Video object detection and matching

The automatic analysis of video has come a long way from using global features to detection and matching of objects. We seek to understand two common classes of objects—text and vehicles. In unconstrained video, text occurs in a variety of ways including artificially embedded captions and those that naturally occur in the scene. We present a novel algorithm for detecting text in video frames. The constraints are that the text object is horizontally-aligned and the stroke pixels have uniform color. The color features derived by the detection algorithm are then used in a novel algorithm for finding similar text objects between two video frames. We believe that finding similar text objects is useful for finding related video clips and finding objects and scenes with similar text marks. Quantitative performance evaluation is important for optimizing, evaluating and comparing object detection algorithms. We present a set of six measures for quantifying different aspects of performance. Using these measures, the text detection algorithm was evaluated and compared to another text detection algorithm. Recognition of vehicles using visual features is useful for surveillance and traffic studies. We modified our text-detection method to find text marks on moving vehicles. We also explored a strategy for recognizing vehicles that repeatedly pass through a scene, utilizing feature clustering and a multi-object alignment framework.