\subsection{Haar Feature-based Cascade Classifier for Object Detection}
The object detector described below has been initially proposed by Paul Viola
\cvCPyCross{Viola01}
and improved by Rainer Lienhart
\cvCPyCross{Lienhart02}
. First, a classifier (namely a \emph{cascade of boosted classifiers working with haar-like features}) is trained with a few hundred sample views of a particular object (i.e., a face or a car), called positive examples, that are scaled to the same size (say, 20x20), and negative examples - arbitrary images of the same size.
After a classifier is trained, it can be applied to a region of interest
(of the same size as used during the training) in an input image. The
classifier outputs a "1" if the region is likely to show the object
(i.e., face/car), and "0" otherwise. To search for the object in the
whole image one can move the search window across the image and check
every location using the classifier. The classifier is designed so that
it can be easily "resized" in order to be able to find the objects of
interest at different sizes, which is more efficient than resizing the
image itself. So, to find an object of an unknown size in the image the
scan procedure should be done several times at different scales.
The word "cascade" in the classifier name means that the resultant
classifier consists of several simpler classifiers (\emph{stages}) that
are applied subsequently to a region of interest until at some stage the
candidate is rejected or all the stages are passed. The word "boosted"
means that the classifiers at every stage of the cascade are complex
themselves and they are built out of basic classifiers using one of four
different \texttt{boosting} techniques (weighted voting). Currently
Discrete Adaboost, Real Adaboost, Gentle Adaboost and Logitboost are
supported. The basic classifiers are decision-tree classifiers with at
least 2 leaves. Haar-like features are the input to the basic classifers,
and are calculated as described below. The current algorithm uses the
The feature used in a particular classifier is specified by its shape (1a, 2b etc.), position within the region of interest and the scale (this scale is not the same as the scale used at the detection stage, though these two scales are multiplied). For example, in the case of the third line feature (2c) the response is calculated as the difference between the sum of image pixels under the rectangle covering the whole feature (including the two white stripes and the black stripe in the middle) and the sum of the image pixels under the black stripe multiplied by 3 in order to compensate for the differences in the size of areas. The sums of pixel values over a rectangular regions are calculated rapidly using integral images (see below and the \cvCPyCross{Integral} description).
\ifPy
A simple demonstration of face detection, which draws a rectangle around each detected face:
\begin{lstlisting}
hc = cv.Load("haarcascade_frontalface_default.xml")
\cvarg{directory}{Name of the directory containing the description of a trained cascade classifier}
\cvarg{orig\_window\_size}{Original size of the objects the cascade has been trained on. Note that it is not stored in the cascade and therefore must be specified separately}
\end{description}
The function loads a trained cascade
of haar classifiers from a file or the classifier database embedded in
OpenCV. The base can be trained using the \texttt{haartraining} application
(see opencv/apps/haartraining for details).
\textbf{The function is obsolete}. Nowadays object detection classifiers are stored in XML or YAML files, rather than in directories. To load a cascade from a file, use the \cvCPyCross{Load} function.
\fi
\cvCPyFunc{HaarDetectObjects}
Detects objects in the image.
\ifC
\begin{lstlisting}
typedef struct CvAvgComp
{
CvRect rect; /* bounding rectangle for the object (average rectangle of a group) */
int neighbors; /* number of neighbor rectangles in the group */
\cvarg{scaleFactor}{The factor by which the search window is scaled between the subsequent scans, 1.1 means increasing window by 10\%}
\cvarg{minNeighbors}{Minimum number (minus 1) of neighbor rectangles that makes up an object. All the groups of a smaller number of rectangles than \texttt{min\_neighbors}-1 are rejected. If \texttt{minNeighbors} is 0, the function does not any grouping at all and returns all the detected candidate rectangles, which may be useful if the user wants to apply a customized grouping procedure}
\cvarg{flags}{Mode of operation. Currently the only flag that may be specified is \texttt{CV\_HAAR\_DO\_CANNY\_PRUNING}. If it is set, the function uses Canny edge detector to reject some image regions that contain too few or too much edges and thus can not contain the searched object. The particular threshold values are tuned for face detection and in this case the pruning speeds up the processing}
\cvarg{minSize}{Minimum window size. By default, it is set to the size of samples the classifier has been trained on ($\sim20\times20$ for face detection)}
\cvarg{maxSize}{Maximum window size to use. By default, it is set to the size of the image.}
The function finds rectangular regions in the given image that are likely to contain objects the cascade has been trained for and returns those regions as a sequence of rectangles. The function scans the image several times at different scales (see \cvCPyCross{SetImagesForHaarClassifierCascade}). Each time it considers overlapping regions in the image and applies the classifiers to the regions using \cvCPyCross{RunHaarClassifierCascade}. It may also apply some heuristics to reduce number of analyzed regions, such as Canny prunning. After it has proceeded and collected the candidate rectangles (regions that passed the classifier cascade), it groups them and returns a sequence of average rectangles for each large enough group. The default parameters (\texttt{scale\_factor} =1.1, \texttt{min\_neighbors} =3, \texttt{flags} =0) are tuned for accurate yet slow object detection. For a faster operation on real video images the settings are: \texttt{scale\_factor} =1.2, \texttt{min\_neighbors} =2, \texttt{flags} =\texttt{CV\_HAAR\_DO\_CANNY\_PRUNING}, \texttt{min\_size} =\textit{minimum possible face size} (for example, $\sim$ 1/4 to 1/16 of the image area in the case of video conferencing).
\ifPy
The function returns a list of tuples, \texttt{(rect, neighbors)}, where rect is a \cross{CvRect} specifying the object's extents
\cvarg{cascade}{Hidden Haar classifier cascade, created by \cvCPyCross{CreateHidHaarClassifierCascade}}
\cvarg{sum}{Integral (sum) single-channel image of 32-bit integer format. This image as well as the two subsequent images are used for fast feature evaluation and brightness/contrast normalization. They all can be retrieved from input 8-bit or floating point single-channel image using the function \cvCPyCross{Integral}}
\cvarg{sqsum}{Square sum single-channel image of 64-bit floating-point format}
\cvarg{tilted\_sum}{Tilted sum single-channel image of 32-bit integer format}
\cvarg{scale}{Window scale for the cascade. If \texttt{scale} =1, the original window size is used (objects of that size are searched) - the same size as specified in \cvCPyCross{LoadHaarClassifierCascade} (24x24 in the case of \texttt{default\_face\_cascade}), if \texttt{scale} =2, a two times larger window is used (48x48 in the case of default face cascade). While this will speed-up search about four times, faces smaller than 48x48 cannot be detected}
\end{description}
The function assigns images and/or window scale to the hidden classifier cascade. If image pointers are NULL, the previously set images are used further (i.e. NULLs mean "do not change images"). Scale parameter has no such a "protection" value, but the previous value can be retrieved by the \cvCPyCross{GetHaarClassifierCascadeScale} function and reused again. The function is used to prepare cascade for detecting object of the particular size in the particular image. The function is called internally by \cvCPyCross{HaarDetectObjects}, but it can be called by the user if they are using the lower-level function \cvCPyCross{RunHaarClassifierCascade}.
\cvarg{cascade}{Double pointer to the released cascade. The pointer is cleared by the function}
\end{description}
The function deallocates the cascade that has been created manually or loaded using \cvCPyCross{LoadHaarClassifierCascade} or \cvCPyCross{Load}.
\cvCPyFunc{RunHaarClassifierCascade}
Runs a cascade of boosted classifiers at the given image location.
\cvdefC{
int cvRunHaarClassifierCascade( \par CvHaarClassifierCascade* cascade,\par CvPoint pt,\par int start\_stage=0 );
}
\begin{description}
\cvarg{cascade}{Haar classifier cascade}
\cvarg{pt}{Top-left corner of the analyzed region. Size of the region is a original window size scaled by the currenly set scale. The current window size may be retrieved using the \cvCPyCross{GetHaarClassifierCascadeWindowSize} function}
\cvarg{start\_stage}{Initial zero-based index of the cascade stage to start from. The function assumes that all the previous stages are passed. This feature is used internally by \cvCPyCross{HaarDetectObjects} for better processor cache utilization}
\end{description}
The function runs the Haar classifier
cascade at a single image location. Before using this function the
integral images and the appropriate scale (window size) should be set
using \cvCPyCross{SetImagesForHaarClassifierCascade}. The function returns
a positive value if the analyzed rectangle passed all the classifier stages
(it is a candidate) and a zero or negative value otherwise.
\fi
\fi
\ifCpp
\cvclass{FeatureEvaluator}
Base class for computing feature values in cascade classifiers.
\cvarg{filename}{Name of file from which classifier will be load. File may contain as old haar classifier (trained by haartraining application) or new cascade classifier (trained traincascade application).}
\end{description}
\cvCppFunc{CascadeClassifier::read}
Reads the classifier from a FileStorage node. File may contain a new cascade classifier (trained traincascade application) only.
\cvarg{rectList}{The input/output vector of rectangles. On output there will be retained and grouped rectangles}
\cvarg{groupThreshold}{The minimum possible number of rectangles, minus 1, in a group of rectangles to retain it.}
\cvarg{eps}{The relative difference between sides of the rectangles to merge them into a group}
\end{description}
The function is a wrapper for a generic function \cvCppCross{partition}. It clusters all the input rectangles using the rectangle equivalence criteria, that combines rectangles that have similar sizes and similar locations (the similarity is defined by \texttt{eps}). When \texttt{eps=0}, no clustering is done at all. If $\texttt{eps}\rightarrow+\inf$, all the rectangles will be put in one cluster. Then, the small clusters, containing less than or equal to \texttt{groupThreshold} rectangles, will be rejected. In each other cluster the average rectangle will be computed and put into the output rectangle list.
