An Investigation into Segmenting Traffic Images Using Various Types of Graph Cuts

In computer vision, graph cuts are a way of segmenting an image into multiple areas. Graphs are built using one node for each pixel in the image combined with two extra nodes, known as the source and the sink. Each node is connected to several other nodes using edges, and each edge has a specific weight. Using different weighting schemes, different segmentations can be performed based on the properties used to create the weights. The cuts themselves are performed using an implementation of a solution to the maximum flow problem, which is then changed into a minimum cut according to the max-flow/min-cut theorem. In this thesis, several types of graph cuts are investigated with the intent to use one of them to segment traffic images. Each of these variations of graph cut is explained in detail and compared to the others. Then, one is chosen to be used to detect traffic. Several weighting schemes based on grayscale value differences, pixel variances, and mean pixel values from the test footage are presented to allow for the segmentation of video footage into vehicles and backgrounds using graph cuts. Our method of segmenting traffic images via graph cuts is then tested on several videos of traffic in various lighting conditions and locations. Finally, we compare our proposed method to a similarly performing method: background subtraction.