Tracking the Dynamic Distribution of People in Indoor Space with Noisy Partitioning Sensors.

Song Wang,Xiaoyang Sean Wang,Yan Huang

DOI: https://doi.org/10.1109/mdm.2012.56

2012-01-01

Abstract:The term "indoor" here refers generally to enclosed space partitioned into subspaces with connecting doors or gates. Examples include the inside of office buildings, amusement parks, and indoor shopping malls. In many applications, it is desirable to keep track of the distribution of people within the enclosed space. These applications range from smart house with automatically controlled air-conditioning and lighting, shopping assistance allocation, to business intelligence. Contact sensors are accurate but obstructive. Non-contact sensors such as automated visual recognition and RFID tags can be expensive in calibration or cost in order to obtain accurate readings. An interesting cost optimization problem is to understand how to obtain relatively accurate dynamic distribution of people from inaccurate point sensors using correlations among the point sensor readings. The paper formalizes a framework that uses a flow model and a particle-filter learning algorithm for this optimization problem based on the hypotheses that (1) there is an underlying stochastic "flow model" of people moving within the space, and (2) with the help of this flow model, counting of people can be made more accurate by taking advantage of the continuous, albeit inaccurate, point sensor readings. The main challenge is that the performance of particle filters deteriorates rapidly with the number of doors in the indoor space. We propose a divide and conquer method that uses relatively more accurate sensors at a few strategically chosen locations to achieve overall good accuracy. Experimental results given herein show that the algorithm is effective.