A wavelet-based freeway incident detection algorithm with adapting threshold parameters

This paper presents a wavelet-based novel freeway automated incident detection algorithm with varying threshold parameters considering the level of traffic flow. In this approach, new test statistics for incident detection are extracted from occupancy and speed data using discrete wavelet transform, which decomposes traffic measurements into different resolution-time components. Unlike conventional incident detection algorithms, which apply fixed threshold values and often result in undesirably high false alarm rates, our proposed algorithm varies its threshold values adaptively based on the level of traffic volume. We have derived the mathematical relationship between the false alarm probability and the threshold value of our proposed decision function. For a given target false alarm rate, the threshold values can be changed adaptively depending on the traffic levels of normal traffic conditions. Also, we propose the new feature selection technique to measure the quality of different features that may be used to discriminate between normal and incident traffic conditions. Using both simulated data set and real-life incident data set, the performance of our proposed algorithm was compared with existing popular approaches such as California algorithm, Minnesota algorithm, conventional neural networks algorithm, and a wavelet-based neural-net algorithm. Experimental results show that the proposed wavelet-based algorithm consistently outperformed others with a higher detection rate, lower false alarm rate, and shorter mean time to detection. It is conclusive that the proposed algorithm is a superior alternative to existing algorithms.