Project

Chameleon

Chameleon is a wall-mounted sensor system that uses machine learning to classify building activity.

We present the system as an alternative that can make smart sensing within a building easier to scale.  The device is capable of eliminating the need for constant maintenance and heavy upfront infrastructure costs while keeping user privacy concerns at the center of its design.  Its novel architecture allows it to be used within rooms with different space layouts, ventilation systems, windows and usage patterns such as offices, classrooms and homes. 

The system uses the latest version of the MIT terMITe sensors developed in the City Science group. The new hardware adds capabilities for sensing CO2 particles and movement. These types of sensors are inherently less intrusive than cameras or other optical systems, making them a valuable alternative for privacy preserving building infrastructure. 

Chameleon is a wall-mounted sensor system that uses machine learning to classify building activity.

We present the system as an alternative that can make smart sensing within a building easier to scale.  The device is capable of eliminating the need for constant maintenance and heavy upfront infrastructure costs while keeping user privacy concerns at the center of its design.  Its novel architecture allows it to be used within rooms with different space layouts, ventilation systems, windows and usage patterns such as offices, classrooms and homes. 

The system uses the latest version of the MIT terMITe sensors developed in the City Science group. The new hardware adds capabilities for sensing CO2 particles and movement. These types of sensors are inherently less intrusive than cameras or other optical systems, making them a valuable alternative for privacy preserving building infrastructure. 

The device uses novel machine learning algorithms to create models that are able to classify room activity states.  Chameleon combines Spectral Clustering and Recurrent Neural Networks to automatically update classification models every seven days. The clustering algorithm is able to discretize different activity states such as a space being used by many people while having good ventilation or a poorly ventilated space being used by only one person.  Clusters are then used to train a Neural Network so that live classification is carried our more efficiently. 

We deployed and tested the architecture in an office and a classroom with entirely different geographical location and layouts. The system was in operation for one month, completing four cycles of retraining and classification. Our experiment evaluations show that the system has classification accuracies that range from 87% - 99%, indicating that the approach is an interesting alternative for smart building infrastructure that can be scaled while keeping privacy concerns at the center of its design.