Research And Implementation Of Dynamic Gesture Recognition Based On Embedded Platform

Dynamic gesture algorithm is currently an important sub-direction of human-computer interaction research.With the popularization of the concept of intelligent life,the number of intelligent terminal devices has soared,and the demand for human-computer interaction using dynamic gestures is also increasing.This paper mainly focuses on the research on dynamic gesture recognition based on embedded platform.The goal is to implement a dynamic gesture recognition system that can recognize specific gestures on the neural network processor Hi3516DV300 launched by Huawei.The main research work is as follows:(1)Design a hand keypoint detection algorithm suitable for deployment on the embedded platform.The backbone network of this module is a lightweight DLANet,and the output head of this module is designed as a double-head output: The Gaussian heatmap output head and the Gaussian heatmap quantization error head make the groundtruth of the training data for the network closer to the true value and improve the accuracy of the network model.Combined with the post-processing method of gradient offset,the keypoint detection results are more effective.(2)Combined with long short-term memory network and attention mechanism,a dynamic gesture recognition algorithm based on skeleton suitable for deployment in embedded platform is designed.First,decouple the gesture keypoint information of the input sequence into two parts: hand shape change information and hand motion trajectory information,and then use the designed network structure to extract features from the two parts of information,and finally obtain the gesture recognition result of the input sequence.(3)Quantify the gesture keypoint detection algorithm and dynamic gesture recognition algorithm model trained on the PC and deploy it on the embedded NPU chip Hi3516DV300 provided by Huawei to realize dynamic gesture recognition system on embedded devices.The gesture keypoint detection algorithm is trained and tested on data collected by Huawei.Experiments show that the algorithm can achieve better keypoint detection effects on both the PC and the embedded platform for data without self-occlusion or self-occlusion.The dynamic gesture recognition algorithm is trained and tested using the DHG dataset.Experiments show that the algorithm can achieve a high classification accuracy on both the PC and the embedded platform,and achieve a better dynamic gesture classification effect.Finally,a complete and effective dynamic gesture recognition system is implemented on the embedded NPU chip Hi3516DV300.