基于嵌入式平台的车载手势交互系统研究
摘 要：目的 解决基于YOLOv5手势识别模型在嵌入式平台运行速度较慢的问题。方法 改进了YOLOv5目标检测模型的网络结构，使用轻量型网络ShuffleNet替换YOLOv5的主干网络，将其部署到嵌入式平台，利用tensorRT对模型进行加速，并通过手势控制界面的准确率来判断改进YOLOv5模型的性能。结论 经过加速后的改进的YOLOv5模型在实时识别中可以实现对界面的控制，界面对静态手势的响应识别准确率为95.13%，对动态手势的响应识别准确率达86.25%，平均准确率为90.69%，推断时间相对原始YOLOv5模型降低了9.6%，训练后的权重大小只有原始YOLOv5权重的1/4。
关键词：手势识别；YOLOv5；嵌入式平台；ShuffleNet
中图分类号：TP391.41 文献标识码：A
Research on Vehicle-Gesture Interaction System based on Embedded Platform
Chenwangang1,2, Yanbo1, Dengliang2, Xujiexin1,2, Wangjinlei2
(1. Changzhou Xingyu Automotive Lighting Systems Co. Ltd, Changzhou 213022, China; 2. College of Mechanical and Electrical Engineering, Hohai University, Changzhou 213022, China)
Abstract: It aims to solve the problem that YOLOv5 gesture recognition model runs slowly on the embedded platform. It improves the network structure of the YOLOv5 model,replace the backbone network of YOLOv5 with a lightweight network ShuffleNet, deploy to the embeded platform, use tensorRT to accelerate the model, and judge the performance of the improved YOLOv5 model gesture control of the interface. The accelerated and improved YOLOv5 model can realize real-time recognition and control of the interface. The response accuracy of the interface to static gestures was was 95.13%, the response accuracy of dynamic gestures was 86.25%, and the average accuracy was 90.69. %, the reasoning time is reduced by 9.6% compared with the original YOLOv5 model, and the weight after training is only 1/4 of the original YOLOv5 weight.
Key words: gesture recogniton; YOLOv5; embedded platform; ShuffleNet
深度学习[1-2]的发展让目标检测的应用更加广泛。目标检测方法分为one-stage和two-stage。one-stage目标检测算法，其特点是一步到位，速度相对较快，典型的有YOLO，SSD。Two-stage目标检测算法的特点是识别错误率低，漏识别率也较低，但速度较慢，代表的是Faster RCNN。目前较广泛应用的是Redmon[3]等提出的多目标检测模型YOLO(You Only Look Once) 。经过不断的改进，YOLO在实时性和准确性上都有很好的表现。Leng Jianwei[4]在人机交互过程中利用YOLOv2模型定位手势，能够实现较好的识别速度和准确性。张锦[5]等人在原YOLOv5主干网络中加入MCA注意力模块，mAP超过了YOLOv3+SPP和原始YOLOv5，在安全帽佩戴检测任务中表现优秀。但是识别精度的提升往往带来网络深度和复杂度的增加，这限制了深度网络在嵌入端的表现。
