| Summary: | 碩士 === 逢甲大學 === 交通工程與管理所 === 93 === Abstract
When a traffic accident occurs, the arguments on accident responsibility also appeared. In order to clarify the responsibility and settle the arguments, accident appraisal committees are authorized to review these cases and determine the responsibility for causing the accident. However, comparing with over 15,000 accident appraisal cases per year in Taiwan, there are rather few experts in reviewing these cases. Therefore, in order to ease the burden of these experts, an efficiency expert system for accident authentication is worthy of developing. Besides, because highly professional knowledge and experience accumulated from long-term training is needed for conducting accident appraisal, it is also important for the system to effectively educate junior reviewers with these knowledge and experience. Basing on that, Chiou and Fang (2004) have already developed an expert system for appraisal of two-car crash accidents. This study aims to develop an expert system for the appraisal of motorcycle and car crash accidents, which are the second largest appraisal cases next to two-car accidents, by employing artificial neural network.
A total of 450 motorcycle and car crash accident appraisal cases from 2000 to 2002 are selected by excluding the cases with inconsistent appraisal results between local committee and reviewing committee. Since the right-of-way is most important variable in accident appraisal, this study employs decision tree technique to conclude the judgment of right-of-way according to different crash types. The input variables for ANN are selected by using contingent table from all variables, such as speeding, drinking alcoholic, right of the way. The output variable is set as the degree of responsibility that the party involved is assessed to take. These cases are randomly divided into to two sets: 70% for training and 30% for validating. Then, different network structures and settings of parameters are tested and analyzed for proposing a comprising model. For the sake of comparison, a statistical discriminating analysis model for accident appraisal is also developed and calibrated.
The results show that the ANN model can achieve 88.1% and 76.3% of correctness rate in training and validating, respectively. The correctness rates of discrimination analysis model are only 54.76% in training and 52.59% in validating. Obviously, it indicates that the ANN model is more suitable to be the expert system for accident appraisal. Moreover, in order to measure the influence of each input variable on judging the accident responsibility, an index named as general influence index (GI) is calculated by the ANN trained weights. The most influential variable is related direction, with GI=0.245, followed by movement (GI=0.178), right of way (GI=0.175), damage spot (GI=0.158), speeding (GI=0.152). It is also in accordance with the prior knowledge in accident appraisal.
|
|---|
