engleski

Recent advances in traffic sign detection

Traffic sign detection from the driver's perspective has attracted significant and sustained interest both in academic and industrial communities in the last decade. State of the art approaches in the field achieve high detection recalls by applying machine learned binary classifiers at all image locations and scales in a sliding detection window approach. However, experiments have shown that high precision (low false positive rate) and accurate localizations of the detected signs can not be obtained by straightforward approaches. Most of this presentation shall be devoted to recent developments by which we have alleviated if not overcome the two problems above. The next major challenge is making a transition towards simultaneous dealing with heterogeneous traffic sign classes, which calls for efficient multi-class detection approaches. In the second, shorter part, I will present key issues that must be addressed in order to achieve that goal, and show some preliminary results along these lines.

computer vision; object detection

This talk presented the research on traffic sign detection carried out in the frame of the scientific project Mapping and Assessing the State of Traffic Infrastructure. URL http://www.zemris.fer.hr/~ssegvic/mastif/index_en.shtml