Loading...
Detection of Traffic Incidents Using Machine Learning Techniques
ElSahly, Osama Mohamed
ElSahly, Osama Mohamed
Date
2023-04
Authors
Advisor
Type
Dissertation
Degree
Citations
Altmetric:
Description
A Doctor of Philosophy Dissertation in Engineering Systems Management by Osama Mohamed ElSahly entitled, “Detection of Traffic Incidents Using Machine Learning Techniques”, submitted in April 2023. Dissertation advisor is Dr. Akmal Abdelfatah. Soft copy is available (Dissertation, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Abstract
This dissertation proposes new models for detecting traffic incidents on freeways using machine learning algorithms to classify traffic data collected from the freeway system. These models are generic and consider multiple factors that affect incident detectability simultaneously. The models were trained and tested on simulated traffic data that represent normal and incident conditions using the well-known microsimulation software VISSIM. The proposed models, which include the Random Forest (RF) and Multilayer Feedforward Artificial Neural Network (MLF), consider four factors: the congestion level, the distance between the upstream and downstream detector stations, the location of the incident relative to the detector stations, and the severity of the incident. The results showed that the developed models achieved excellent performance, surpassing existing models in the literature. During training, the MLF model achieved a detection rate (DR) of 95.96%, a mean time to detect (MTTD) of 0.89 minutes, and a false alarm rate (FAR) of 1.01%. During testing, the MLF model achieved a DR of 100%, MTTD of 1.6 minutes, and FAR of 1.29%. Similarly, the RF model achieved a DR of 96.97%, MTTD of 1.05 minutes, and FAR of 0.62% during training, and a DR of 100%, MTTD of 1.17 minutes, and FAR of 0.862% during testing. The results revealed that incident detection systems may have difficulty detecting incidents with minor severity during low traffic volumes. The FAR decreased with the increase in the Demand to Capacity ratio (D/C), while the MTTD increased with the increase in D/C. Additionally, higher incident severity resulted in lower MTTD values, while the distance between the incident location and upstream detector had the opposite effect. The FAR decreased as the incident moved farther from the upstream detector but increased with the distance between detectors. Larger detector spacings were associated with longer detection times. The proposed models can significantly improve traffic performance on freeways, especially during incidents, and benefit both local and international transportation agencies. The study's contribution lies in developing efficient and reliable incident detection models that consider multiple variables simultaneously, improving traffic safety, reducing congestion, saving lives and properties, and reducing pollution.