Vehicle Queue Length Estimation Using Uncalibrated Cameras

Vehicle queue lengths at city road intersections are important parameters in designing intelligent transport systems, and are particularly useful in traffic modeling and in designing smart signals. In this paper, we propose a simple and scalable method for estimation of vehicle queue lengths that works in real time. The proposed solution is based on cameras (with unknown intrinsics) that are commonly found on most city roads, and does not require the installation of any sensors. The approach consists of three stages. In the first stage, we automatically identify the road region in the camera view. This is achieved through unsupervised segmentation that uses a convolutional neural network. In the second stage, we identify the presence of the queue by using low level features such as corners. In the third stage, we estimate the physical length (in metres) of the queue. Our experiments show that the proposed method is effective, with less than 10% error in the length estimates.