In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
will defend his dissertation
Occupant Detection and Tracking in Smart Buildings using Unobtrusive Sensing
Non-intrusive occupant identification enables numerous applications in smart buildings such device-free climate control and adjustment in buildings which would not only enhance occupants' comfort but also save energy as we show in this thesis. To achieve that, we propose a set of methods to identify occupants by sensing their body shape, movement, behaviors, and the path taken as they walk through the door or a network of doors. We mount three ultrasonic ping sensors, one on top to sense height and two on the sides of the door to sense width. We extract a feature set from the occupant's walk to non-intrusively identify him. We cluster the occupants using their waist girth and the time spent under the door. We use DBSCAN for clustering because it discovers the number of clusters and also takes into consideration the precision of the sensors. We deployed our system in a classroom for a month, and 20 people participated. Our current model is able to identify 20 occupants with an accuracy of 95\%. This technology though it is 4 times better than the state of the art is not applicable to realistic building settings where the number of occupants is far bigger than 20. To achieve building-level occupant identification, we propose a set of improvements to the sensing platform which led to increasing the sampling rate from 30Hz to over 145Hz by performing sensor sampling optimizations and sampling parallelization. Then, we propose a novel methodology that rather than relying on our clustering algorithm to point to the identity of the user, we only filter down possible candidates and model the paths users take in the building using a Markov Model. We also develop a set of algorithms that detect a set of "behaviors" that users perform as they walk through the door such as holding a phone, wearing a backpack, carrying a handbag or wearing heels. We show that such behaviors skew the data making the user appear as a different one and therefore leads in misidentification. We propose developed a set of algorithms that first detects these common "behaviors" and correct the data to mitigate their effect on the model. We were able to accurately identify 100 people in a building using a network of 5 doors. This makes the use of such a system useful in buildings as it enables buildings to be aware of their users and their preference which leads to increased comfort and energy savings of the around 10\% beyond regular occupancy-based energy saving systems.
Date: Monday, April 2, 2018
Time: 3:00 PM
Place: PGH 501D
Advisor: Dr. Omprakash Gnawali
Faculty, students, and the general public are invited.