چكيده به لاتين
The ability of traffic microsimulation software in modeling different traffic phenomena, and offering high levels of information and details have made them unique tools for analyzing complex traffic systems. Numerous microsimulation tools have been introduced in the recent decade. Car-following and lane-changing models are known as two of the most critical underlying behavioral models of these tools that describe the longitudinal movement of the vehicle (along a single lane) and its lateral behavior, respectively. Despite the high variation, almost all of these models are developed based on homogeneous, lane-based traffic flow theory, and also based on some key assumptions such as homogenous vehicles types, movement of vehicles along the center line of a lane, being influenced only by the leader vehicle(s) in front, and a discrete lateral movement through a lane-changing model. Due to the underlying assumptions of these models, they are not appropriate for modeling traffic disordered systems where there is no lane discipline and drivers using lateral spaces to move forward. Therefore, using lane-based driving behavior models in traffic microsimulation tools for modeling heterogeneous traffic flow could lead to unrealistic results. Hence, developing microscopic traffic models for such systems are required. On the other hand, the availability of appropriate trajectory data as well as efficient algorithms for developing and calibration of the non-lane-based model is important.
Accordingly, in the present research, having a comprehensive insight into the driving behavior modeling, four phases consisting of vehicles’ trajectory data collection, reconstructing and modifying the trajectory data, developing a non-lane-based driving model and calibration and validation of the model have been considered. In data collection phase, a software was developed for extracting the trajectory data from the videos collected by a quad-copter, in which a series of efficient algorithms for video stabilization, vehicles detection and tracking in sequences of pictures. Comparing the results with the manually extracted data indicated the good performance and high accuracy of the developed algorithms for automatic extraction of the data. In the reconstruction and modification phase, an algorithm was developed based on wavelet transformations in which the modification of the data is applied in two steps. In the first step, the outliers are identified and then modified locally. Finally, in the second step, the noises in the data are reduced. The comparison of the proposed methodology with the latest method in the literature shows better performance of the proposed one. This method can reduce more noise while it preserves the useful information and the trajectory structure.
In the model development phase, a general framework is presented based on the field theory. This model can be applied in either the homogeneous or heterogeneous traffic flows. In this model, the peripheral environment of the driver is considered as a potential field, elements of which (including adjacent vehicles, traffic signs, and roadway edges) impose forces to the driver, and the superposition of perceived forces determines how the driver moves longitudinally and laterally. By using this approach, simultaneous modeling of the longitudinal and lateral movements are feasible. The Simulation results and evaluation of the different parts of the proposed model indicate the rationality of the model and its proper performance. Furthermore, the structure of the model is flexible such that new motivations (forces caused by new elements in the perceived field of the driver) can be added to the model with little modification.
Finally, in the calibration phase, an efficient algorithm for estimating parameters of microscopic models was presented. In this algorithm, mutation and crossover operators in the genetic algorithm have been replaced by learning and sampling from a probabilistic model. The comparison of the proposed algorithm with the genetic algorithm, which is widely used in the calibration of microsimulation models, indicate better performance of the proposed methodology in terms of accuracy and time consumption (cost) compared to the genetic algorithm.
