Research And Implementation Of Kinect-based Human Gesture Recognition Algorithm

Human body gesture recognition technology has been widely used in many life fields.However,in actual research,the final recognition accuracy is not high due to problems such as environment,background,and illumination.The depth image and skeleton information obtained by Kinect are color-independent and insensitive to light.Therefore,this thesis uses Kinect2.0 to construct the human body identity and posture image database required for the experiment,and collects three data types of color,depth and skeleton information.A human body identity and posture modeling method based on the coordinates of key nodes was proposed,and the distance method was used to extract the features.The model required less data and had high robustness.On this basis,the BP neural network identity and attitude recognition simulation model is built.The experimental results show that the human identity recognition rate is 99.17% and the average attitude recognition rate is 98.00%,which proves that the proposed algorithm has the dual advantages of less training samples and high recognition accuracy.On the other hand,training a large-scale network requires a large amount of resources,and the required tagged data sets are difficult to obtain.In view of this situation,this thesis builds an improved Res Net network migration learning model.Multi-scale feature fusion is introduced into the original network to improve the ability of feature expression,Adma algorithm and global average pooling method are used to optimize the model,and then combined with transfer learning technology to improve the performance of the network.Finally,the average recognition rate of human posture is 98.13%.Through comparative analysis,this method is better than the traditional algorithm in accuracy,and still has a high recognition rate for small data sets.