DEVELOPING HETEROGENEOUS TRAFFIC STATE MEASUREMENT TECHNIQUES BASED ON IMAGE PROCESSING AND MODELING

Abstract

In Dhaka city traffic operating condition is predominantly non-lane-based and heterogeneous in nature. Accurate measurement of traffic state is the first and principal task for describing and improving such traffic condition. For this, video-based image processing algorithms are preferred over other alternatives. This research initiates with the development of a novel traffic detection algorithm for detecting non-lane-based heterogeneous traffic. The algorithm uses mainly two different parameters: (1) static background; and (2) threshold parameter. The algorithm achieves high accuracy of detection through small computational effort. The qualitative analysis shows that compared to others the proposed algorithm can detect vehicles accurately. The quantitative analysis of the algorithm shows stable Precision-Recall Relationship. The RMSE values computed from ground truth and estimated traffic parameters show that the proposed algorithm outperforms other state of the art algorithms. Given the difficulty of obtaining variable traffic demands in the real world and the cost and time associated with the field data collection, a controlled environment is needed where the demand can be changed artificially and the overall network performance can be observed. Microscopic simulators (e.g. VISSIM) are capable of providing such controlled environment. In this research a generic calibration tool, VISCAL has been developed, for microscopic simulation parameters in VISSIM environment. The optimization system of the tool is based on three heuristic algorithms: (a) GA; (b) SPSA; (c) SA. VISCAL can be used to calibrate any type (rural, urban etc.) and extent (large, medium etc.) of network. Again macroscopic traffic flow models play an irreplaceable role in real-time traffic state estimation & prediction and represent the traffic states with the help of aggregated variables. The two most frequently used macroscopic models are the first-order cell transmission model (CTM) and the second-order METANET model. In this research a 1st order CTM model was developed for non-lane-based heterogeneous urban traffic condition. At first the nature of the fundamental traffic relationships was systematically investigated based on the traffic data. From regression analysis it was concluded that 3rd degree polynomial structure shows the best fit with vii the measured traffic data. As classical CTM tends to simulate inappropriate flow and speed behavior due to simple and linear consideration of FD, a new model was proposed in this research. Detail investigation shows that, the non-linear FD plays the most important role in developing CTM model, for estimating traffic state accurately in heterogeneous traffic operating condition. Again this research proposes a new second-order macroscopic traffic flow model having the following special features: (1) both the flow and speed dynamics have a normally distributed stochastic term; (2) the FD in the speed dynamics follows Zhang‘s one-parameter polynomial structure and (3) the parameters of the FD are variable over the links. In the model calibration stage, simultaneous optimization of FD parameters and driver-related parameters were conducted. The optimized parameters capture the existing traffic conditions of the respective links quite well. The link-specific FD parameters and the stochastic traffic state influencing terms improve the model performance the most, followed by the Car Following parameter. Proposed model performs most poorly in the absence of FD in the speed dynamics. Thus it can be concluded that FD affects the traffic states very seriously for heterogeneous composition and cannot be dropped off from the speed dynamics for simplicity in control design. The predicted speed, flow and density from the developed macroscopic model were compared with those from a microscopic simulation model, VISSIM. Four levels of traffic demands viz. light, moderate, heavy and excessive demand levels were applied to evaluate the compatibility of the two models. Based on the performance of the models and comparative analysis, the following main conclusions are found: i) The prediction of traffic states from the proposed stochastic METANET-based model is generally consistent with that from VISSIM simulation over the whole range of traffic demand levels used in this research. ii) The MAEs in traffic state estimation of various links do not show any distinct trend with the change of traffic demand levels, thus indicating that the developed macroscopic model performs quite satisfactorily for different traffic demand levels.