Many traffic engineering problems consist of evaluating a number of alternate schemes in terms of some measures of performance, such as total travel time, volume to capacity ratio, average speed, total fuel consumption, total emissions, etc. These evaluations generally require prior knowledge of the temporal trip making behaviour of drivers by origin and destination. This thesis presents the development, application, and evaluation of two models capable of inferring these temporal origin - destination (O-D) traffic demands on the basis of observed link traffic flows and assumptions regarding drivers' route choices. In particular, this thesis presents the development and evaluation of a Least Squared Error (LSE) and a Least Relative Error (LRE) model, each of which is capable of estimating either static demands, a time series of static demands, or dynamic demands. Furthermore, the potential of using probe data, from route guidance system (RGS) equipped vehicles, to enhance these estimated dynamic O-D demands is examined.

Both of the LSE and LRE models' mathematical formulations are presented. The LSE model formulation parallels that of a least squared regression as the error function is composed of the sum of the squared absolute difference between the observed and estimated link flows. In contrast, the LRE model is formulated on the basis that the link flow error, when measured relative to the observed flow, is to be minimized instead. Iterative solution algorithms, that are modifications of the Jacobi and Gauss-Seidel techniques, are proposed to solve each of the model formulations. It is shown, by way of the application of these iterative algorithms to several example networks, that the estimated O-D demands, which result from these iterative solution techniques, are consistent with the model formulations and with the analytical solutions. Furthermore, it is shown for several examples that, when multiple solutions exist which each exactly replicate the observed link flows, and no prior O-D demand information is specified, both the LSE and LRE models estimate demands that closely approximate the maximum likelihood solution.

The proposed iterative solution algorithms have been incorporated into a computer model called QUEENSOD. This model can be practically applied to real networks using current computer memory constraints. This thesis describes the application of the LSE and LRE models to a 35 km section of multilane urban freeway in Toronto, Canada, in which alternate parallel routes exist. Dynamic 15 minute O-D demands were estimated for the eastbound direction for the period from 5 am to 11 am. Despite FTMS detector data being available for only 45% of the network, a correlation coefficient of approximately 98% was obtained for both models. This value reflects the high linear correlation between estimated and observed link traffic flow data for this network. The statistical analysis of the expected quality of O-D demands, which are estimated solely on the basis of RGS probe vehicle data, indicated that even for levels of market penetration of 30 - 50%, the O-D estimates are unlikely to be of sufficient quality to be of practical benefit.