| Summary |
Yi-lin Chen, 2006.06
Feng Chia University - Graduate Institute of Traffic and Transportation Engineering and Management
Freeway system plays an important role in intercity transportation. Once traffic jams or accidents happen, the traffic would be seriously deterred and even causes other accidents. Thus, it is essential to establish an effectively incident detection system, which can provide correct and prompt alarmed incident signal to enhance the efficiency of accidents responsive actions. Genetic Fuzzy Logic Controller (GFLC) can not only self-learn the optimal combination of fuzzy rules and shapes of membership functions, but also performs very well in many applications. Based on this, the study aims to develop a GFLC-based freeway incident detection system.
The traffic information real-time detected by vehicle detection devices include volume, speed and occupancy, which are commonly used to develop incident detection system. An incident is said to be detected if a significant gap of these traffic information has been identified within different time horizons or at different vehicle detectors (upstream or downstream of the incident spot). Since the potential rules will rapidly get enlarged as the number of state variables increases, generally, the number of state variables would not exceed three. Therefore, this study develops four GFLC-based incident models based on detected traffics. They are volume model, speed model, occupancy model and integrated model, where volume model only considers the volume variables of different time horizons and vehicle detectors, speed model only considers speed variables, occupancy model only considers occupancy variables and integrated model simultaneously considers these three variables of different vehicle detectors. In addition, in order to consider as many as variables with subject to the number constraints, principal component method is used to choose three principal components, which are the linear combination of nine variables, as state variables of GFLC. This model is named as principle components model. For comparison, an artificial neural network (ANN)-based incident system is also developed, which simultaneously considers all these nine variables.
To investigate the performance and applicability of proposed models, the 20-seconds traffic data of a total of 30 accidents on national No.1. Freeway in Taiwan are collected and used to conduct a case study. Three commonly used index: detection rate (DR), false alarm rate (FAR) and mean time for detection (MTD) are adopted to measure the performances of models. The results show that the principal component model outperforms among all these incident detections models with DR=100%, FAR=0.92% and MTD=17.6 seconds, followed by ANN model with DR=96.67%, FAR=1.43% and MTD=16.0 seconds and integrated model with DR=93.33%, FAR=1.26% and MTD=18.2 seconds. Other three GFLC models based on only one kind of traffic information (volume, speed and occupancy) perform relatively inferior. |
