Recognition Of Double Interaction Based On Joint Point Data

Two-person interactive behavior recognition is an important research topic in the field of computer vision,and it has broad application prospects in the fields of intelligent video surveillance and human-computer interaction.Most existing two-person interactive recognition is based on RGB video to perform human detection and tracking and behavior analysis tasks.Although RGB video data can provide rich color and texture information,the data is highly susceptible to changes in lighting,morphology,posture,and occlusion.These problems will seriously restrict the robustness and operation efficiency of detection and tracking algorithms.Compared with RGB video information,the joint point data information has the advantages of clear and simple characteristics and less influenced by appearance factors.Microsoft Kinect somatosensory camera can be combined with bone tracking technology to obtain real-time human joint point data,and is almost not affected by the background and lighting.This thesis conducts an in-depth study on the recognition of two-person interaction behavior based on joint data:Firstly,for the problem of two-person interaction recognition method based on RGB video,which is easily affected by changes in illumination,body occlusion and environmental changes,a two-person interaction recognition algorithm based on joint point data is proposed.The Histograms of 3D Jointsfeature representation method of three-dimensional human posture is used and extendedto the expression of two-person interaction behavior.This method uses an improved spherical coordinate system to divide the three-dimensional space into n spaces,and select 12 joints information to construct a compact representation of the human body.The Gaussian function is used to vote the three-dimensional skeletal joints into the adjacent space.The joint point data is used to calculate the HOJ3 D features of each person performing the action,and then is aggregated into k gesture visual dictionaries torepresenting the key gestures of the action.Then,HMM is chosen to train the action model and recognize it.The algorithm makes full use of the data of the joint points,and has good robustness to small posture changes.Aiming at the traditional method of human interaction recognition,the recognition accuracy of complex interactions may not be greatly improved,and most recognition algorithms treat double interaction as a whole,and do not consider the interaction between individuals.a two-person interactive behavior recognition algorithm based on joint point data is proposed.This method calculates the HOJ3 D features for two single persons separately,and then calculates the HOJ3 D features for the two persons as a whole.At the same time,the joint distance features of the two persons are calculated to represent the interactive information between the two persons.The two features are imaged and sent to the convolutional neural network for fusion and recognition.This method avoids the interference of the original data very well,and the constructed features retain a lot of spatio-temporal information and interactive information of the interaction behavior.This method achieves a better recognition effect for the complex two-person interaction behavior.Finally,in view of the problem that the joint point data is imaged and the convolutional neural network is used for identification,the video timing relationship cannot be effectively modeled.An interactive behavior recognition method for skeleton image sequences is proposed.First,calculate the joint point distance feature of a single frame,and quantize it into a grayscale image every three frames.Then send each grayscale image to the convolutional neural network to extract the depth features.Finally the extracted depth features are sent to the long-term and short-term memory networks for time series modeling to complete the recognition of the interaction between two people.This method makes full use of the space and time information of the interaction between two people,and realizes the accurate recognition of the complex interaction between two people.