| It is a crucial manner of natural Human-Computer Interaction to perform a noncontact operation on devices by recognizing gesture movements.With the advantages of abundant target information,good privacy protection,and light robustness,gesture recognition based on the high-resolution radar is currently a hot research direction in the field of gesture recognition,with potential application demand in smart homes,autonomous driving,and other fields.As an alternative to traditional Human-Computer Interaction,the existing high resolution radar-based gesture recognition technology focuses on addressing the issue of gesture recognition accuracy,but it is not deep enough to research the high robustness and low latency of recognition algorithms.To solve the above challenges,this dissertation condenses the main scientific problems that need to be solved by analyzing the physical characteristics and high-resolution radar signal signatures of hand gestures,and carries out in-depth research on high-resolution radar gesture recognition approaches.The main research contents and contributions of the dissertation include the following aspects:1.To address the issue of large intra-class sample variance caused by the divergences among different users’ gestures,low personality-sensitive gesture recognition method is proposed.Based on the gesture signal model,the proposed method applies data enhancement to highlight the motion trajectory information in a frame of a gesture sample,and establishes Focus on Generalization loss function according to the visual experimental results of the data sources’ influence on the recognition performance.The proposed method improves gesture recognition accuracy in the cross-source evaluation mode.2.Aiming at the issues of multi-signal coupling and sample diversity in gesture radar signals,adaptive generalized feature learning method is proposed.By analyzing the phenomenon of multi-signal coupling in different kinds of gestures,a two-pathway feature extractor is proposed to extract the generalized features of coarse-grained gestures and fine-grained gestures.Meanwhile,adaptive individual cost loss function is established based on the quality of gesture samples from different data sources.The proposed method further improves the generalized performance of cross-source gesture recognition.3.To address the problem of the recognition performance degradation caused by the limited amount of information in frame-grained input samples,a gesture feature set extraction approach based on higher resolution radar is proposed.By analyzing the gesture signal model based on higher resolution radar,the proposed method reconstructs seven feature operators that describe the fine signatures of gestures from three aspects,respectively in motion trajectory,behavior,and direction.Combining the advantages of radar’s high Pulse Repetition Frequency,the proposed method establishes a sequential feature set based on a gesture frame.The proposed method achieves high-precision gesture framelevel recognition.4.Aiming at the challenge that frame-level gesture recognition is difficult to satisfy both high-precision and low-latency,negative latency gesture recognition method is proposed.The method constructs a pre-classification criterion of gestures by analyzing the motion law of common gestures,and also establishes a two-stage cascaded gesture recognition framework according to the characteristics of each gesture subset obtained by pre-classification.The proposed method reduces the dependence of recognition accuracy on the sequential information in input samples,and achieves high-precision and low-latency gesture recognition. |
