This thesis explores software algorithm for implementing a people counting and matching system to be used on a bus. A special camera is used, known as a texel camera, that generates depth and color information for a scene. This added information greatly facilitates both the tasks of matching and counting. Although people counting is a relatively mature field, there are several situations in which current technologies are not able to count correctly. Several of these difficult situations are tested with 82% counting accuracy. The idea of matching people on a bus is also developed. The goal is not to identify a specific person on a bus, but to find the time that a specific person is on the bus, and what bus stops were used. There are several aspects of this matching problem that differentiate it from other classification tasks that have been researched. In this thesis, multiple measurements are used to classify a person and sequence estimation techniques explored. The techniques developed classify with 92% accuracy, even with a relatively large number of people on a bus.

This thesis explores pattern recognition in the dynamic setting of public transportation, such as a bus, as people enter and later exit from a doorway. Matching the entrance and exit of each individual provides accurate information about individual riders such as how long a person is on a bus and which stops the person uses. At a higher level, matching exits to entries provides information about the distribution of traffic flow across the whole transportation system. A texel camera is implemented and multiple measures of people are made where the depth and color data are generated. A large number of features are generated and the sequential floating forward selection (SFFS) algorithm is used for selecting the optimized features. Criterion functions using marginal accuracy and maximization of minimum normalized Mahalanobis distance are designed and compared. Because of the particular case of the bus environment, which is a sequential estimation problem, a trellis optimization algorithm is designed based on a sequence of measurements from the texel camera. Since the number of states in the trellis grows exponentially with the number of people currently on the bus, a beam search pruning technique is employed to manage the computational and memory load. Experimental results using real texel camera measurements show good results for 68 people exiting from an initially full bus in a randomized order. In a bus route simulation where a true traffic flow distribution is used to randomly draw entry and exit events for simulated riders, the proposed sequential estimation algorithm produces an estimated traffic flow distribution which provides an excellent match to the true distribution.