下載/瀏覽Download

Download Report

Transcript 下載/瀏覽Download

A Novel One-Pass Neural Network
Approach for Activities Recognition
in Intelligent Environments
出處: Hui Li, Qingfan Zhang, Peiyong Duan,
Sch. of Control Sci. & Eng., Shandong Univ.,Jinan,
Intelligent Control and Automation, 2008.
WCICA 2008. 7th World Congress on
報告人:向峻霈
Outline
• Introduction
• Human behavior recognition systems
• The novel one-pass neural networks architecture
and algorithm
• Experiment and result
• Conclusion
Introduction
• Weiser envisions a world of intelligent
environments, which can observe, interact with
and react to humans in meaningful ways
• Intelligent environments even take decisions
and execute actions on their own
Introduction
• 一般的系統狀態可利用兩種方式可識別不同的高層次
活動
▫ information
▫ Sensors
• 可以利用下列方式獲取數據
▫ 攝影機、麥克風和視覺追蹤模組
▫ Sensors裝置
Human behavior recognition systems
• there are two main approaches employed in
human behavior recognition systems which
▫ 數據統計
▫ 類神經網絡
The novel one-pass neural networks
architecture and algorithm
• The activities recognition system uses unobtrusive
and simple sensors
• S = (s1,s2,..si,..sn-1,sn)
• O = (o1,o2,..oj,..op-1,op)
The novel one-pass neural networks
architecture and algorithm
• A.Collecting and Classifying Learning Data
The novel one-pass neural networks
architecture and algorithm
• B.Neural Networks Learning Algorithm
▫ only the weight ratios of input-hidden layer need
to learn
▫ The leaning of weight ratios adopts One-Pass
learning technology.
The novel one-pass neural networks
architecture and algorithm
• B1. Local Weight Ratios Learning Algorithm
kj is the numbers of input-output pairs to one activity
The novel one-pass neural networks
architecture and algorithm
• B2. Global Weight Ratios Learning Algorithm
The novel one-pass neural networks
architecture and algorithm
• B3. Neural Networks Weight Ratios Algorithm
▫ According to local weight ratios and global weight
ratios the neural networks weight ratios can be
calculated using the following equation:
The novel one-pass neural networks
architecture and algorithm
• C. Neural Networks Output Calculation
ABNORMAL BEHAVIOR RECOGNITION
APPROACH
Experiment and result
• Simulating a user bedroom, the system adopts
10 low-level sensors to detect the users activities
• 200 were training data and 100 were test data
Experiment and result
96%.
95%.
training results
92%
78%
test results
Conclusion
• This paper conducted an abnormal behavior
recognition approach
• Our plan for the future work is to do verification
in the lab conditions