| Hand gestures can be said to be another important communication way besides language,which contain rich semantic information and are widely used in many fields,such as intelligent driving,virtual reality,sign language recognition and so on.With the rapid development of electronic information technology,gesture recognition technology has attracted more and more attention.Gesture recognition technology can be divided into static gesture recognition and dynamic gesture recognition.Static gestures only contain the spatial characteristics of gestures,that is,they can only recognize the "state"of gestures at a certain time,but cannot track the transformation of the "state" of gestures.For example,when the hand is in the "fist" state,the gesture action can be correctly recognized by the static gesture recognition method,but for the semantic action of the palm sliding from left to right,the static gesture recognition method cannot correctly recognize it.Because dynamic gesture recognition can acquire the temporal information of gestures,i.e.grasp the continuous change of gestures,it can deal with more complex gesture semantic actions,which has broader application prospects.Dynamic gesture recognition can be divided into two main methods:recognition based on visual information and recognition based on auxiliary equipment.Moreover,the recognition method based on auxiliary equipment mainly uses multi-sensor wearing equipment to directly determine the spatial position of the key points of palm and finger,so as to obtain the changes of the hand's key points with time,and ultimately use these information to infer the semantic information expressed by gesture action.Most of these auxiliary devices connect gesture recognition system with user's hand through wired or wireless communication technology,and transmit user's gesture information to recognition system intact and unambiguously.Their typical devices,such as data gloves,are of poor practicability(comfort,security)and ease of use due to equipment dependence.In recent years,more and more attention has been paid to the visual-based gesture recognition technology which can recognize gesture movements without devices.This thesis focuses on dynamic gesture recognition and tracking algorithm based on visual information.The main contents are as follows:1)A dynamic gesture recognition method based on two-dimensional convolutional neural network is proposed.For conventional gesture recognition method,it is usually necessary to extract the key points of gesture accurately,and to design spatio-temporal features manually or to model temporal information,which seems a bit complex.Convolutional neural network has a strong feature extraction ability.Based on training datasets,it can learn the features that are beneficial to the target task easily without human intervention.Recognition algorithm can learn the spatial and temporal features of gesture autonomously with CNN,which enormously simplifies the gesture recognition process,and improves the recognition accuracy.2)A gesture tracking method based on multi-vision is proposed.Gesture tracking under monocular vision is prone to the situation that the target is temporarily lost and cannot be accurately tracked in the follow-up process.In multi-vision,the accuracy of tracking can be guaranteed to the greatest extent under the non-extreme situation(the target is lost for a long time under all cameras).Compared with single camera target tracking,multi-camera target tracking can obtain information of different visual views of the target.When the target occlusion occurs at a certain view with the result that a small part of the tracking target or even the target totally is missed by the corresponding camera,while other cameras can still capture the complete tracking target,multi-camera tracking can abandon the target image captured by the occluded camera and only use other cameras to obtain the target information,thus effectively solving occlusion problem of the single camera target tracking. |
PDF Full Text Request