Incident Detection by Fractal Dimension Analysis of Loop Detector Data

Abstract

This paper describes a research project which aimed to demonstrate the feasibility of using Fractal Dimension analysis of speed, occupancy and flow data for automatic incident detection (AID).

Non-recurrent congestion resulting from accidents, breakdowns and other incidents accounts for about 60% of the delays on freeways (Dia and ROSE, 1997). Therefore, the sooner an appropriate incident response is implemented, the less impact the incident will have on road user safety, congestion and the environment.

Various models have been developed for AID from a variety of theoretical backgrounds and data sources. However, most of these models have limitations, namely high false alarm rates or difficulties with portability and configuration. Artificial neural networks have had the most success, with low false alarm rates and relatively easy configuration.

The use of fractal dimension analysis is becoming widespread. Experts in fields as diverse as Medicine (Hara et al, 1995), Physics (Mouradian and Soruescaut, 1991), Seismology (Tosi et al, 1999), Economics (Richards, 2000), Meteorology (Suresh et al, 1999) and Ecology (Wigley et al, 1999) are using fractal dimension analysis to quantify various phenomena. Fractal analysis has been used to model traffic flow (Torok and Kertesz, 1996), but does not appear to have been used for incident detection.

Two fractal models were developed and tested on a data set of 100 incidents collected by VicRoads for the development of artificial neural network incident detection models (Dia and Rose, 1997). A similar methodology to that presented by Dia and Rose (1997) was used in this project so that the results of the fractal models could be compared with those of the ARRB/VicRoads and the Artificial Neural Network Models.