| Like face recognition and voice recognition,gesture recognition is also a common humancomputer interaction method,which has been widely used in daily life and has high application value.The existing gesture recognition methods are through video surveillance,wearing devices or wireless devices,among which the first two methods can achieve good results in recognition accuracy and efficiency,but limited by the environment or privacy and other restrictions.The gesture recognition method based on Channel State Information(CSI)in Wi-Fi is not subject to these limitations and is commonly used in various situations,which is beneficial to the research of gesture recognition.However,in the cross-scene gesture recognition problem,each new scene requires a large number of data samples to build a classifier.But the signal transmission of wireless devices such as Wi-Fi is highly susceptible to the influence of external layout,etc.The layout of different scenes or the location of different homes can have a great impact on the signal data recording,so that more samples are needed for the recognition of gestures in the new scene.Therefore,this paper uses Wi-Fi CSI data to study cross-scene gesture recognition under a few number of samples.The main work is as follows:(1)A Wi-Fi CSI gesture segment segmentation method based on dual-stage positioning is proposed to address the problem of false segments caused by long sequences of continuously recorded CSI signal data due to gesture continuity and external factors such as device recording.First,the method calculates the probability of a gesture by regressing the data in each time unit,and then after data smoothing,it effectively removes the outliers and separate the different gesture segments that seem to be continuous.Second,the method further determines the position of each data segment by combining random candidate frames with non-maximal suppression and the idea of exact regression.After experimental validation,the method can achieve 98% segmentation accuracy on the existing gesture data.(2)A multi-scene common gesture feature extraction algorithm based on information fusion is proposed to address the difficulty of extracting common gesture features between different scenes in the existing cross-scene recognition problem.Firstly,the method corrects the data offset by collecting scene information.Secondly,the method greatly expands the gesture information weight by calculating the multidimensional information of the used data.Finally,the common gesture features in different scenes are further obtained by changing the weights of different feature components in the process of feature extraction.After experimental verification,the method can effectively learn multi-scene common gesture features and can improve the accuracy of cross-scene gesture recognition methods.(3)A cross-scene gesture recognition algorithm based on a embedded two-branch network is proposed for the existing cross-scene recognition problem in which more samples are required in the target scene for model training and recognition.The method first pairs the gestures of the source and target scenes,then extracts the temporal and spatial features of the data through the internal double-branching structure,and then calculates the similarity degree of the data sample pairing through the external double-branching structure,which greatly improves the recognition accuracy at a very low cost of target scene data collection.After experimental validation,the method can achieve 96% recognition accuracy in the cross-scene recognition problem.From the above work,it can be seen that the method proposed in this paper has high accuracy and robustness in cross-scene recognition tasks of gestures by means of Wi-Fi CSI data,and can be effectively applied to human-computer interaction and smart homes in existing life. |
