面向嵌入式平台的手势轨迹识别
Gesture trajectory recognition for embedded platform
王绎茗 1高美凤1
作者信息
- 1. 江南大学物联网工程学院,江苏无锡214122
- 折叠
摘要
针对大部分嵌入式平台计算能力较弱、无法实时运行基于神经网络的应用程序的情况,提出了一种基于Anchor-Free轻量化卷积神经网络(CNN)的手势交互方法.该方法缓解了Anchor-Based检测的弊端,通过神经网络实时检测人体手部,并利用动态时间规整(DTW)算法对手势轨迹进行分类.神经网络参数量仅有0.22 M,在自建手部检测数据集上的平均精度均值(mAP)交并比(IoU)=0.50:0.95可以达到68％.在RK3568嵌入式平台上,每帧推理和后处理时间仅有31 ms,轨迹分类耗时仅有43 ms,CPU使用率仅有34％,满足实时性要求.
Abstract
Aiming at the fact that most embedded platforms have weak computing power and cannot run neural network-based application (App)in real-time,a gesture interaction method based on Anchor-Free lightweight convolutional neural network(CNN)is proposed.This method alleviates the drawbacks of Anchor-Based detection,detects human hands in real-time through neural network,and uses dynamic time warping(DTW)algorithm to classify gesture trajectories.The quantity of neural network parameters is only 0.22 M.On the self-built hand detection dataset,the mean average precision(mAP)intersection over union(IoU)=0.50:0.95 can reach 68％.On the RK3568 embedded platform,for inference and post-processing per frame is only 31 ms,time comsuming of trajectory classification is only 43 ms,and CPU usage rate is only 34％,which meets real-time requirements.
关键词
轻量化卷积神经网络/无锚框/目标检测/手势分类/嵌入式系统Key words
lightweight convolutional neural network/Anchor-Free/target detection/gesture classification/embedded system引用本文复制引用
出版年
2024
NETL
NSTL
万方数据