Zitat:14.1.2 Pedestrian detection
While a lot of the research on object detection has focused on faces, the detection of other
objects, such as pedestrians and cars, has also received widespread attention (Gavrila and
Philomin 1999; Gavrila 1999; Papageorgiou and Poggio 2000; Mohan, Papageorgiou, and
Poggio 2001; Schneiderman and Kanade 2004). Some of these techniques maintain the same
focus as face detection on speed and efficiency. Others, however, focus instead on accuracy,
viewing detection as a more challenging variant of generic class recognition (Section 14.4)
in which the locations and extents of objects are to be determined as accurately as possible.
(See, for example, the PASCAL VOC detection challenge, http://pascallin.ecs.soton.ac.uk/
challenges/VOC/.)
An example of a well-known pedestrian detector is the algorithm developed by Dalal
and Triggs (2005), who use a set of overlapping histogram of oriented gradients (HOG) descriptors
fed into a support vector machine (Figure 14.8). Each HOG has cells to accumulate
magnitude-weighted votes for gradients at particular orientations, just as in the scale invariant
feature transform (SIFT) developed by Lowe (2004), which we discussed in Section 4.1.2 and
Figure 4.18. Unlike SIFT, however, which is only evaluated at interest point locations, HOGs
are evaluated on a regular overlapping grid and their descriptor magnitudes are normalized
using an even coarser grid; they are only computed at a single scale and a fixed orientation. In
order to capture the subtle variations in orientation around a person’s outline, a large number
of orientation bins is used and no smoothing is performed in the central difference gradient
computation—see the work of Dalal and Triggs (2005) for more implementation details.

Figure 14.8d shows a sample input image, while Figure 14.8e shows the associated HOG
descriptors.
Once the descriptors have been computed, a support vector machine (SVM) is trained
on the resulting high-dimensional continuous descriptor vectors. Figures 14.8b–c show a
diagram of the (most) positive and negative SVM weights in each block, while Figures 14.8f–
g show the corresponding weighted HOG responses for the central input image. As you can
see, there are a fair number of positive responses around the head, torso, and feet of the
person, and relatively few negative responses (mainly around the middle and the neck of the
sweater).
The fields of pedestrian and general object detection have continued to evolve rapidly
over the last decade (Belongie, Malik, and Puzicha 2002; Mikolajczyk, Schmid, and Zisserman
2004; Leibe, Seemann, and Schiele 2005; Opelt, Pinz, and Zisserman 2006; Torralba
2007; Andriluka, Roth, and Schiele 2009, 2010; Dollar, Belongie, and Perona 2010 ` ).
Munder and Gavrila (2006) compare a number of pedestrian detectors and conclude that
those based on local receptive fields and SVMs perform the best, with a boosting-based approach
coming close. Maji, Berg, and Malik (2008) improve on the best of these results using
non-overlapping multi-resolution HOG descriptors and a histogram intersection kernel SVM
based on a spatial pyramid match kernel from Lazebnik, Schmid, and Ponce (2006).
When detectors for several different classes are being constructed simultaneously, Torralba,
Murphy, and Freeman (2007) show that sharing features and weak learners between
detectors yields better performance, both in terms of faster computation times and fewer
training examples. To find the features and decision stumps that work best in a shared manner,
they introduce a novel joint boosting algorithm that optimizes, at each stage, a summed
expected exponential loss function using the “gentleboost” algorithm of Friedman, Hastie,
and Tibshirani (2000).
In more recent work, Felzenszwalb, McAllester, and Ramanan (2008) extend the histogram
of oriented gradients person detector to incorporate flexible parts models (Section 14.4.2).
Each part is trained and detected on HOGs evaluated at two pyramid levels below the overall
object model and the locations of the parts relative to the parent node (the overall bounding
box) are also learned and used during recognition (Figure 14.9b). To compensate for inaccuracies
or inconsistencies in the training example bounding boxes (dashed white lines in
Figure 14.9c), the “true” location of the parent (blue) bounding box is considered a latent
(hidden) variable and is inferred during both training and recognition. Since the locations
of the parts are also latent, the system can be trained in a semi-supervised fashion, without
needing part labels in the training data. An extension to this system (Felzenszwalb, Girshick,
McAllester et al. 2010), which includes among its improvements a simple contextual model,
was among the two best object detection systems in the 2008 Visual Object Classes detection
challenge. Other recent improvements to part-based person detection and pose estimation in-
668 Computer Vision: Algorithms and Applications (September 3, 2010 draft)
(a) (b) © (d)
Figure 14.9 Part-based object detection (Felzenszwalb, McAllester, and Ramanan 2008)
c 2008 IEEE: (a) An input photograph and its associated person (blue) and part (yellow)
detection results. (b) The detection model is defined by a coarse template, several higher
resolution part templates, and a spatial model for the location of each part. © True positive
detection of a skier and (d) false positive detection of a cow (labeled as a person).
clude the work by Andriluka, Roth, and Schiele (2009) and Kumar, Zisserman, and H.S.Torr
(2009).
An even more accurate estimate of a person’s pose and location is presented by Rogez,
Rihan, Ramalingam et al. (2008), who compute both the phase of a person in a walk cycle and
the locations of individual joints, using random forests built on top of HOGs (Figure 14.11).
Since their system produces full 3D pose information, it is closer in its application domain to
3D person trackers (Sidenbladh, Black, and Fleet 2000; Andriluka, Roth, and Schiele 2010),
which we discussed in Section 12.6.4.
One final note on person and object detection. When video sequences are available, the
additional information present in the optic flow and motion discontinuities can greatly aid in
the detection task, as discussed by Efros, Berg, Mori et al. (2003), Viola, Jones, and Snow
(2003), and Dalal, Triggs, and Schmid (2006).