**8. Conclusions**

18 Will-be-set-by-IN-TECH

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11

corner.

0 1 2 3 4 5 6 7 8 9 10

LRD

0 1 2 3 4 5 6 7 8 9 10

 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

computed errors (see text) of WaldBoost and EnMS.

error

orig 0.01 0.05

orig 0.01 0.05

err "=X" EnMS err ">X-2" EnMS err ">X-6" EnMS err ">2" EnMS err "=X" WB err ">X-2" WB err ">X-6" WB err ">2" WB

0.20 .015

0.20 0.02

Fig. 8. The graphs show AUC on y-axis (Area Under ROC) versus the average number of weak classifiers evaluated per image position as measured on the MIT+CMU frontal face dataset. The individual lines are for original WaldBoost detectors without neighborhood suppression (full line) and the other lines are with added neighborhood suppression with different target false negative rates. Good results should be in the left (fast) bottom (accurate)

0.02 .025 0.03 .035 0.04 .045 0.05 .055 0.06

.025 0.03 .035 0.04 .045 0.05 .055 0.06 .065 0.07 .075

0 50 100 150 200 250 300 350

speed-up

Fig. 9. The graphs show frontal face localization error (y-axis) for different speed-ups achieved by WaldBoost and EnMS. The speed-up is measured as reduction of the number of weak hypotheses evaluated on average per image position relative to the full length of the classifier (length is the same for WaldBoost and EnMS). The lines represent differently

0 2 4 6 8 10 12 14

LRP

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

orig 0.01 0.05 0.20

orig 0.01 0.05 0.20

LBP

Haar

This chapter focused on methods of real-time object detection with classifiers. It has been demonstrated that the object detection methods working in real-time are feasible and can be implemented on a variety of platforms, such as personal computer processors, GP-GPU platforms, or even in programmable hardware.

In order to achieve real-time performance, an efficient implementation platform and efficient implementation itself is necessary, but further enhancement through algorithmic acceleration is needed as well. Two examples of such acceleration are presented in the chapter: exploitation of information about neighborhoods of the already classified positions in the image and early suppression of non-maxima of the classifier responses. The approach of exploitation of the neighborhoods in the image is based on the idea that classification of the overlapping sub-images in the image - the neighborhoods - may share some properties and information. One of the possible ways to share such information is through re-using the weak classifiers used during classification of one location through WaldBoost for predicting results in the other neighboring locations. This prediction is done through a machine learning process similar to WaldBoost where the difference to WaldBoost is that the training process actually reuses the already selected weak classifiers that were used at the original location. While this process works well only in close neighborhoods, it brings a significant speed-up.

Pre-processing that rules out some parts of the image from the detection process can significantly speed up the detection process. Important future research certainly includes machine-learning based pre-processing methods and research of under-sampling in scanning methods that can also improve detection performance possibly without any adverse effects on precision. Future research also includes algorithmic improvements of acceleration methods, such as improvement in the processor assignment in GP-GPU, improved scanning trajectories in neighborhood exploitation, or further improvements in feature extraction.
