About the research
State departments of transportation (DOTs) typically install a relatively large number of cameras across freeways for surveillance purposes. It is estimated that there will be approximately a billion of these traffic cameras worldwide by 2020. However, most of these cameras are used for manual surveillance purposes only. Hence, there is a significant need to develop automatic anomaly detection algorithms that use the data from these cameras.
This study was divided into two broad topics involving the detection of freeway traffic anomalies from cameras: detecting traffic congestion from camera images and detecting traffic incidents from camera videos. Two modern deep learning techniques, the traditional deep convolutional neural network (DCNN) and the you only look once (YOLO) models, were used to detect traffic congestion from camera images and compared with a shallow model, support vector machine (SVM).
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. The deep models were found to perform well in challenging conditions too, such as nighttime or poor weather conditions.
To detect traffic incidents from camera videos, the research team proposed a semi-supervised approach for detecting incident trajectories. Results showed that the proposed semi-supervised trajectory classification outperformed the traditional semi-supervised techniques and its supervised counterpart by a significant margin.