多场景下基于传感器的行为识别
Sensor-based behavior recognition in multiple scenarios
安健 1程宇森 2桂小林 1戴慧珺1
作者信息
- 1. 西安交通大学计算机科学与技术学院,陕西西安 710049;西安交通大学陕西省计算机网络重点实验室,陕西西安 710049
- 2. 西安交通大学计算机科学与技术学院,陕西西安 710049
- 折叠
摘要
针对基于传感器的行为识别任务中识别场景单一且固定的问题,提出一种多场景下基于传感器的行为识别迁移模型,由基于传感器的动态感知算法(dynamic perception algorithm,DPA)和自适应场景的行为识别迁移方法(adaptive scene human recognition,AHR)两部分组成,解决在固定场景下对传感器的依赖性以及在场景转换时识别模型失效的问题.DPA提出两阶段迁移模式,将行为识别阶段和模型迁移阶段同步推进,保证模型在传感器异动发生后仍能持续拥有识别能力.进一步提出AHR场景迁移方法,实现模型在多场景下的行为识别能力.实验验证该模型具有更优的适应性和可扩展性.
Abstract
Aiming at the problem of identifying a single and fixed scene in the sensor-based behavior recognition task,a sensor-based behavior recognition migration model was proposed.The model consisted of dynamic perception algorithm(DPA)for sen-sors and adaptive scene human recognition(AHR)to solve the problems of dependence on sensors in fixed scenes and recognition of model failure during scene transformation,respectively.Among them,a two-stage transfer mode was proposed in DPA,which promoted the behavior recognition stage and the model migration stage simultaneously to ensure that the model continued to have recognition capabilities after the sensor change occurred.AHR scene migration method was further proposed to achieve the behavior recognition ability of the model in multiple scenarios.The experiment verifies that the model has better adaptability and scalability.
关键词
传感器/行为识别/迁移学习/动态感知算法/自适应场景/两阶段迁移模式/场景转换Key words
sensors/behavior recognition/transfer learning/dynamic perception algorithm/adaptive scene/two-stage transfer mode/scene transformation引用本文复制引用
基金项目
国家重点研发计划基金项目(2018YFB1800304)
河南省重大公益基金项目(201300210400)
陕西省重点研发计划基金项目(2020GY-033)
中央高校基本科研业务费基金项目(xzy012020112)
出版年
2024
NETL
NSTL
万方数据