The research of the urban accident severity with Artificial Neural Networks

• Main Authors: Young-Chien Chou, 周雍傑
• Other Authors: Tzuoo-Ding Lin
• Format: Others
• Language: zh-TW
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/62258881992449685150

碩士 === 國立成功大學 === 交通管理學系 === 88 === The serious traffic accident in downtown areas of the city results from the rapid growth of passenger cars and motorcycles, it also causes huge social cost expenditure. The negative consequences have been concerned by government officials and transportation researchers. Most previous studies always focused on the relationship between cause and effect of the accident frequencies and the following factors including road design and traffic flow to discuss the severity of accident based on the aggregate data. In this study, the disaggregate data is used to explore the severity of the accidents. The injury severity of the driver can be divided into three levels: property damage only, injury, and death. The Artificial Neural Network (ANN) model is adapted to forecast the severity of the accidents. And the sensitivity analysis is used to evaluate the influence of factors on accident severity. The performance of ANN model was affected by the elements including "training & testing set", "input variable", "hidden layer", "learning rule", "transfer function", and "momentum factor" etc. These elements were considered to develop the best ANN model. The variables were selected by the contribution graph approach, and the multiple neurons and single neuron ANN models were developed to compare the similarities & differences. This research shows that the accident severity predicting correct rate can achieve 94.44%(death 50%, injury 76.56%, and property damage only 91.82%) for multiple neurons and 96.26%(death 50%, injury 89.21%, and property damage only 98.31%) for single neuron ANN model at intersection. Meanwhile, the accident severity predicting correct rate can achieve 87.55% (death 50%, injury 74.44%, and property damage only 90.7%) for multiple neurons and 98.43% (death 50%, injury 92.59%, and property damage only 100%) for single neuron ANN model at road section. The sensitivity analysis shows that, both at intersection and road section, the proportion of death on cloudy and rainy days is higher than that on sunny days, but the proportion of property damage only on cloudy and rainy days is lower; The proportion of death in dry road surface is lower than that in others (e.g. wet) road surface, but the proportion of property damage only in dry road surface is lower. The proportion of death at intersection with roadblock (e.g. road construction, illegal parking) is higher than that at intersection without roadblock.