Human Action Recognition Based On DBN-HMM

Action recognition is human-centered,and it is the basis of human-computer interaction that accords with human habits.Human body motion contains a large amount of information.In human-machine cooperative assembly,the machine can better cooperate with people in assembly by improving the recognition of human action and improve assembly flexibility,human behavior recognition makes the machine able to discriminate and understand the intention of the human movement,so as to better carry out human-computer interaction and achieve man-machine cooperation assembly.Action recognition is gradually applied to all aspects of people's life and work,and it has far-reaching application value.In this paper,the three aspects of action feature extraction,classifier design and continuous motion segmentation and recognition are studied.Established human body motion description model.Based on the human body 3D information,the ten corresponding angles of the human body are taken as the static feature,and the limb angle change between adjacent frames is taken as the dynamic feature.The static feature is combined with the dynamic feature to describe the motion,and the limb angle model is implemented.For human action,the model has strong characterization capabilities.The appropriate frame is selected and the redundant frame is removed by the frame selection algorithm according to the cumulative distance of the angle distances between adjacent frames,thereby further improving the recognition effect and reducing the amount of calculation.An action recognition system based on deep belief network and hidden Markov model is constructed.In the structure of the deep belief network,the conditionally restricted Boltzmann machine is used instead of the traditional restricted Boltzmann machine,which can enhance the deep confidence network's ability to model time-varying data,and uses feedback-adjusted output layers for global fine-tuning.Using the deep belief network to estimate the observation probability in the hidden Markov model.The deep confidence network model based on the condition-bound Boltzmann machine can abstract high-level abstract features with historical information,and can be better applied to the recognition of time series actions compared with the traditional action recognition methods.Experiments were performed with UTKinect Action and MSR Action3D database's action data.The recognition results show that the algorithm in this paper has higher recognition results.On the basis of single action recognition,sliding window method and dynamic programming method are combined to segment and recognize continuous actions.The initial segmentation points are detected by scoring mechanism and sliding window method.On this basis,the dynamic programming algorithm is used to optimize the location of the segmentation points.Finally,the segmentation results are reasonably optimized and selected.In this paper,the segmentation of continuous action is estimated by the logarithmic likelihood value of the trained action model.It is better to judge the sequence by the prior knowledge obtained by training than the traditional unsupervised method.