Automatic incident detection (AID) is crucial for reducing non-recurrent congestion caused by traffic incidents. In this paper, a data-driven AID framework is proposed that can leverage large-scale historical traffic data along with the inherent topology of the traffic networks to obtain robust traffic patterns. Such traffic patterns can be compared with the real-time traffic data to detect traffic incidents in the road network. Our AID framework consists of two basic steps for traffic pattern estimation. First, we estimate a robust univariate speed threshold using historical traffic information from individual sensors. This step can be parallelized using MapReduce framework thereby making it feasible to implement the framework over large networks. Our study shows that such robust thresholds can improve incident detection performance significantly compared to traditional threshold determination. Second, we leverage the knowledge of the topology of the road network to construct threshold heatmaps and perform image denoising to obtain spatio-temporally denoised thresholds. We used two image denoising techniques, bilateral filtering and total variation for this purpose. Our study shows that overall AID performance can be improved significantly using bilateral filter denoising compared to the noisy thresholds or thresholds obtained using total variation denoising.

The second research objective involved detecting traffic congestion from camera images. Two modern deep learning techniques, the traditional deep convolutional neural network (DCNN) and you only look once (YOLO) models, were used to detect traffic congestion from camera images. A shallow model, support vector machine (SVM) was also used for comparison and to determine the improvements that might be obtained using costly GPU techniques. The YOLO model achieved the highest accuracy of 91.2%, followed by the DCNN model with an accuracy of 90.2%; 85% of images were correctly classified by the SVM model. Congestion regions located far away from the camera, single-lane blockages, and glare issues were found to affect the accuracy of the models. Sensitivity analysis showed that all of the algorithms were found to perform well in daytime conditions, but nighttime conditions were found to affect the accuracy of the vision system. However, for all conditions, the areas under the curve (AUCs) were found to be greater than 0.9 for the deep models. This result shows that the models performed well in challenging conditions as well.

The third and final part of this study aimed at detecting traffic incidents from CCTV videos. We approached the incident detection problem using trajectory-based approach for non-congested conditions and pixel-based approach for congested conditions. Typically, incident detection from cameras has been approached using either supervised or unsupervised algorithms. A major hindrance in the application of supervised techniques for incident detection is the lack of a sufficient number of incident videos and the labor-intensive, costly annotation tasks involved in the preparation of a labeled dataset. In this study, we approached the incident detection problem using semi-supervised techniques. Maximum likelihood estimation-based contrastive pessimistic likelihood estimation (CPLE) was used for trajectory classification and identification of incident trajectories. Vehicle detection was performed using state-of-the-art deep learning-based YOLOv3, and simple online real-time tracking (SORT) was used for tracking. Results showed that CPLE-based trajectory classification outperformed the traditional semi-supervised techniques (self learning and label spreading) and its supervised counterpart by a significant margin. For pixel-based incident detection, we used a novel Histogram of Optical Flow Magnitude (HOFM) feature descriptor to detect incident vehicles using SVM classifier based on all vehicles detected by YOLOv3 object detector. We show in this study that this approach can handle both congested and non-congested conditions. However, trajectory-based approach works considerably faster (45 fps compared to 1.4 fps) and also achieves better accuracy compared to pixel-based approach for non-congested conditions. Therefore, for optimal resource usage, trajectory-based approach can be used for non-congested traffic conditions while for congested conditions, pixel-based approach can be used.