Recognition of Human Behavior Patterns Using Depth Information and Gaussian Feature Maps

The representation of human behavior patterns is challenging due to the complex dependencies between features gathered by a sensor and their spatial and temporal context. In this work we propose a new Gaussian feature map representation that uses the Kinect depth sensor, can easily be integrated in home environments, and allows learning unsupervised behavior patterns. The approach divides the living space into grid cells and models each grid cell with a Gaussian distribution of features like height, duration, magnitude and orientation of the velocity. Experimental results show that the method is able to recognize anomalies regarding the spatial and temporal context.