摘要： 无接触手势识别技术作为一种自然的人机交互方式，可以应用于手机、平板和可穿戴设备。为了高效实现超声手势识别嵌入式系统中采用的“一对其余”多分类随机森林算法，提出一种其推理过程的嵌入式软件实现方案。设计更精简的模型节点数据结构，以降低手势模型占用的存储空间。为节省系统能耗并缩短运行时间，利用分支定界的方法及时排除不可能产生正确解的手势类型，在保证识别率的条件下避免不必要的FLASH读取和决策树判定过程。实验结果表明，与传统的随机森林算法相比，该方案在FPGA上运行的实测时间缩短约60%，一次推理的平均判定次数低至243。

关键词: 随机森林, 手势识别, 嵌入式软件, 分支定界, FLASH存储

Abstract: Contactless gesture recognition enables natural human-machine interactions,and is applicable to mobile phones,tablets,and wearable devices.For embedded ultrasonic gesture recognition systems,the One-vs.-Rest(OvR) multi-class random forest algorithm is frequently used,and this paper proposes an efficient embedded software implementation scheme for the reasoning process of the algorithm.In this scheme,a more compact model node data structure is designed to save the storage space occupied by the gesture model.To reduce the energy consumption and the uptime of the system,the branch and bound method is used to eliminate in time the types of gestures that cannot draw the right conclusion.With the recognition rate assured,redundant FLASH reading and decision tree judgments are avoided.The experimental results show that compared with the traditional random forest algorithm,the proposed scheme reduces the actual uptime on the FPGA by 60% with the average number of decisions in an inference decreased to 243.

Key words: random forest, gesture recognition, embedded software, branch and bound, FALSH storage

中图分类号: 

• TP391