The application of fuzzy logic and neural networks to freeway incident detection

The purpose of this research is to improve the reliability of automatic freeway incident detection by applying the fuzzy logic system and neural networks. A Fuzzy Logic Incident Patrol System (FLIPS) is proposed to combine a fuzzy logic system and the learning capabilities of neural networks to form a connectionist model. The proposed system can be constructed automatically from training examples to find the optimum input/output term sets, input/output membership functions, and fuzzy logic rules.;The output of FLIPS is an incident detection index which represents the possibility of an incident. These continuous indices can be sent to a freeway traffic operations center for further analysis. One example of such further analysis is the conversion of the continuous incident index value into a binary value predicting the absence or presence of an incident. The Fuzzy Logic Adaptive Threshold System (FLATS) is developed for this purpose. The FLATS produces an adaptive threshold for the conversion task based on the current incident detection index provided by FLIPS. FLATS itself is also a fuzzy logic system with a recurrent neural network structure. Therefore, the threshold serves both as input and output parameters for FLATS. Like FLIPS, the FLATS can also be generated automatically from training examples.;In total, four artificial intelligence incident detection models have been developed to evaluate the benefits of the proposed integrated learning process and adaptive threshold technique. The evaluation results show that the integrated learning process and adaptive threshold technique can significantly improve incident detection performance. The FLATS-I, which incorporates proposed integrated learning process and adaptive threshold technique, exhibits superior incident detection performance.;Finally, the research conducted comparisons of the FLATS-I with conventional incident detection algorithms (i.e., the McMaster algorithm, the California algorithm, and the filtering technique) using the same incident database collected in Toronto, Canada. The evaluation results indicate that the FLATS-I provides promising improvements in the incident detection and the persistence check technique used in the McMaster algorithm may not be an effective method to improve incident detection.