Temporospatial Context-Aware Vehicular Crash Risk Prediction
MetadataShow full item record
With the demand for more vehicles increasing, road safety is becoming a growing concern. Traffic collisions take many lives and cost billions of dollars in losses. This explains the growing interest of governments, academic institutions and companies in road safety. The vastness and availability of road accident data has provided new opportunities for gaining a better understanding of accident risk factors and for developing more effective accident prediction and prevention regimes. Much of the empirical research on road safety and accident analysis utilizes statistical models which capture limited aspects of crashes. On the other hand, data mining has recently gained interest as a reliable approach for investigating road-accident data and for providing predictive insights. While some risk factors contribute more frequently in the occurrence of a road accident, the importance of driver behavior, temporospatial factors, and real-time traffic dynamics have been underestimated. This study proposes a framework for predicting crash risk based on historical accident data. The proposed framework incorporates machine learning and data analytics techniques to identify driving patterns and other risk factors associated with potential vehicle crashes. These techniques include clustering, association rule mining, information fusion, and Bayesian networks. Swarm intelligence based association rule mining is employed to uncover the underlying relationships and dependencies in collision databases. Data segmentation methods are employed to eliminate the effect of dependent variables. Extracted rules can be used along with real-time mobility to predict crashes and their severity in real-time. The national collision database of Canada (NCDB) is used in this research to generate association rules with crash risk oriented subsequents, and to compare the performance of the swarm intelligence based approach with that of other association rule miners. Many industry-demanding datasets, including road-accident datasets, are deficient in descriptive factors. This is a significant barrier for uncovering meaningful risk factor relationships. To resolve this issue, this study proposes a knwoledgebase approximation framework to enhance the crash risk analysis by integrating pieces of evidence discovered from disparate datasets capturing different aspects of mobility. Dempster-Shafer theory is utilized as a key element of this knowledgebase approximation. This method can integrate association rules with acceptable accuracy under certain circumstances that are discussed in this thesis. The proposed framework is tested on the lymphography dataset and the road-accident database of the Great Britain. The derived insights are then used as the basis for constructing a Bayesian network that can estimate crash likelihood and risk levels so as to warn drivers and prevent accidents in real-time. This Bayesian network approach offers a way to implement a naturalistic driving analysis process for predicting traffic collision risk based on the findings from the data-driven model. A traffic incident detection and localization method is also proposed as a component of the risk analysis model. Detecting and localizing traffic incidents enables timely response to accidents and facilitates effective and efficient traffic flow management. The results obtained from the experimental work conducted on this component is indicative of the capability of our Dempster-Shafer data-fusion-based incident detection method in overcoming the challenges arising from erroneous and noisy sensor readings.
Cite this version of the work
Pouya Mehrannia (2020). Temporospatial Context-Aware Vehicular Crash Risk Prediction. UWSpace. http://hdl.handle.net/10012/15954