[Algorithm Research] Target Detection

Author: gnuhpc

Source: http://www.cnblogs.com/gnuhpc/

The target detection method was initially proposed by Paul Viola [vila01] and improved by Rainer lienhart [lienhart02. first, a cascade boosted classifier is obtained by training the classifier using the Harr feature of the sample (about several hundred sample images. The training samples are divided into positive samples and inverse samples. The positive samples refer to the target samples to be checked (such as faces or automobiles), and the negative samples refer to any other images, all the sample images are normalized to the same size (for example, 20x20 ).

After the classifier is trained, it can be applied to the detection of the areas of interest (the same size as the training sample) in the input image. The output of the classifier in the target area (car or face) is 1. Otherwise, the output is 0. To detect the entire image, you can move the search window in the image and check each position to determine the possible target. In order to search for objects of different sizes, the classifier is designed to be able to change the size, which is more effective than changing the size of the image to be examined. Therefore, to detect an unknown target object in an image, the scanner usually needs to scan the image several times in a search window of different proportions.

In a classifier, "cascade" means that the final classifier is composed of several simple classifier Cascade. In image detection, the detected window uses each level of classifier in sequence,
In this way, most of the candidate regions in the previous detection layers are excluded, and all the regions detected by each level of classifier are the target regions. Currently, four types of boosting technologies are supported: discrete Adaboost, real Adaboost, gentle AdaBoost and logitboost. "Boosted" means that each layer of the cascade classifier can select a boosting algorithm (weight voting) and use the self-training of the basic classifier. A basic classifier is a decision tree classifier with at least two leaf nodes. The Haar feature is the input of the basic classifier. It is mainly described as follows. Current algorithms mainly use the following Harr features.

The features used by each specific classifier use shapes, locations in the areas of interest, and proportional coefficients (the proportional coefficients here are different from those used during detection, although the product value of the two coefficients will be obtained at the end. For example, in the case of the third feature (2C), the response is calculated to cover the entire rectangle of all features (including two white rectangles and one black rectangle) the sum of pixels is three times the sum of pixels in the black rectangle.
. The pixel and sum in each rectangle can be quickly calculated using the integral image. (View the following and the description of cvintegral ).

The demo version of haarfacedetect can be used to check the working status of the target detection.

Author: gnuhpc

Source: http://www.cnblogs.com/gnuhpc/