Research And Application Of Dynamic Gesture Recognition Algorithm Based On Educational Robot

Educational robots have shown positive effects in stimulating users’ interest in learning,hands-on ability and innovation;as one of the main ways for educational robots to interact with people,dynamic gesture recognition has the characteristics of natural interaction,intuitive image and low learning cost.However,educational robots based on dynamic gesture recognition still have problems of low recognition accuracy,slow speed,and low matching between gestures and functions,which lead to poor overall interaction experience.Therefore,it is of great practical importance to study how to effectively improve the dynamic gesture recognition performance of educational robots and optimize the human-robot interaction logic.By analyzing the performance and price of common educational robots,the TonyPi robot is selected as the experimental platform,and then the MobileNet v3-Small model is improved,and T-S MobileNet is proposed as follows:(1)To meet the need of dynamic gesture data,the original model is transformed in 3D;(2)To reduce the inference time of the model,the structure of the model is modified;(3)T-S attention is proposed to make the model establish the attention matrices of spatial and channel dimensions respectively about the temporal dimension in the first stage to improve the model accuracy;(4)To further improve the model operation efficiency,the redundant layers are removed without affecting the model accuracy.Finally,the ERG dynamic gesture dataset is established according to the functional needs,and 21 functions are designed and written using multi-threading techniques to improve the matching between gestures and functions and optimize the overall interaction experience of the educational robot.To verify the effectiveness of the proposed methods,T-S MobileNet is trained and tested using the large-scale public dataset Jester and the self-built dataset ERG.Compared with the original model,the accuracy of the improved model is improved by1.14%(94.97%)and 0.97%(98.92%)for the Jester and ERG datasets,respectively,with more accurate gesture recognition.The inference speed of TonyPi educational robot with the improved model is improves by 12.40%(1.36 FPS)and 11.81%(1.42FPS)in the Jester and ERG datasets,respectively,and the robot runs more efficiently.Combined with all the improvement strategies,the TonyPi educational robot with dynamic gesture recognition has better accuracy,faster speed and 2.03%(94.29%)higher average correct rate,the overall interactive experience is excellent and can meet the needs of daily education.