Modeling of road geometry and traffic accidents by hierarchical object-based and deep learning methods using laser scanning data

Road traffic accidents are global concerns since they affect human life, economy, and road transportation systems. Rapid information acquisition and insight discovery are key tasks in transportation management. Specifically, extraction of geometric road features such as slopes and superelevation are essential information to understand the effects of road geometry on road traffic accidents. However, to understand these effects clearly and accurately, proper modeling techniques should be used. This study aims to develop methods to extract geometric road features (e.g., vertical gradients, superelevation, width, design speed) and establish associations between those features and road traffic accidents including frequency and accident severity. There was a need for efficient segmentation algorithm, optimization strategy, feature extraction and classification, and robust statistical and computational intelligence models to accomplish the set aims. Experimental results regarding road geometry extraction indicated that the proposed methods could achieve relatively high accuracy (~ 85% - User’s Accuracy) of road detection from airborne laser scanning data. Our method improved the overall accuracy of classification by 7% outperforming the supervised ƙ nearest neighbor method. In addition, the results also showed that the proposed hierarchical classification method could extract geometric road elements with an average error rate of 6.25% for slope parameter and 6.65% for superelevation parameter, and it is transferable to other regions of similar environments. On the other hand, the geometric regression model predicted the number of accidents in the North- South Expressway with a reasonable accuracy (R2 = 0.64). This model also could identify the most influential factors contributing to the number of accidents. Experiments on deep learning models showed that the recurrent neural network performs better than the feed forward neural networks, statistical bayesian logistic regression, and convolutional neural networks. This study also suggests that transfer learning could improve the forecasting accuracy of the injury severity by nearly 10%