Comparative appraisal of adaptive ANN-based freeway incident detection models

Abstract

Ahslroct A number o/arl$cial newal network (2 ")-based incidenl dclcclion models hove been tested independenlly over IIrr post decade. This paper aim lo evalrmle the incidenl derectlon capabilities of Ihree pronfismg ANN-bawd delectlon models. These rnodels were developed on an original freeway sile ln Singapore ond fhcn adapted lo a tiew freeway sile in Cal!fornla Aparl/ron h i r inclden1 deleclion peflonitances, their adapfaflon slmfegies and network sizes have also been conpared Resulls of this study show that allhough sulli-layer /ced-jonUard (MLF) models I r m lhe best lacident defection pejornrance on the development sile. conslniclivc probabillslic neural network (CPh'N) niodcls are most adaplable and epclent. Moreover, CPNN ,nodel lrns a rnsclr srnaller network size, making 11 enslrr lo ln,plen~enl il /or real-lime applicofion. The reslrlts ruggrsr lhal CPNN nwdel has lk highesr polenlid for use in an operattonal automatic incident detection splemJ0r freeways. hi frccwny traffic monitoring and conlml, an itnportnttl ndivily is lhc detection and VCrificSliM of incidcnts. In tcchnicnl terms, incidents ore defined ns randoin a i d non-recurring events such as accidents. disablcd vehicles, spilled loads, temporary maintenance and construction acfivitics, and othcr unusual events that dismnt lhc oomnl flow of traffm A-urate and earlv Artificial neural nehvorks have been widely applied to numerous pattern recognition problems including freeway incident detection A review of existing literature shows that neural network-bnscd incident detection models were developed mainly based on multi-layer feed-forward n e u d network (MLF) and adaptive resonance theory (ART). These works included tho pioncering work using MLF [3.4], 11s well as other works using MLF and Fuzzy' Logic [7,8.9]. Fmm these works, it is evidmt that MLF has definite advantages over conventional incident detection teohniqucs in providing a high detection rats and a low fdse alarm rate. However, the MLF is limited by its model adaptation capability. Once trained, the MLF is unable to give similarly goad performance when presented with data fmm another site that has different traffic patterns. This hns motivated sevsral recent studies to explore new neural network models that have higher adaptation capability, while at the W I