selecting home appliances with smart glass based on
TRANSCRIPT
Osaka UniversityAdvanced Telecommunications Research
Institute InternationalOsaka UniversityAdvanced Telecommunications Research
Institute International
*Quan Kong(Osaka University)Takuya Maekawa(Osaka University)
Taiki Miyanishi(ATR)Takayuki Suyama(ATR)
Selecting Home Appliances with Smart Glass based on Contextual Information Ubicomp 2016
Osaka UniversityAdvanced Telecommunications Research
Institute International
Introduction- Control home appliances
Free your hands
AndIn more direct way
Osaka UniversityAdvanced Telecommunications Research
Institute International
Approaches for Home Appliances Control
Voice Gesture
Wearable CameraIR
Wearable Camera
Shi F., 06
“Alex, open the curtain on the kitchen’s north wall.”
IR Emitter
Ben Z., SUI ’14
Amazon echo
Osaka UniversityAdvanced Telecommunications Research
Institute International
Our Approach
“ON”
+Living RoomTelevision
Home Network
・Activity・Indoor position
Osaka UniversityAdvanced Telecommunications Research
Institute International
Feature of Our Approach – Context-aware appliance selectionGoogle Glass
Camera ・Orientation・Accelerator・Light sensor…
Context Information
Indoor positions
Activities
non-parametric unsupervised learning
Why context information ?Related to the home appliances Distinguish between different appliances
Osaka UniversityAdvanced Telecommunications Research
Institute International
System Overview- Sensors Used in Our System
ScreenLight sensor
Orientation
Google Glass SmartphoneWi-Fi
Microphone
Accelerator
Camera
Appliance Image(First person view)
Camera
Activity Information
Microphone
Accelerator
Light sensor
Wi-Fi
Indoor position Information
Orientation
Head direction
Osaka UniversityAdvanced Telecommunications Research
Institute International
Update the model by using the collected history daily data
Use the trained appliance selection model in the daily life
Select the prepared name of appliance in glass application
Stand in front of the appliance and take a 10 seconds video
System Overview- Initialization and Model Update
1
2
Images extracted from the video used as the training data for initializing 3
4
5
wrong estimation is corrected by the user and used as training data
Initi
aliz
atio
n
Osaka UniversityAdvanced Telecommunications Research
Institute International
Proposed Method
Extract the attention time’s image feature and estimate the activity & position (IGMM)
Detect the user’s attention using orientation data 1
2Extracted above information as the input of appliance selection model (MKL) 3
Appliance Image Sensor Data
Position
Activity
+
Deep Convolutional
NeuralNetwork
Extract CNN fc6
CNN featurefc6:4096
dimensional
Feature Extraction
Train(ed)ActivityModel
Train(ed)Positional
ModelTrain Data(test data) Feature Data
0.981 Act10.765 Act20.543 Act3… …
MultipleKernel
Learning
Classification
Estimated Appliance
Context feature
Clustering
0.923 Pos10.865 Pos20.443 Pos3… …
Attention Detection
Osaka UniversityAdvanced Telecommunications Research
Institute International
Proposed Method – Image feature extraction with DCNN
image feature
CaffeNetReference Model: ILSVR-2012
Osaka UniversityAdvanced Telecommunications Research
Institute International
Proposed Method-Unsupervised activity recognition and indoor positioning
Wi-Fi
- light- Acc- microphone
PositionIGMM
ActivityIGMM
FeatureExtraction
Test data
𝐷 : the invers of the distance between the test data and each cluster
𝐷
𝐷Bedroom Kitchen
Cook Sleep- Wi-Fi- microphone- light- Acc
Feature extraction of IGMM input
Accelerator
Microphone Wi-Fi
Light sensor3-axis combination signal
Average MFCC components
Average of illumination
Signal strength values
Learning Activity and Position Model・Use non-parametric learning approach IGMM for activity and position clustering
Osaka UniversityAdvanced Telecommunications Research
Institute International
Proposed Method – Appliance selection using MKL
Cameraimage
Wi-Fi
- light- Acc- microphone
Selected Appliance
PositionIGMM
ActivityIGMM
Deep convolutional
neuralnetwork
Multiple kernel
learning
Image features
context features
A linear combination of multiple base kernels for image and context feature
describe a different property of the datawith multiple kernels
Multiple Kernel Learning𝒌𝒊𝒎𝒈,∗: 𝑝𝑜𝑙𝑦𝑛𝑜𝑚𝑖𝑎𝑙 𝑘𝑒𝑟𝑛𝑒𝑙 𝑓𝑜𝑟 𝑖𝑚𝑎𝑔𝑒 𝒌𝒄𝒐𝒏𝒕𝒆𝒙𝒕,∗: 𝑟𝑎𝑑𝑖𝑎𝑙 𝑏𝑎𝑠𝑖𝑠 𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛(𝑓𝑜𝑟 𝑐𝑜𝑛𝑡𝑒𝑥𝑡)
Decision Function : f 𝑥∗ 𝑎 𝑒 𝑘 ,∗ 𝑒 𝑘 ,∗ 𝑏
Osaka UniversityAdvanced Telecommunications Research
Institute International
Device for data collection Google Glass, Nexus 5 (in pocket) Sampling rate : 30HzSemi-naturalistic collection
protocol Activities follow the instruction 3 users X 10 sessions activities
prepare meals
eat meals
wash dishes
watch TV
…
sleep
Random
Activities
floor plan and appliances
toilet faucet
bedroom air conditioner
bedroom lighting
lounge lighting
TV
lounge air conditioner
kitchen faucet
kitchen curtain lounge curtain
kitchen lighting
drawer
front door
fan
Evaluation – Data set
Osaka UniversityAdvanced Telecommunications Research
Institute International
Evaluation Result - Leave-one-session out cross validation
Proposed: activity + position + camera
Proposed w/o pos:activity+ camera
Proposed w/o act: position + camera
Proposed w/ cam: only camera
SVM w/ cam: only camera
SVM all: activity + position + camera
Effect of context
F-measureRecallPrecision
10%
0.845
0.813
0.857
0.928
0.894
0.955
0.844
0.812
0.862
0.897
0.878
0.936
0.844
0.812
0.859
0.912
0.886
0.945
SVM all
SVM w/ cam
Proposed w/ cam
Proposed w/o act
Proposed w/o pos
Proposed
Osaka UniversityAdvanced Telecommunications Research
Institute International
Evaluation Result – Confusion MatrixVisual confusion matrix of Proposed w/ cam Visual confusion matrix of Proposed
• air conditioner and lighting were relatively poor • can’t distinguish between kitchen lighting and
bedroom lighting • drawer performed not well
• air conditioner and lighting were increased about 14% on average of F-measure
• F-measure improved by about 10% on total average
Osaka UniversityAdvanced Telecommunications Research
Institute International
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2 4 6 8 10 12
Proposed
Proposed (reuse)
Proposed (imgaenet)
Evaluation Result – Transition of Average F-measures
sessions
F-m
easu
re
Reusing other users’ data
Appliance Image Sensor Data
Position
Activity
+
Appliance Image Sensor Data
Position
Activity
+
Appliance Image Sensor Data
Position
Activity
+
User1User2
User3
-collect online images for each category
-find top-k similar images for each appliance
Utilizing online image database
32% 28%
Osaka UniversityAdvanced Telecommunications Research
Institute International
Conclusion
We proposed a new method of appliance selection with a smart glass based on position and activity contextual information
The effectiveness of contextual information in an appliance selection task has been confirmed in a real experiment environment.
Context based method can also be used to enhance the performance of such other appliance selection approaches as speech, gaze direction, and beacon- based approaches