Activity Recognition in Traffic Data

I did this project in collaboration with my graduate student Jim Howard. The problem we addressed was how to more accurately predict (or forecast) traffic data. Accurate traffic forecasting is of great interest for commercial, security, and efficiency applications. For example, consider the problem of predicting the occupancy of a large building. If you know that a room will be occupied at a certain day and time, you can pre-heat or pre-cool the room just prior to its use, instead of always keeping the room at a set temperature. Forecasting traffic data can lead to improvements in control systems for traffic lights, heating and cooling systems for buildings or even determining if anomalous events are potential security threats. We developed a method to automatically detect and model anomalies in sensor data, and train the system to recognize an anomaly when it is happening, to generate a more accurate forecast.

We developed a method to learn better forecasting models, from occupancy data collected by sensors. We used a variety of existing datasets for vehicle traffic and also building occupancy. We also collected our own data on building occupancy. We placed 50 passive infrared sensors mounted on the ceilings and in some rooms of a large campus building. Data was collected for one academic year, and there were more than 23 million sensor readings. To acquire readings, the sensors were polled every second and recorded data if motion was detected. Data is on GitHub at https://github.com/ahwhoff/mines-sensor-network.

Our method used a “Bayesian Combined Forecasting” approach, which used an ensemble of forecasting models. These models included Seasonal ARIMA, Historic average, Time Delayed Neural Network, and Support Vector Regression. The BCF approach improved the forecasting accuracy over using any of the component models.

For full details, see: J. Howard, W. Hoff, “Forecasting Building Occupancy Using Sensor Network Data,” Proc. of 2nd International Workshop on Big Data, Streams and Heterogeneous Source Mining (BigMine 13), pp. 87-94, August 2013, Chicago, Illinois. (pdf)