Application of the Cascades

Several experiments were made to evaluate the performance of the proposed approach with two different kinds of images, namely, reflectance and depth images. Both types are acquired by the 3D laser range finder and are practically light invariant. About 200 representation of the objects were taken in addition to a wide variety of negative examples without any target object. The detection starts with the smallest classifier size, e.g., $16 \times 40$ pixel for the human classifier, $23 \times 30$ for the volksbot classifier. The image is searched from top left to bottom right by applications of the cascade. To detect objects on larger scales, the detector is rescaled. An advantage of the Haar-like features is that they are easily scalable. Each feature requires only a fixed number of look-ups in the integral image, independent of the scale. Time-consuming picture scales are not necessary to achieve scale invariance. Fig. show examples of the detection.

To decrease the false detection rate, we combine the cascades of the depth and reflectance images. There are two possible ways for combining: Either the two cascades run interleaved or serial and represent a logical ``and'' [13]. The joint cascade decreases the false detection rate close to zero. To avoid the reduction of the hit rate, several different off-screen rendered images are used, where the virtual camera is rotated and the apex angle is changed [13].