Department of Statistics, UCLA

Real-time object detection is essential for many computer vision applications. Many rapid detection algorithms are based on using cascades of tests. But existing design criteria for cascades either ignore the time complexity of the tests or make over-simplified assumptions about them. This paper gives a criterion for designing a time-efficient cascade that explicitly takes into account the time complexity of tests (as evaluated by computer run time) including the time for pre-processing. We design a greedy algorithm to minimize this criterion (noting that the full problem is NP-complete). Finally, we illustrate our method on the task of text detection in city scenes. This gives a text detection algorithm that runs at 0.025 seconds per 320×240 image, which is equivalent to 40 frames per second. This is a speed up factor of 2.5 compared to our previous text detector. It gives a realtime system which can be used for applications to help the blind and visually impaired.