Integrated Pedestrian Detection and Localization Using Stereo Cameras - Digital Signal Processing for In-Vehicle Systems and Safety - page 215

Fig. 16.1 Select candidates strictly ( continuous line bounding boxes ); use looser criterion, more
candidates were found ( dashed line bounding boxes )
models will lead to better accuracy. However, improvement in this approach
sometimes comes with additional processing time which usually slows down the
entire system's speed [ 1 ].
In most real-world applications, speed and accuracy are crucial issues and
should be addressed simultaneously. Of course, time-consuming methods are not
recommended. At the same time, the simplest and fastest methods are not robust
enough by themselves. An example is illustrated in Fig. 16.1 . We apply a very
simple pedestrian detector described in [ 2 ] on the street view image. When
selecting the candidates using strict standards, as shown with continuous line
bounding boxes, many true occurrences for pedestrians were missed. As we make
the selection standard a little looser, some missed true occurrences were success-
fully found. But the second approach has a drawback. The false number increased.
This means that a detector using a simple feature and coarse model is not, by itself,
discriminative enough. The inadequacy, however, could be compensated to some
extent by using other cues from the image and background knowledge.
Several studies have tried to use other cues for pedestrian detection. Leibe et al. [ 3 ]
proposed the use of scene geometry to improve object detection. By assuming that
pedestrians can only be possibly supported by the ground plane, some false detection
results could be filtered out. In another work, Gavrila and Munder [ 4 ]presenteda
system which involves a cascade of modules wherein each unit utilizes complemen-
tary visual criteria to narrow down the image searching space. These two were both
excellent works; however, additional cues are mainly used to get rid of false results but
unable to support a true one.