面向嵌入式平台的轻量化神经网络手势识别方法

针对传统基于图像分割和特征提取的手势识别算法在复杂背景下识别准确率低、灵活性差的问题，基于目标检测神经网络的手势识别算法可以有效提高复杂环境下手势识别的准确性。受嵌入式处理器体积和功耗的限制，常用的目标检测神经网络在嵌入式上的识别速度较低，不能满足实时手势识别的要求。在SSD目标检测的基础上对其进行优化，使用MobileNetv3网络实现特征提取，目标检测方面则是使用SSD-lite结构，其使用深度可分离卷积替代普通卷积，实现了轻量化MobileNetv3-SSDLite手势识别算法的设计。针对手势识别的要求，制作了包含不同手势的数据集，利用它在服务器上完成了模型的训练。为了满足嵌入式的算力限制，通过模型的量化压缩将float64的网络参数量化为int8，并压缩网络结构，提高网络在嵌入式上的推理速度，实现基于嵌入式的手势识别。实验结果表明，基于嵌入式的MobileNetv3-SSDLite手势识别算法可以达到平均准确率99.61%，且识别速度达到每秒50帧以上，满足实时手势识别的要求。

Lightweight neural network hand gesture recognition method for embedded platforms

Compared with the traditional gesture recognition algorithms based on image segmentation and feature extraction in complex backgrounds which have low ?recognition ?accuracy ?and ?poor flexibility, the gesture recognition algorithm based on target detection neural network can effectively improve the accuracy of gesture recognition in complex environments. Restricted by the size and power consumption of embedded processors, the recognition speed of commonly used target detection neural networks on embedded processors is low and cannot meet the requirements of real-time gesture recognition. In this paper, we optimize the SSD target detection and use MobileNetv3 network to achieve feature extraction and SSD-lite structure for target detection, thus to use depth-separable convolution instead of ordinary convolution to realize the design of lightweight MobileNetv3-SSDLite gesture recognition algorithm. For the requirements of gesture recognition, we make a dataset containing different gestures and complete the training of the model on the server using the dataset. In order to meet the arithmetic limitation of embedded processor, we quantize the float64 network parameters into int8 by quantization compression of the model, and compress the network structure to improve the inference speed of the network on embedded processor to realize the embedded-based gesture recognition. The experimental results show that the embedded-based MobileNetv3-SSDLite gesture recognition algorithm can achieve an average accuracy of 99.61% and a recognition speed of above 50 frame/s, which meets the requirements of real-time gesture recognition.