Research On Human Multi-posture Recognition Algorithm In Computer Vision

With the continuous development of computer vision technology and related hardware equipment,human posture recognition has gradually entered human life,and it has been widely used in video surveillance,motion-assisted analysis,home care,body games and other fields,with considerable economic and social value.Traditional artificial design feature methods require researchers to have rich experience,which has heavy workload and poor robustness.The emerging deep learning methods can automatically learn action features and recognize them,which can solve this problem well.However,due to the complex background,changeable movements,occlusion and illumination of the target,human posture recognition still has some limitations in practical application scenarios.The purpose of this thesis is to improve the accuracy and robustness of human posture recognition algorithm by constructing a new convolutional neural network model.The specific research is as follows:(1)Firstly,five kinds of moving target detection methods are analyzed and studied.Compared with other detection algorithms,ViBe algorithm has better detection clarity and real-time performance,but in the environment of complex background or strong dynamic target,there are also problems of target edge incompleteness,regional faults and internal holes.In order to further improve the integrity of target detection,an improved ViBe algorithm with self-adaptive threshold adjustment combined with self-defined evaluation function is proposed.The experimental results show that the improved algorithm can provide a clearer prospect for the input of convolution neural network model,which is conducive to the subsequent attitude classification and recognition.(2)Although the three-dimensional convolution neural network has been gradually applied in the field of human posture recognition,there are still some problems such as missing target motion information,imperfect feature extraction,and easy to misdetect similar actions.In order to extract more abundant and detailed motion features,this thesis uses two three-dimensional convolution operations continuously,and prevents model over-fitting which may be caused by multiple convolutions through BN algorithm and dropout technology.In order to improve the applicability of the algorithm,a spatial pyramid pooling layer is added before the full connection layer to enable the network process any resolution image.Finally,this thesis constructs a multi-convolution 3D CNN model which integrates BN algorithm,dropout technology and space pyramid pooling technology.(3)Recognition tests are performed on KTH,UCF101 and self-built video libraries.Different feature combinations are used as input of the 3D CNN model.The binary images with clear targets are obtained by the improved ViBe algorithm.The experimental results show that the feature combination of "ViBe binary graph + optical flow graph + three frame difference graph" as model input can achieve higher recognition accuracy,especially for data sets with complex background,multiple action types and small difference.Therefore,the 3D CNN model constructed can effectively improve the accuracy of human posture recognition,and has good application value.