Nonintrusive video analytics for traffic flow observation are a flexible alternative to human operators and conventional inductive loop detectors. Ang, Shen, and Duraisamy present an automated low-cost dual-camera system for monitoring traffic flow at a “T” intersection.
In their method, a form of background subtraction segments foreground objects from the background, and statistical methods compare feature distribution histograms of blobs in the image in order to accomplish object detection. The central--and, unfortunately, weakest--aspect of the method is the trajectory estimation tables that classify vehicle events. The very structured rule-based approach is rather brittle; in more complex situations, it is unlikely that it will provide a general framework for reasoning about trajectories. Furthermore, because the cameras do not have an overlapping field of view, the system relies on assumptions regarding the expected timing of events, which weakens and decreases the flexibility of its overall design. Although this condition seems to be a necessary constraint, it will inevitably hamper the extension of the design to handle more complex traffic flow, such as incident detection. Nevertheless, the approach works in real-time and adapts well to different types of vehicles.
This well-written paper concisely presents the relevant issues pertaining to tracking vehicles, and effectively summarizes the current state of the art. The authors show a convincing level of familiarity with their field. Despite the design’s shortcomings, the paper reads well and is a valuable introduction to traffic surveillance.