to complete the goals of person pose recognition and identity recognition.The main work of this thesis includes:(1)The depth image of the human body and the three-dimensional coordinates of the 32 bone joint points are obtained with the Kinect device,and the Kalman filter is used to eliminate the jitter error of the joint.After the preprocessing is completed,the shoulder width and arm width of the human body are extracted from the joint data.The12-dimensional skeleton features such as length and leg length form feature vectors,and the extracted skeleton features can clearly reflect the differences between individuals after experimental verification.(2)The factors affecting the extraction of skeleton features were analyzed and verified,and it was found that the posture of the human body had the greatest impact on the process of skeleton feature extraction.Therefore,the design was designed to identify and classify the posture before identification.In this thesis,the HOG feature of human pose is extracted from the depth image,and the principal component analysis method is introduced to reduce the dimension of the data.Finally,the pose feature of the100-dimensional feature vector is extracted for pose recognition.(3)BP neural network and SVM multiple classifiers are constructed to recognize attitude and identity.Because these two classification models are greatly affected by initial parameters,PSO-BP and PSO-SVM improved models are designed to improve the model with particle swarm optimization algorithm.The data samples in NTU rgb-d data set are selected to form the data set for attitude recognition and identity recognition verification.After experimental verification,it is found that the performance of the classifier model improved by particle swarm optimization algorithm is improved by more than 10 percentage points,and the recognition rate of attitude and identity is more than 88%.It also verifies the feasibility of the attitude recognition and identity recognition method proposed in this thesis.(4)The PSO-BP model is used as the gesture recognition model,and the PSO-SVM model is used as the identity recognition model to build an identity recognition system based on human gestures.Kinect equipment is used to collect relevant data containing 30 different human bodies and 8 postures to form a training set.And the test set,the training of the recognition model in the system was completed,and the effect of the classification system was tested.The final gesture recognition rate was92.8%,and the identity recognition accuracy rate was 91.7%,and good recognition results are obtained.