Efficient Traffic Sign Detection Using Bag of Visual Words and Multi-scales SIFT

摘要

Automatic traffic sign detection is important in many applications such as GPS based navigation systems, advanced driver assistance systems, and self-driving cars. Recently, several researches have shown that bag of visual words (BoVW) method is really an interesting and potential choice for this detection problem. However, it is difficult for using this approach in practice due to the high computational cost. To find the exact boundaries of objects, this approach has to scan a large number of image sub-windows over location and scale (e.g. there are approximately 60,000 32×32 pixels sub-windows for an 320×240 pixels image). In this paper, we propose an efficient approach, which use multi-scales SIFT features and coarse-to-fine search strategy, to improve speed of BoVW. We argue that multi-scales SIFT features can be used for quickly detecting the coarse boundaries of objects. Then, the further searching stage only need to concentrate on these discovered boundaries. By this way, the number of image sub-windows is efficiently reduced. The experimental results show that our proposed method significantly improves detection speed without trading off performance.