| With the rapid development of artificial intelligence,With its simple,easy to understand and non-contact characteristics,gesture has rapidly become an upsurge of exploration in the field of human-computer interaction.Compared with the early data gloves,the vision based non-contact gesture recognition interaction method has obvious advantages.Especially in the human-computer interaction system,the information image collected in real time by the visual sensor can quickly detect the position area of the human gesture,and then automatically control the corresponding functional modules through the gesture information,which has the characteristics of high accuracy and fast speed.However,due to the variability of gesture itself,the complexity of background and different illumination,the accuracy of gesture recognition has been impacted.With the rapid development of deep learning like a rocket,gesture recognition has also achieved amazing results in accuracy.In terms of human-computer interaction,this paper has done the following work on the basis of existing technology by investigating a large number of materials.Aiming at the problem of detection accuracy of gesture recognition,This paper studies and compares the target detection model fast r-cnn and the target detection model yolov4 tiny.Two convolutional neural network detection algorithms are used to detect gestures respectively,Experimental verification and comparative analysis are carried out on self-made standard data set and complex background data set.Considering the recognition results,confidence and target box positioning of the algorithm,yolov4 tiny performs better than fast r-cnn.For the classification of gesture recognition,This paper studies and analyzes the lightweight convolutional neural network algorithm Mini xception.Recognize gestures on self-made standard data sets and complex background data sets respectively,The experimental results show that,mini xception network model on the standard data set,the average recognition accuracy of 0-9 gestures is 96.43%,The average recognition time is 38.4 milliseconds,On the complex background data set,the average recognition accuracy of 0-9 gestures is 92.62%,The average recognition time is 43.8milliseconds.The recognition accuracy and average recognition time of mini xception network model on standard data set meet the needs of practical application.Aiming at the problem of gesture recognition accuracy in complex background,This paper optimizes the mini xception network model of lightweight convolutional neural network algorithm.Introduce the transfer learner and integrate yolov4 tiny into the mini xception model,generate a new network model YT＿mini-Xception。After experimental verification,YT＿Mini xception network model,the average accuracy of 0-9 gesture recognition on complex background data set is 96.64%,The average recognition time was 39.8 milliseconds,and the expected goal was achieved.Based on the improved YT＿mini-xception model,an intelligent physiotherapy system based on gesture recognition optimization algorithm is designed and implemented in this paper.The system transmits the real-time recognized human gesture information to the STM32 microcontroller of the lower computer through the upper computer,and automatically controls the hyperthermia parameters by controlling the output signal to realize the human-computer interactive physiotherapy platform of intelligent physiotherapy integration. |
