Incremental detection of text on road signs

This paper presents a framework for incremental detection of text from road signs. The approach efficiently incorporates tracking and detection mechanisms into the same framework. The proposed approach first finds a set of discriminative feature points and clusters them into different regions. We then select candidate sign planes by a combination of color and vertical plane models. Within detected road sign planes, the framework selects candidate text regions again base on feature points. The feature points serve a dual purpose: correspondence for tracking if text has been detected in the region and cues of candidate regions for text detection. The framework further verifies candidate text regions using more sophisticated features. Once a text region is confirmed, the tracking algorithm will continuously track the region. The text region grows as more text around it is detected from frame to frame. Experimental results have demonstrated the feasibility of the proposed framework in incrementally detecting text on road signs over the time from video sequences captured from a moving vehicle. This research was partially supported by the CMU/CM-CRL.