Inferring High-Level Behavior from Low-Level Sensors

1. Inferring High-Level Behavior from Low-Level Sensors Donald J. Patterson, Lin Liao, Dieter Fox, and Henry Kautz

2. Outline • Authors • Objective • Basic idea • Implementation outline • Experiment • Result • Conclusion • Related work

3. Authors • Donald J Patterson • Assistant professor of Donald Bren School of Information and Computer Sciences at the University of California, Irvine. • Lin Liao • PhD student of the University of Washington. • Dieter Fox • Associate Professor in the Department of Computer Science & Engineering at the University of Washington • Henry Kautz • University of Rochester

4. Objective • The authors would want to recognize and predict the high-level intentions and complex behaviors that cause particular physical movements through space. • Such higher-order models would both enable the creation of new computing services that autonomously respond to a person’s unspoken needs, and support much more accurate predictions about future behavior at all levels of abstraction.

5. In this paper… • This paper presents an approach to learning how a person uses different kinds of transportation in the community. • A key to inferring high-level behavior is fusing a user’s historic sensor data with general commonsense knowledge of real-world constraints. • The authors introduce a three-part model in which a low-level filter continuously corrects systematic sensor error, a particle filter uses a switching state-space model for different transportation modes, and a street map guides the particles through the high-level transition model of the graph structure. • They additionally show how to apply Expectation-Maximization (EM) to learn typical motion patterns of humans in a completely unsupervised manner.

6. The model of the world • The model of the world is a graph G = (V, E) which has a set of vertices and a set of directed edges. • The state of an object:

7. Basic idea • To estimate the location and transportation mode of a person we apply Bayes filters, a probabilistic approach for estimating the state of a dynamic system from noisy sensor data. • Uncertainty is handled by representing all quantities involved in the estimation process using random variables. • The authors assumed that the state space conforms to the first-order Markov independence assumption.

8. Basic idea (cont.) • The model assumes that velocities are drawn randomly from these Gaussians,where the probability of drawing from a particular Gaussian depends on the mode.

9. Basic idea (cont.) • In the current approach, the probabilities for the Gaussians in thedifferent transportation modes were set manually based on external knowledge. • The motion mode at time only depends on the previous mode and thepresence of a parking lot or bus stop.

10. Particle Filter Based Implementation

11. Expectation-Maximization (EM) algorithm • Each E-step estimates expectations (distributions) over the hidden variables using the GPS observations along with the current estimate of the model parameters. • Then in the M-step the model parameters are updated using the expectations of the hidden variables obtained in the E-step.

12. E-step • Θ: the parameters of the graph-based model we want to estimate • Θ(i-1): the estimation thereof at the i-1-th iteration of the EM algorithm. • The E-step estimates

13. M-step • The goal of the M-step is to maximize the expectation of logp(z1:t, x1:t | Θ) over the distribution in the E-step by updating the parameter estimations.

14. Experiments • The test data set consist of logs of GPS data collected by one of the authors. The data contain position and velocity information collected at 2-10 second. • The data was hand labeled with one of three modes of transportation: foot, bus, or car. • 29 episodes which represent a total of 12 hours of logs were divided chronologically into three groups which formed the sets for three-fold cross-validation for learning.

16. Mode Estimation and Prediction

17. Location Prediction

18. Conclusions • Authors demonstrated that good predictive user-specific models can be learned in an unsupervised fashion. • The combination of general knowledge and unsupervised learning enables a broad range of “self-customizing” applications.

19. Future Work • Making positive use of negative information. • Learning daily and weekly patterns. • Modeling trip destination and purpose. • Using relational models to make predictions about novel events.

20. Related Work • Extracting Places and Activities from GPS Traces Using Hierarchical Conditional Random Fields.L. Liao, D. Fox, and H. Kautz. International Journal of Robotics Research, 2007 • Location-Based Activity Recognition.L. Liao, D. Fox, and H. Kautz. NIPS-05. • Bayesian filtering for location estimation. D. Fox, J. Hightower, L. Liao, D. Schulz, and G. Borriello. IEEE Pervasive Computing, 2003. • Opportunity Knocks: a System to Provide Cognitive Assistance with Transportation Services. D. J. Patterson, L. Liao, K. Gajos, M. Collier, N. Livic, K. Olson, S. Wang, D. Fox, and H. Kautz. UBICOMP-04. • Learning and Inferring Transportation Routines. L. Liao, D. Fox, and H. Kautz. AAAI-04. • Voronoi Tracking: Location Estimation Using Sparse and Noisy Sensor Data.L. Liao, D. Fox, J. Hightower, H. Kautz, and D. Schulz. IROS-03.

21. Related Topic to Study • Particle filter • EM algorithm • Hidden Markov model (HMM)