Bilge Mutlu, Andreas Krause, Jodi Forlizzi, Carlos Guestrin, and Jessica Hodgins Human-Computer Interaction Institute, Carnegie Mellon University Robust,

Slides:



Advertisements
Similar presentations
Automatic Photo Pop-up Derek Hoiem Alexei A.Efros Martial Hebert Carnegie Mellon University.
Advertisements

Distinctive Image Features from Scale-Invariant Keypoints David Lowe.
ECG Signal processing (2)
Image classification Given the bag-of-features representations of images from different classes, how do we learn a model for distinguishing them?
Cost-effective Outbreak Detection in Networks Jure Leskovec, Andreas Krause, Carlos Guestrin, Christos Faloutsos, Jeanne VanBriesen, Natalie Glance.
Submodular Dictionary Selection for Sparse Representation Volkan Cevher Laboratory for Information and Inference Systems - LIONS.
Watching Unlabeled Video Helps Learn New Human Actions from Very Few Labeled Snapshots Chao-Yeh Chen and Kristen Grauman University of Texas at Austin.
Iowa State University Department of Computer Science Artificial Intelligence Research Laboratory Research supported in part by grants from the National.
Feature Selection Presented by: Nafise Hatamikhah
Near-optimal Nonmyopic Value of Information in Graphical Models Andreas Krause, Carlos Guestrin Computer Science Department Carnegie Mellon University.
A Practical Approach to Recognizing Physical Activities Jonathan Lester Tanzeem Choudhury Gaetano Borriello.
1 Efficient planning of informative paths for multiple robots Amarjeet Singh *, Andreas Krause +, Carlos Guestrin +, William J. Kaiser *, Maxim Batalin.
Nonmyopic Active Learning of Gaussian Processes An Exploration – Exploitation Approach Andreas Krause, Carlos Guestrin Carnegie Mellon University TexPoint.
Vision-based Control of 3D Facial Animation Jin-xiang Chai Jing Xiao Jessica Hodgins Carnegie Mellon University.
7/17/2002 Greg Grudic: Nonparametric Modeling 1 High Dimensional Nonparametric Modeling Using Two-Dimensional Polynomial Cascades Greg Grudic University.
1 Automated Feature Abstraction of the fMRI Signal using Neural Network Clustering Techniques Stefan Niculescu and Tom Mitchell Siemens Medical Solutions,
+ Doing More with Less : Student Modeling and Performance Prediction with Reduced Content Models Yun Huang, University of Pittsburgh Yanbo Xu, Carnegie.
Activity Recognition from User- Annotated Acceleration Data Presented by James Reinebold CSCI 546.
© 2013 IBM Corporation Efficient Multi-stage Image Classification for Mobile Sensing in Urban Environments Presented by Shashank Mujumdar IBM Research,
Learning from Multiple Outlooks Maayan Harel and Shie Mannor ICML 2011 Presented by Minhua Chen.
Hierarchical Exploration for Accelerating Contextual Bandits Yisong Yue Carnegie Mellon University Joint work with Sue Ann Hong (CMU) & Carlos Guestrin.
Sean Ryan Fanello. ^ (+9 other guys. )
Transfer Learning From Multiple Source Domains via Consensus Regularization Ping Luo, Fuzhen Zhuang, Hui Xiong, Yuhong Xiong, Qing He.
Comparison of Boosting and Partial Least Squares Techniques for Real-time Pattern Recognition of Brain Activation in Functional Magnetic Resonance Imaging.
Gwangju Institute of Science and Technology Intelligent Design and Graphics Laboratory Multi-scale tensor voting for feature extraction from unstructured.
1 Li Li [WSC17] Institute of Integrated Sensor Systems Department of Electrical and Computer Engineering Multi-Sensor Soft-Computing System for Driver.
Learning Visual Bits with Direct Feature Selection Joel Jurik 1 and Rahul Sukthankar 2,3 1 University of Central Florida 2 Intel Research Pittsburgh 3.
Mining Discriminative Components With Low-Rank and Sparsity Constraints for Face Recognition Qiang Zhang, Baoxin Li Computer Science and Engineering Arizona.
Kaihua Zhang Lei Zhang (PolyU, Hong Kong) Ming-Hsuan Yang (UC Merced, California, U.S.A. ) Real-Time Compressive Tracking.
Crowdsourcing for Spoken Dialogue System Evaluation Ling 575 Spoken Dialog April 30, 2015.
Kernel Methods A B M Shawkat Ali 1 2 Data Mining ¤ DM or KDD (Knowledge Discovery in Databases) Extracting previously unknown, valid, and actionable.
MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation 1 Dynamic Sensor Resource Management for ATE MURI.
Today Ensemble Methods. Recap of the course. Classifier Fusion
Vehicle Segmentation and Tracking From a Low-Angle Off-Axis Camera Neeraj K. Kanhere Committee members Dr. Stanley Birchfield Dr. Robert Schalkoff Dr.
CROSS-VALIDATION AND MODEL SELECTION Many Slides are from: Dr. Thomas Jensen -Expedia.com and Prof. Olga Veksler - CS Learning and Computer Vision.
Paired Sampling in Density-Sensitive Active Learning Pinar Donmez joint work with Jaime G. Carbonell Language Technologies Institute School of Computer.
1 Research Question  Can a vision-based mobile robot  with limited computation and memory,  and rapidly varying camera positions,  operate autonomously.
Interactive Learning of the Acoustic Properties of Objects by a Robot
5 Maximizing submodular functions Minimizing convex functions: Polynomial time solvable! Minimizing submodular functions: Polynomial time solvable!
Classification (slides adapted from Rob Schapire) Eran Segal Weizmann Institute.
CS558 Project Local SVM Classification based on triangulation (on the plane) Glenn Fung.
© Devi Parikh 2008 Devi Parikh and Tsuhan Chen Carnegie Mellon University April 3, ICASSP 2008 Bringing Diverse Classifiers to Common Grounds: dtransform.
Sparse Granger Causality Graphs for Human Action Classification Saehoon Yi and Vladimir Pavlovic Rutgers, The State University of New Jersey.
Chapter 8. Learning of Gestures by Imitation in a Humanoid Robot in Imitation and Social Learning in Robots, Calinon and Billard. Course: Robots Learning.
Data Mining and Decision Support
Feature Selction for SVMs J. Weston et al., NIPS 2000 오장민 (2000/01/04) Second reference : Mark A. Holl, Correlation-based Feature Selection for Machine.
A Kernel Approach for Learning From Almost Orthogonal Pattern * CIS 525 Class Presentation Professor: Slobodan Vucetic Presenter: Yilian Qin * B. Scholkopf.
Next, this study employed SVM to classify the emotion label for each EEG segment. The basic idea is to project input data onto a higher dimensional feature.
Neural networks (2) Reminder Avoiding overfitting Deep neural network Brief summary of supervised learning methods.
Shape2Pose: Human Centric Shape Analysis CMPT888 Vladimir G. Kim Siddhartha Chaudhuri Leonidas Guibas Thomas Funkhouser Stanford University Princeton University.
High resolution product by SVM. L’Aquila experience and prospects for the validation site R. Anniballe DIET- Sapienza University of Rome.
Signal and Image Processing Lab
Guillaume-Alexandre Bilodeau
Action-Grounded Push Affordance Bootstrapping of Unknown Objects
Near-optimal Observation Selection using Submodular Functions
Boosted Augmented Naive Bayes. Efficient discriminative learning of
Reflectance Function Approximation
Alan Qi Thomas P. Minka Rosalind W. Picard Zoubin Ghahramani
Vijay Srinivasan Thomas Phan
Dynamical Statistical Shape Priors for Level Set Based Tracking
Video-based human motion recognition using 3D mocap data
Machine Learning Feature Creation and Selection
Chao Xu, Parth H. Pathak, et al. HotMobile’15
ECE539 final project Instructor: Yu Hen Hu Fall 2005
Multi-Sensor Soft-Computing System for Driver Drowsiness Detection
Classification Breakdown
Machine Learning in Practice Lecture 27
Wellington Cabrera Advisor: Carlos Ordonez
One-shot learning and generation of dexterous grasps
High-Level Vision Object Recognition II.
Presentation transcript:

Bilge Mutlu, Andreas Krause, Jodi Forlizzi, Carlos Guestrin, and Jessica Hodgins Human-Computer Interaction Institute, Carnegie Mellon University Robust, Low-cost, Non-Intrusive Sensing and Recognition of Seated Postures

Why seated postures? Automobile Classroom Wheelchair Home Office

Using posture information Today’s talk

Pellegrini and Iocchi., 2006 Kinesthetic Motion-capture markers or conductive- elastomer-embedded fabrics Existing approaches

Kinesthetic Motion-capture markers or conductive- elastomer-embedded fabrics Vision-based Image sequences from a single camera or multiple cameras Tognetti et al., 2005 Existing approaches

Kinesthetic Motion-capture markers or conductive- elastomer-embedded fabrics Vision-based Image sequences from a single camera or multiple cameras Pressure-sensing-based Pressure readings from the seating surfaces Existing approaches Han et al., 2001

Poor generalization Good performance in classifying “familiar” subjects, poor performance with “unfamiliar” subjects due to high dimensionality. High cost High-fidelity pressure sensors are expensive. Slow performance Processing high-fidelity sensor data demands computational power, which leads to slow processing. Challenges Robust generalization Low-cost Near-real-time performance

Our solution Robust generalization Up to 87% accuracy in classifying 10 postures with new subjects. Low cost Using 19 pressure sensors instead of Reducing sensor cost from $3K to ~$100. Near-real-time performance 10Hz on a standard desktop computer Novel methodology Using domain knowledge and near- optimal sensor placement.

Methodology

Learning Algorithm Logistic Regression Sparse representation Cross-validation 10-fold, gender-balanced training and testing samples from different subjects Separate sets Training, testing, and reporting samples from 52 people in 5 trials Implementation in Java ✴ We would like to thank Hong Tan and Lynne Slivovsky for providing their data set for comparison. ✴

Understanding pressure data Modeling

Understanding pressure data Modeling

Understanding our data Modeling

Domain knowledge Modeling

Features Modeling Size and position of bounding boxes Distances to the edges of the seat Distance and angle to between bounding boxes Parameters of the ellipses that fit the bottom area Pressure applied to the bottom area

Features Modeling Classification accuracy

Separability test Modeling

Feature elimination Modeling

Methodology

Dimensionality Reduction Sensor granularity

Dimensionality Reduction Sensor granularity

How to place sensors? F, feature variables V, locations and granularities A subset A of V that maximizes information gain about F where H is entropy NP-Hard optimization problem We use near-optimal approximation algorithm Dimensionality Reduction IG(A;F) = H(F) - H(F | A) F V A ⊆ V

Near-optimal placement Dimensionality Reduction

Sensor placements Dimensionality Reduction

Near-optimal placement Dimensionality Reduction Classification accuracy

Methodology

Prototyping

Evaluation of prototype 20 naive participants 10-fold cross validation testing with %5 of the data 78% accuracy In classifying 10 postures 10 Hz real-time performance On a standard desktop computer

Methodology

Conclusions Generalizability Up to 87% (with a base rate of 10%) achieved with unfamiliar subjects. Low cost Higher classification accuracy than existing systems using less than 1% of the sensors. ~ $100 sensor cost compared to the commercial sensor for $3K (33 times reduction in price). Near-real-time performance At 10Hz on a standard desktop computer.

Applications Automobile Classroom Wheelchair Home Office

Future challenges Transferring learning across chairs A “transformation map” could be created Only static postures Temporal dimension needs to be considered The set of ten postures The set of postures should come from the activity Next Steps

Summary of Contributions A non-intrusive, robust, low-cost system that recognizes seated postures with generalizable, near-real-time performance. A novel methodology that uses domain-knowledge and near- optimal sensor placement strategy for classification. This work was supported by NSF grants IIS , DGE , CNS , Intel Corporation and Ford Motor Company.

From Postures to Activities Reading the paper Watching TV Reading paperwork Watching TV + eating Sleeping Talking on the phone Reading a book Craftwork Reading the paper + watching TV Reading the paper + eating Next Steps