Integrated Pedestrian Detection and Localization Using Stereo Cameras - Digital Signal Processing for In-Vehicle Systems and Safety - page 222

Fig. 16.5 PR curve for the detection performance
detected in practice. For a fair comparison, we also take top ten ranked hypotheses
as system output. The comparison of the three systems' performance on the 133
stereo pairs is plotted in Fig. 16.5 .
In most cases, the detector with deformable-part-based model has maintained a
precision near 0.5. By integrating depth information, our proposed system
outperforms the baseline detector significantly and closes to the best one.
We also compute an average precision for the three methods to show the overall
performance. The results are 0.2325 (our method), 0.1738 (baseline), and 0.2530
(UoCTTI), respectively.
Without any optimization of speed, on a 2.83-G Intel Core 2 Quad CPU with 4 G
RAM, the average speed of the three methods are 1.73 s (our method), 1.7 s (baseline),
and 8.4 s (UoCTTI) on a single 640 by 480 image. The UoCTTI detector is quite time-
consuming. It uses nearly five timesmore than our baseline detector. Since theUoCTTI
detector also uses HOG as low-level feature set, its disadvantage in runtime may
mainly boil down to the complicated model it uses. Therefore, even if it is powerful,
it could not be used in some applications before the runtime issue could be resolved.
By carefully selecting the efficient cues, our integration system could also be
very fast. Though this runtime performance is not good enough for some
applications, it still has room for improvement. Currently, in our system, the most
time-consuming part is the HOG feature pyramid computation and sliding window
searching. Since these two kinds of processing can be done much more faster by
using GPU programming, our integration system still have the potential to be used
in real-time applications.