On accurate and efficient statistical counting in sensor-based surveillance systems

Sensor networks have been used in many surveillance systems, providing statistical information about monitored areas. Accurate counting information (e.g., the distribution of the total number of targets) is often important for decision making. As a complementary solution to double-counting in communication, this paper presents the first work that deals with double-counting in sensing for wireless sensor networks. The probability mass function (pmf) of target counts is derived first. This, however, is shown to be computationally prohibitive when a network becomes large. A partitioning algorithm is then designed to significantly reduce computation complexity with a certain loss in counting accuracy. Finally, two methods are proposed to compensate for the loss. To evaluate the design, we compare the derived probability mass function with ground truth obtained through exhaustive enumeration in small-scale networks. In large-scale networks, where pmf ground truth is not available, we compare the expected count with true target counts. We demonstrate that accurate counting within 1%-3% relative error can be achieved with orders of magnitude reduction in computation, compared with an exhaustive enumeration-based approach.