Fixed and mobile sensor based generalized additive models for freeway incident detection.

摘要

Generalized additive models (GAM) to detect lane-blocking and shoulder incidents are developed based on traffic measures estimated from fixed and mobile sensors. The generalized additive model, a nonparametric model, is a generalization of the generalized linear model, allowing appropriate functional forms of independent variables to be proposed. Generalized additive models allow flexible functions to be fitted and therefore their functional forms are revealed in the parametric estimate of generalized additive models. This capability of GAM serves as a powerful interpretive tool to examine the affect of each traffic measure on the probability of an incident. Fixed sensor based incident detection models are developed for lane-blocking and shoulder incidents on the Interstate 25 freeway in Colorado and the Interstate 880 freeway in California. Separate lane-blocking and shoulder incidents models are also developed for the Interstate 880 freeway to examine the characteristic differences between lane-blocking and shoulder incidents, as they relate to incident detection. Characteristics of incidents, model development, including significant variables selection, and model interpretation are also examined. Based on performance measures including detection rate, false alarm rate and mean time to detect, the nonparametric GAM and the parametric estimate of GAM, with only five variables for lane-blocking incidents and six variables for all incidents, outperform several neural network based models using 16 to 24 variables. In this research, the effect of type and length of freeway segments on model performance is also examined.; Mobile sensor, and fixed and mobile sensor based incident detection models are developed for lane-blocking and shoulder incidents on the Interstate 25 freeway. The performance of mobile sensor based model shows the potential use of mobile sensor as an alternative data source. Using mobile sensor as an additional data source to fixed sensor data helps reduce the false alarm rate of the incident detection model.; The performance of the incident detection models developed is unbiasedly validated using bootstrap method. The bootstrap performance examined includes mean detection rate, incident state detection rate, false alarm rate, mean time to detect, and their 95 percent confidence interval. The bootstrap performance may provide a good estimate of model performance in the field.