KR101433472B1 - Apparatus, method and computer readable recording medium for detecting, recognizing and tracking an object based on a situation recognition - Google Patents

Abstract

The present invention relates to an object detection, recognition and tracking apparatus based on context recognition for detecting, recognizing and tracking objects based on context recognition for surveillance intrusion detection, and more particularly, An object feature point extraction module for extracting numerical feature vector values from element information; And an object detection and tracking module for detecting and tracking an object by classifying an image according to the type of an operation pattern of the object with respect to the minutiae point of each object extracted from the object minutiae point extraction module, When the object is detected, a different threshold is applied to each cluster for the entire target image.

Description

[0001] APPARATUS, METHOD AND COMPUTER READABLE RECORDING MEDIUM FOR DETECTING, RECOGNIZING AND TRACKING AN OBJECT BASED ON A SITUATION RECOGNITION [0002] BACKGROUND OF THE INVENTION [0003]

The present invention relates to an apparatus for detecting, recognizing and tracking an object infiltrating a specific space, and more particularly, to an apparatus for detecting an object based on context recognition for detecting surveillance intrusion, Based object detection, recognition and tracking apparatus, method and computer-readable recording medium.

Surveillance intrusion detection is evolving into an auxiliary or main function security control system in airports, offices and apartments for personal security and social security situations. Analysis of human perception and behavior pattern in surveillance intrusion detection based on limited space is an important factor in security, so pattern analysis of situation recognition and data mining is essential. Recently, the demand of various real - time security services along with the arrival of a high - level knowledge and information society to improve the quality of life has demanded the advent of convergence security technology through the fusion of intelligent sensor and situational awareness.

On the other hand, the need for security has been strongly emphasized recently, and the importance of control technology for physical access security is particularly emphasized. As a technology for physical access control most commonly used at present, there is a problem that there is no fundamental countermeasure against loss of a card in the method of a key pad and an RFID card.

In order to solve these problems, there is a great interest in object recognition technology in dynamic images, and fingerprint recognition technology is widely popular among biometrics technologies. However, there is a gradual decrease in recognition rate due to physical contact, vulnerability to children and the elderly, and countermeasures against fingerprint duplication. In addition, although the reed recognition technology has a relatively high recognition rate, it requires an expensive input device for a high-resolution image and requires a user's mastery in use.

On the other hand, most current intrusion detection systems have weaknesses in noise images due to changes in illumination and external factors, and thus they are very difficult in the field of physical access market.

Therefore, by using the dynamic image-based and context-adaptive programming techniques to maximize the non-contact and user-friendly interface, which is an advantage of object detection technology, and to overcome the vulnerability due to dynamic changes of illumination and objects, Technology development is urgently required.

It is an object of the present invention to provide an intrusion detection system capable of providing an intrusion detection system that requires high reliability by minimizing the influence of a change in a lighting environment (illumination direction, brightness contrast, contrast, and the like) A recognition-based object detection, recognition and tracking device, a method, and a computer-readable recording medium.

Also, an object of the present invention is to provide an algorithm and system for minimizing the influence of external illumination and noise, and to develop an active vision based high reliability object tracking system through development of feature point extraction technology and tracking program for an object having a certain performance A method and a computer readable recording medium capable of providing an object detection, recognition,

In order to achieve the above-described object of the present invention and to achieve the specific effects of the present invention described below, the characteristic structure of the present invention is as follows.

According to an aspect of the present invention, an object feature point extraction module extracts a feature vector value digitized from characteristic element information of an object extracted from an input object image; And an object detection and tracking module for detecting and tracking an object by classifying an image according to the type of an operation pattern of the object with respect to the minutiae point of each object extracted from the object minutiae point extraction module, When the object is detected, a different threshold is applied to each cluster for the entire target image.

Advantageously, said object detection and tracking module detects and tracks objects using Bayesian classifiers based on the 'Harr' feature.

Preferably, the object feature point extraction module includes: a region of interest model generation unit that generates a region of interest model for setting a search region of each feature point of the object; And a detector generator for generating a detector for each feature point for detecting a feature point in the search area.

Preferably, the ROI generation unit sets the coordinates of the set number of feature points as a vector, normalizes the center points of the coordinates so that the distance between the two feature points is 1, finds the most unique feature point of the object, The normalized mean vector of coordinates and the area vector that calculates the distance changeable from the average position of the feature points is set as the region of interest model.

Preferably, the object feature point extracting module further includes a fast Fourier calculator for performing a fast Fourier transform on the search area.

Preferably, the object feature point extraction module further includes a post-processing unit that excludes a portion of the spectral image processed through the fast Fourier calculator that is less than a predetermined threshold value from the feature candidate region of the object.

Preferably, the object feature point extracting module further includes an image preprocessing unit that processes the input image so that the distribution of the lightness value is uniformized by the histogram smoothing method.

According to another aspect of the present invention, there is provided an object feature extraction method for extracting a feature vector value digitized from feature information of an object extracted from an input object image; And detecting and tracking the object by classifying the image according to the shape of the operation pattern of the object with respect to the extracted feature points of the object, wherein the step of detecting and tracking the object comprises: A different threshold value is applied to each cluster.

Preferably, detecting and tracking the object detects and tracks the object using a Bayesian classifier based on the 'Harr' feature.

Preferably, the object feature point extracting step includes: a ROI model generating step of creating a ROI model for setting a search area of each feature point of the object; And a detector generating step of generating a detector for each feature point for detecting a feature point in the search area.

Preferably, the generating of the ROI model includes: setting coordinates of a set number of feature points as a vector; Normalizing the center point of the coordinates so as to find the most unique characteristic point of the object so that the distance between the two characteristic points is 1; And setting a normalized average coordinate vector of the minutiae of each object and a region vector in which a changeable distance at an average position of the minutiae points is calculated as a region of interest model.

Advantageously, the object feature point extracting step further includes a fast Fourier transforming step of performing a fast Fourier transform on the search area.

Preferably, the object feature point extracting step further includes a post-processing step of excluding, from the feature candidate region of the object, a portion of the spectral image processed through the fast Fourier transform that is less than a predetermined threshold value.

Preferably, the object feature point extracting step further includes an image preprocessing step of processing the input image so that the distribution of lightness values is uniformized by a histogram smoothing method.

Meanwhile, the information for providing the object detection, recognition, and tracking method based on the situation recognition may be stored in a recording medium readable by a server computer. Such a recording medium includes all kinds of recording media in which programs and data are stored so that they can be read by a computer system. Examples include ROMs (Read Only Memory), Random Access Memory, CD (Compact Disk), DVD (Digital Video Disk) -ROM, magnetic tape, floppy disk, optical data storage device, (For example, transmission over the Internet). Such a recording medium may also be distributed over a networked computer system so that computer readable code in a distributed manner can be stored and executed.

As described above, according to the present invention, an intrusion detection system that requires high reliability is provided by minimizing the influence of a change in a lighting environment (illumination direction, brightness contrast, contrast, etc.), which is a fatal problem of existing object detection and tracking technology There is an advantage to be able to do.

According to the present invention, there is provided an algorithm and system for minimizing influence of external illumination and noise, thereby providing an active vision-based high reliability object tracking system through development of feature point extraction technology and tracking program for an object having a constant performance There is an advantage to be able to do.

In addition, according to the present invention, access security is performed through object detection and tracking using image information of an object for non-depressed image information measurement. Therefore, the object detection and tracking technology in the input image according to the present invention can create effects such as human-centered user interface, automation of factory inspection, operation of household appliances through motion recognition, have.

In addition, according to the present invention, there is an advantage that it is effective for managing access to a person using a monitoring function, can generate high efficiency with a low cost, has an intrusion detection function, and can perform reliable system management.

Meanwhile, market requirements for security facilities and systems in recent building or authorized areas are increasing, and by developing a system according to the present invention, many fields such as hospitals, welfare facilities, protection facilities, personal security facilities, There are advantages that can be applied to.

In addition, according to the present invention, it is possible to add and configure functions in accordance with the purpose and environment through a customer at a large discount mart or a distributor through customer flow through flow analysis and purchasing behavior analysis.

BRIEF DESCRIPTION OF THE DRAWINGS A brief description of each drawing is provided to more fully understand the drawings recited in the description of the invention.
1 is a block diagram illustrating a detailed structure of an object feature point extraction module of a system according to an embodiment of the present invention.
2 is a block diagram illustrating a detailed structure of an object detection and tracking module of a system according to an embodiment of the present invention.
3 is a flowchart of a feature point detection of an object according to an embodiment of the present invention.
FIGS. 4 and 5 are flowcharts of object detection and tracking in the surveillance intrusion detection according to an exemplary embodiment of the present invention.
6 is a diagram illustrating minutiae and non-minutiae data collection according to an embodiment of the present invention.
7 is a diagram illustrating a text file storing feature vectors according to an embodiment of the present invention.
8 is a view showing an image used in an experiment according to an embodiment of the present invention.
9 is a diagram illustrating a Fourier transform for a feature region according to an embodiment of the present invention.
10 is a diagram illustrating a Fourier transform for a location approximately 4 pixels to the right from a feature region according to an embodiment of the present invention.
FIG. 11 is a diagram showing the result of measurement of the number of comparison blocks and the Fourier transform time relationship according to the embodiment of the present invention.
12 is a diagram showing minutiae points of an extracted object according to an embodiment of the present invention.
13 is a diagram showing a structure of a learning system according to an embodiment of the present invention.
14 is a diagram illustrating an object detection and tracking process in a surveillance according to an embodiment of the present invention.
FIG. 15 is a view showing an experiment result of object detection rate per sample according to an embodiment of the present invention. FIG.
16 is a diagram illustrating a clustering method in a first direction according to an embodiment of the present invention.
17 is a diagram illustrating a clustering method in a second direction according to an embodiment of the present invention.
18 is a diagram illustrating an object image recognition process according to an embodiment of the present invention.
FIG. 19 is a diagram showing a post-processing process by merger and adjustment according to the embodiment of the present invention. FIG.
20 is a diagram illustrating the installation of a stereo / zoom camera in a building according to an embodiment of the present invention.
FIG. 21 is a diagram illustrating a movement path trace flow according to an embodiment of the present invention.
22 is a diagram illustrating stereo matching modeling according to an embodiment of the present invention.
23 is a diagram illustrating a tracking process of a moving object according to an embodiment of the present invention.
24 is a view showing an image preprocessed using histogram smoothing according to an embodiment of the present invention.
25 is a diagram showing minutiae points of a detected object according to an embodiment of the present invention.
26 is a diagram illustrating a spectral image acquisition process through preprocessing and fast Fourier operations according to an embodiment of the present invention.

The following detailed description of the invention refers to the accompanying drawings, which illustrate, by way of illustration, specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that the various embodiments of the present invention are different, but need not be mutually exclusive. For example, certain features, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the invention in connection with an embodiment. It is also to be understood that the position or arrangement of the individual components within each disclosed embodiment may be varied without departing from the spirit and scope of the invention. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present invention is to be limited only by the appended claims, along with the full scope of equivalents to which the claims are entitled, if properly explained. In the drawings, like reference numerals refer to the same or similar functions throughout the several views.

In order to implement the present invention, in order to track a feature point or a feature region of an object in real time, the discrimination from the surrounding region must be distinct. This is always regarded as an important attribute in feature point tracking. In the present invention, preprocessing is performed on the input image so as not to be sensitive to noise and illumination changes in the input object image, and the feature vector values are extracted from the characteristic element information of the extracted object. Optimization of preprocessing improves the speed of object feature detection.

Meanwhile, in the proposed system according to the embodiment of the present invention, such a minutiae point detection step is largely composed of a learning step and a detection step. In the learning phase, the ROI model is set up to set the search area of each feature point of the object and the detector of each feature point to detect the feature point in the search area. In the detection step, the search region of each feature point is set in the input image using the ROI model generated in the learning step, and the feature point is detected using Fast Fourier Transform (FFT) for the search region.

When a fast Fourier transform is applied to the feature region of an object, a spectrum form different from the surrounding region of the object is obtained. By using this property, each feature point of the object is detected, and this method is faster in detecting the feature point than the existing method.

Based on the feature point detection method of these objects, it is possible to detect and track only human objects by using Bayesian classifier when several same objects, ie, human objects, are contained in one input image. The various conditions of the input image are affected by the lighting condition, the motion pattern of the object, the background image, and the noise of the foreground image.

For images with various situations, we classify the objects into n groups in the image. The feature points are detected using the most optimal Bayesian classifier for the images of the classified object. Detecting the feature point as the most optimal Bayesian classifier means detecting the feature point using a threshold set arbitrarily according to the classified image. After that, the object is detected by determining the center of the object through post-processing such as merger and adjustment method. In this process, a method called 2D Haar wavelet transformation is used, which makes the system according to the invention less susceptible to illumination.

Therefore, the method based on the context analysis of the image according to the present invention shows better performance than the conventional method in terms of accuracy of object detection and tracking for various changing environments. In addition, the merging and reconciliation strategies used in the post-processing process are based on the problem of correction of the difference in brightness per detection area and the processing speed of the object detection system, which makes the detection and tracking algorithm of various motion patterns difficult, Can be solved.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT Hereinafter, a detailed description according to a preferred embodiment of the present invention will be described with reference to the accompanying drawings. In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear.

The surveillance system according to the embodiment of the present invention may include an object feature point extraction module 100 and an object detection and tracking module 200.

FIG. 1 is a block diagram illustrating a detailed structure of an object feature point extraction module 100 of a system according to an embodiment of the present invention. FIG. 2 is a detailed view of an object detection and tracking module 200 of a system according to an embodiment of the present invention. Fig.

FIG. 3 is a flowchart of a feature point detection of an object using an object feature point extraction module according to an embodiment of the present invention. FIG. 4 (or FIG. 4-1) FIG.

1, the object feature point extraction module 100 includes an object component detection unit 110, an object component normalization unit 120, a learning database storage unit 130, a fast Fourier calculation unit 140, a post-processing unit 150 A feature point extraction unit 160, and the like.

Referring to the characteristic point detection flowchart of the object using the object characteristic point extraction module of FIG. 3, the function of each building block of the object characteristic point extraction module 100 of FIG. 1 and the characteristic point extraction process of the object object characteristic point extraction module 100 .

The object component detection unit 110 performs a function of selecting a valid position of a unique feature region having only the object to be detected having shape and color information that are distinguished from each other when compared with other object images.

The object component normalization unit 120 includes an image preprocessing unit 121, a ROI generating unit 122, a feature vector extracting unit 123 and a detector generating unit 124, And a detector is generated by using a Bayesian discriminant.

In detail, the image preprocessing unit 121 performs an image processing method that distributes the pixel brightness values that are biased to specific values, such as when the input image is too dark or too bright, to a wider range of the image. In the present invention, when there is a noise image due to light and darkness in the input image using the histogram equalization method, the distribution of the light and dark values is uniformized so that the image can be improved.

In order to detect the feature point of the object, the ROI generation unit 122 searches for a unique ROI that differentiates the object from other background images. In the present invention, a region of interest is constructed to find a region of interest of a minutiae, and utilized for more accurate minutiae detection. At this time, the ROI model is generated by using the analytical knowledge that the unique characteristic of the object to be detected always exists at a certain distance.

On the other hand, the ROI model is generated through the following three steps.

Step 1. Make the vectors of the 16 feature points as shown in <Table 1> below.

Coordinate x of feature point 1 Coordinate y of minutiae 1 Coordinate x of minutiae 2 ...... Coordinates x of 16 minutiae Coordinates of 16 feature points y

Step 2. Normalize the coordinate vector of the feature point. At this time, the center point of the coordinates is normalized so that the distance between the two feature points is 1, and the most unique feature point of the object is found.

Step 3. In the ROI model, a normalized average coordinate vector of each object feature point and a region vector that calculates a changeable distance from the average position of the feature points are set as ROI models. The area vector is a measure of how far each minutiae point of each minutiae coordinate vector is located in the average coordinate vector. At this time, the distance farthest from the average is stored in the area vector.

The feature vector extracting unit 123 uses two class models to classify the feature points using Bayesian. The two class models are feature point models and non-feature point models. In the present invention, as a vector for a minutia model, a 40-dimensional vector is extracted through 2D Haar wavelets having 8 frequencies and 5 directions for 3 × 3 pixels including feature points and surrounding pixels as shown in FIG. 6 . Since the vector for the non-feature point model is to be detected within the feature point interest region, vectorization is performed using 8 pixels 6 pixels away from the feature point.

The detector generating unit 124 vectors the minutiae and non-minutiae using the 2D Haar wavelet as the minutiae detector. At this time, the extracted feature vector is generated by a Bayesian discriminant. Since the feature points and non-feature point patterns are represented by a multivariate normal distribution, the Mahalanobis distance formula related to pattern discrimination is used as a discriminant. Since the Mahalanobis distance formula can be measured using the covariance of the training patterns and the mean vector, each covariance and average vector are generated using the data of the minutiae pattern and the non-minutiae pattern in the learning phase.

The learning database storage unit 130 stores the feature vector values generated in the above step in a file format. This file can be saved as a text file (* .txt) or as an Excel file (* .xls). Regardless of the format of the file, it is stored in a certain format and delimiters so that it can be used as an input value at a later stage. 7 is an example of a text file storing the resultant feature vector value.

The fast Fourier calculator 140 uses the feature vector values stored in the learning database as the input values of the fast Fourier calculations for the feature point detection of the object to be searched. Fast Fourier transform (FFT) is a tool for analyzing frequency components included in arbitrary signals and is applied not only to image processing but also to signal analysis and speech processing.

8 is an image used in an experiment of a fast Fourier operation. 8, fast Fourier transform is performed on the feature region as shown in FIG. FIG. 10 shows a spectral image obtained by performing a fast Fourier transform on a peripheral region of a selected feature region. As shown in FIG. 9 and FIG. 10, it can be seen that the transformed spectrum has a different shape.

When the object is detected and tracked in the surveillance intrusion detection platform, the analysis of the feature region is efficient by using the fast Fourier operation. Since the feature point detection method to be detected according to the embodiment of the present invention uses fast Fourier transform, the feature point detection speed is fast and the probability of failing in tracking the object at a later stage is lowered. For example, in the surveillance intrusion detection system, it is possible to improve the accuracy of object tracking when tracking unique feature points of less than 5 objects in real time.

The post-processor 150 applies a certain threshold to the spectral image to remove noise. That is, the region is used as the feature region of the object only when it is greater than or equal to a preset specific threshold value, and the portion below the threshold is excluded from the feature region of the object. If all the classified data are the minutiae of the object, post-processing may not be necessary. However, this process is intended to eliminate the error of object feature point extraction, which occurs when a fast Fourier operation is applied to an input image.

The feature point extraction unit 160 performs a function of determining the location of the feature points of the object in the object candidate region found through post-processing according to the merger and adjustment strategy. In the present invention, the Bayesian discriminant method for the extracted feature points of the object shows high performance in all the feature points. In the case of the human object, the ratio of high data in the joint region and the head region, especially the legs and arms, is shown. 12 shows the minutiae points of the extracted object.

As described above, the feature points of the object extracted through the object feature point extraction module 100 are classified according to the type of operation pattern in the object detection and tracking step. At this time, another threshold value optimized in the cluster is applied to the Bayesian method.

Such a method according to the present invention is called a situation-based object detection method of an image. The general Bayesian method applies the same threshold for the entire test image, but the context-based method according to the present invention applies an adaptive threshold according to the number of clusters.

FIG. 13 shows a process of learning using two different levels of network. In the first step, non-supervised learning is conducted. In the second step, we use regular supervisory methods using various thresholds. At this time, the input pattern is vector-processed to a gray scale image size of 64x64 pixels. On the other hand, the deformation from the input space to the hidden unit space is nonlinear, and the deformation from the hidden unit space to the output space is linear. The RBF classifier extends the input vector to a higher dimensional space and learns it using K-means. The network input is composed of images that are normalized by the n values entered into the network as a one-dimensional vector and readjusted to 1/2 size.

Figure 14 shows the detection and tracking process of an object in surveillance. For example, it shows a process of detecting and tracking an object using a Bayesian classifier based on the Haar feature according to an embodiment of the present invention. In the present invention, a feature vector is constructed through a 2D Haar wavelet for object detection and is used to learn a Bayesian classifier.

In this way, the method based on the context-based analysis of the image created in the present invention is utilized for object detection and tracking in surveillance, thereby improving the efficiency and accuracy even in a dynamic environment. This method shows a certain performance even when the background brightness or background noise is mixed in the input image.

2, the object detection and tracking module 200 includes an image condition based analysis unit 210, an object image classification unit 220, a situation adaptive Bayesian classification unit 230, an object candidate region detection unit 240, An object extraction unit 250, and the like.

Referring to the flow chart of the object detection and tracking using the object detection and tracking module in the surveillance of FIG. 4, the object detection and tracking module of the object detection and tracking module 200 and the object detection and tracking module 200) will be described in order to detect and track objects in surveillance.

In order to detect and track the object, the input image is processed as shown in FIG. The object detector is designed to detect objects with the minimum size of 64 × 64 pixels rotated to ± 20 degrees in the preprocessed image. The detector uses the path of the image context-based method and the Bayesian context recognition method.

On the other hand, the various situations of the image are affected by the illumination state, the motion pattern, the background image, and the noise of the foreground image. At this time, the image condition is modeled and analyzed by the hybrid network of K-means and RBF. In order to allow the Bayesian classifier to be adjusted according to the changing background image and motion pattern, the situation analysis of the image is performed by assigning the detected object image to one of several image situations. Therefore, the method based on the context analysis of the image shows better performance than the existing method in terms of the accuracy of object detection and tracking for various changing environments.

In the present invention, a post-processing process is used as well as a Haar wavelet based on a Bayesian framework to reduce the influence of various environments. The post-processing process solves the noise problem of the background image through merger and adjustment procedures.

The detection algorithm consists of three stages of modules. The first step is the context-based analysis of the image, and the second step is the application of the Bayesian discriminant with the context adaptive threshold. The third step is the application of merger and adjustment strategies to the candidate areas. In this case, objects are classified by class using K-means and RBF algorithm. The image classified by the situation analysis of the image passes through the Bayesian detector. The Bayesian detector is implemented as a candidate window region classifier with a Haar feature based on a six-level multiresolution method, a feature vector generator using a 2D Haar transformer, and a Bayesian discriminant method.

After object is detected using object detection method, histogram smoothing is processed into detected area. At this time, the detected object image is viewed from the front. However, the behavior patterns of these object fronts can easily be extended to other side or back-side motion patterns. After that, the area is scanned by a 64x64 window.

The 2D Haar wavelet transform is performed for each window. Each window is tested by the Harr feature based on the Bayesian classifier to determine whether the window is a candidate or not. 2D Haar wavelet transforms make the system less susceptible to illumination. Bayesian classifiers based on the Haar feature have candidate regions with different scales. Finally, merger and adjustment strategies are applied and the center of the object is determined.

The context-based analysis unit 210 assigns the detected image to one of the categories of the image. The situation of the image is modeled and analyzed by the hybrid network of K-means and RBF. The RBF classifier expands the input vector into a high dimensional space. The RBF network has a traditional 3-tier back-propagation structure. The hidden units at this stage are trained using k-means. The network input consists of n normalized and resized object images into a network, such as a one-dimensional vector. At this time, the input unit has a real value, and the vector value is normalized.

The object image classifying unit 220 determines whether the object is an object by applying a clustering algorithm. In other words, by improving the method of performing element partitioning of an object, it is verified that the image of the object to be extracted is extracted based on the characteristics of the extracted object components. At this time, the region information of the object is efficiently reduced by using the feature points, and the image of the object is classified using the analyzed feature elements by using the clustering method such as the similarity measurement, PCA, k-NN and correlation method. That is, the center point of each object is found and clustering is performed to separate each part.

Clustering is performed with two directions of the projection vector. 16, the " 1st K-means "is performed using a width as shown in FIG. 16, and the" 2nd clustering " ) Values, we can see the effect of noise reduction after 2nd clustering.

The context adaptive Bayesian classifier 230 applies another threshold value optimized in the cluster to the Bayesian method for the image of the object classified in the above step. This is called the context-based object detection method of the image. The general Bayesian method applies the same threshold for the entire test image, but the context-based method according to the present invention applies an adaptive threshold according to the number of clusters.

The object candidate region detection unit 240 uses a post-processing process as well as a Haar wavelet based on the Bayesian framework to reduce the influence of various environments. The post-processing process solves the noise problem of the background image through merger and adjustment procedures. Image classified by image situation analysis passes through a Bayesian detector. The Bayesian detector is implemented as a candidate window region classifier with a Haar feature based on a six-level multiresolution method, a feature vector generator using a 2D Haar transformer, and a Bayesian discriminant method.

The 2D Haar wavelet transform is performed for each window. Each window is tested by the Harr feature based on the Bayesian classifier to determine whether the window is a candidate or not. 2D Haar wavelet transforms make the system less susceptible to illumination. Bayesian classifiers based on the Haar feature have candidate regions with different scales. At this time, as shown in FIG. 19, the merger and adjustment strategy are applied and the center of the object is determined.

The merging and reconciliation strategy can solve the problem of the correction of the difference of brightness per detection area and the processing speed of the object detection system which makes the object detection and tracking algorithm of various motion patterns which have been raised recently difficult.

The object tracking unit 250 performs security and monitoring of moving objects in a specific area / area. That is, as shown in FIG. 20, a stereoscopic camera and a zoom camera are installed in an area required in a building, and the acquired image information is used to grasp the current position of the person / VIP and control access to the specific area.

At this time, as shown in FIG. 21, the object is tracked by using two-dimensional information of the motion inputted from each region. In addition to the motion information of the object, the zoom camera transmits the facial image information of the approaching person to the server when the access security is extremely needed, and the server grasps the personal details of the person who is about to approach and performs physical access control . In order to track the detected object, the object is tracked using the image information input from the installed stereo / zoom cameras.

As shown in FIG. 22, in the stereo matching processing method, when overlapping objects occur in the input image, the overlapping avoidance is used to distinguish overlapping objects. A pair of images from the same scene is reconstructed through a plurality of lenses spaced apart from each other to restore portions visible in the two images. A method used in a system for tracking existing objects is a model-based method in which an object is acquired in an image using a reference image method and the model is matched based on the obtained object. Therefore, it is a method to track an object by obtaining a moving object under a complicated background, then matching the model with a minimum hausdorff distance, and then finding a position where the result of the matching is the minimum. However, in the present invention, a mahalanobis distance is used as a method of defining a three-dimensional model of an object to be tracked and matching the three-dimensional model to each image.

The stereo matching processing method disclosed in the present invention and the method based on a low-resolution image difference use a difference between successive image frames to extract an object from a certain background. A 2D median filter is used to extract the area of the object using the above method and then to remove noise pixels. The method based on the low-resolution image difference traces the moving object using the image difference after eliminating the noise using low-resolution representation and filtering of the image. Therefore, even if there are many noise images or overlapping objects, have. 23 shows a flow chart of object tracking using the method presented in the present invention.

Hereinafter, a method of detecting, recognizing, and tracking an object based on the context recognition according to an embodiment of the present invention will be described with reference to FIG. 3 to FIG. 5. In the embodiment of the object detection, Detailed description thereof will be omitted.

FIG. 3 is a flowchart of a feature point detection of an object using an object feature point extraction module according to an embodiment of the present invention, and FIGS. 4 and 5 are flowcharts of object detection and tracking using an object detection and tracking module in a surveillance.

3 and 4, a method for detecting, recognizing, and tracking an object based on context recognition according to an exemplary embodiment of the present invention may be performed using the object feature point extraction module and the object detection and tracking module, 3 and the object detection and tracking process of FIGS. 4 and 5 are combined to detect and track objects in surveillance.

3, the learning step and the feature point detection step are performed as described above. In the learning step, a ROI model for setting a search region of each feature point of the object is generated, and a feature point is detected in the search region Thereby generating a detector for each feature point.

The learning step extracts each feature point of an object, obtains a number called a feature vector value according to the formula, and stores the numerical values as a learning file. In order to recognize an object in the learning step proposed in the present invention, feature points that express the characteristics of the object should be selected. A feature point is a unique feature of the object with distinct shape and color information when compared with other object images. These features are characterized by high discriminative power compared to other parts that make gentle low-frequency components such as moving parts in the object or protruding parts that are remarkably protruded because the value changes consist of high-frequency components that are abruptly changed.

On the other hand, in order to detect the minutiae of the object, it is necessary to find the region of interest of the object. In the present invention, the region of interest is performed by a heuristic method based on the horizontal and vertical image histogram analysis, and the region of each important object is found, and then the unique characteristic region of the object is found again. In the method used in the existing intrusion detection system, all the features of the object are detected and then the feature points are extracted. However, the method according to the present invention helps to find each feature point more accurately, but in contrast, There is a drawback.

In order to solve this problem, the present invention uses the ROI method using the fact that the unique characteristic of the object always exists at a constant distance.

The ROI model is generated through the following three steps.

Step 1. Make the coordinates of the 16 feature points a vector.

Coordinate x of feature point 1 Coordinate y of minutiae 1 Coordinate x of minutiae 2 ...... Coordinates x of 16 minutiae Coordinates of 16 feature points y

Step 2. Normalize the coordinate vector of the feature point. That is, normalize the distance between two neighboring feature points to be 1.

Step 3. Regarding the ROI model, the ROIs of the normalized average coordinate vectors of the minutiae of each object and the range of changeable distance from the average position of the minutiae are set as the ROI model. The region vector is a measure of the distance between each feature point of each feature point coordinate vector in the mean coordinate vector. At this time, the distance farthest from the average is stored in the area vector.

In the image preprocessing step, the detected object is extracted to a size of 128 × 128, and histogram equalization is performed to extract an accurate Gabor wavelet vector. The histogram equalization is to adjust the density by redistributing the contrast values so that the histogram of the image is flat. If the histogram of an image is biased to a single part, the characteristics of the image do not appear clearly, and the image discrimination power drops. Therefore, the distribution of the histogram should be uniformly adjusted so that it is suitable for image processing. This is called histogram equalization. In the present invention, more precise vectors can be extracted through the preprocessing.

The feature vector extraction step is performed using a Gabor Wavelet method. The method is suitable for object recognition or detection because it is suitable as a feature vector of an object because of its biological relevance and physical characteristics.

In the present invention, the position of an object feature point is expressed using a graph composed of nodes and jets, which represent distance information as a basic object representation in the feature vector extraction step using a Gabor wavelet. In the present invention, a new data structure called a bunch graph is used, and a characteristic of a general object is expressed using a graph of an object and bunches composed of jets. Therefore, a banned graph constructed using each object data is used for the feature point detection.

On the other hand, in the method used for extracting the feature vector of the existing object, there is a limitation in optimizing the matching requirement at a high speed because a large amount of computation is required to perform the matching process for detecting the feature point of the object. Accordingly, in the present invention, a GA for enlarging, transforming, and rotating the object graph is encoded and used in the matching process to increase the object feature point detection rate.

Feature point detection using bunch is composed as follows. On the other hand, in the present invention, the learning data extracted for detection of a feature point in the tangible structure, that is, a set of hearing vectors, is referred to as a bunch. The following is how to construct bunch.

Step 1. Extract the feature point Gabor vectors of n learning images and store them in the database.

Step 2. Divide the Gabor vectors stored in the database into m clusters by the K-means cluster method.

Step 3. Extract the mean vector of each cluster.

first
Average vector of clusters second
Average vector of clusters third
Average vector of clusters ...... ...... mth
Average vector of clusters

Table 3 shows the configured helium bunch.

In the detector generation step, a Bayesian method is used to generate a feature point detector. At this time, since the minutiae point and non-minutiae point pattern are represented by multivariate normal distribution, the discrimination formula uses the Mahalanobis distance formula included in the multivariate normal distribution formula. Since the Mahalanobis distance formula can be measured using the covariance and mean vector of learning patterns, each covariance and average vector are generated using the data of the minutiae pattern and the nonminute minutiae pattern in the learning phase.

Through such a series of processes, the detector and the ROI model are generated and stored in the learning database. Next, each area of the feature point using the ROI model is set as the next step.

In order to detect feature points, a search area of each feature point is set using an average coordinate vector and a region vector, which are the ROI models generated in the learning step. In this case, the average coordinate vector and the area vector are adjusted according to the object using the distance between the center positions of two feature points of the input image object, and the average coordinate vector is obtained by moving the coordinate axes from the center of two minutiae points to the coordinate axes of the image, To be suitable.

In Equations (1) and (2), X is an average coordinate vector and D is a region vector.

The feature point detection method proposed in the present invention classifies the feature point candidate region through a Bayesian discriminant using an energy function. The energy function is a method used to find the position of the minimum energy. It is assumed that the outline of the object has the minimum energy.

In the feature point detection method proposed in the present invention, a feature point to be detected is a position corresponding to an edge as a characteristic feature of an object. In this case, using the fact that the energy function is closer to the edge as the energy size of the energy function is smaller and the distance from the edge is larger as the energy is larger, the intensity value and slope of the external energy function are used as a more flexible threshold value . Using this characteristic of the minutiae, the discriminant is constructed as shown in Equation (5) below.

G _p (x) and g _n (x) in Eq. (20) are the discriminant functions for the minutiae and non-minutiae regions and Mahalanobis distances for each class. Feature point candidates can be extracted using the above discriminant function. The candidates up to rank 10 of the above equation are obtained and feature points are detected through post-processing.

The detector and the ROI model data are called from the learning database to set the ROIs, and the ROI process is performed to set the ROIs. This can be performed by the Gabor wavelet method. In this case, unlike the prior art, it has a feature of minimizing the time for minutia calculation.

The feature point candidate detection step performs a feature point detection step to detect a feature point candidate using a Fourier transform. This can be performed by a Fast Fourier Transform (FFT) method. In this case, unlike the prior art, it has a feature of minimizing calculation time for finding a minutia candidate.

In the post-processing step, the minutiae points of the object are detected by performing a post-processing process on the detected minutiae candidate (group) through merger and adjustment strategies. This can be performed by the Harris method based on the Bayesian classifier. In this case, unlike the prior art, high detection performance is shown at all the feature points. In the case of the human object, in particular, joint regions such as legs and arms, And the ratio of high data in the area.

The feature points and object candidate regions of each object thus detected are used in an adaptive manner in each situation in an input image in which noise images are mixed in the detection and tracking process of the objects shown in FIGS.

The detection and tracking process of the objects of FIGS. 4 and 5 may include a pyramid-based person detection step 410, a preprocessing step 420, an image context based analysis step 430, a Bayesian classifier application step 450, And a post-processing step 460 of applying an adjustment strategy. The pyramid-based person detection step 410 is a process of detecting an object by dividing an input image into windows of a predetermined size, The person who is the object is detected. This can be accomplished by a " bspline "method called multi-resolution, which is a method that is scanned by a 16 x 16 object window. In this case, unlike the prior art, .

In the preprocessing step 420, a noise image is removed in order to extract a feature vector in an area in which an object exists in the input image, and the histogram smoothing process and the stereo matching modeling are performed. In this case, And a feature vector extraction function.

The situation-based analysis step 430 of the image normalizes the vector values by performing a process of modeling and analyzing by the hybrid network of K-means and RBF, as a process of assigning the detected image to one of the categories of images. In this case, objects are classified by class using K-means and RBF algorithm. The situation of the image is modeled and analyzed by the hybrid network of K-means and RBF. In addition, to analyze the situation of the image, the object image is detected as one of several image situations so that it can be adjusted according to the background image and the operation pattern in which the Bayesian classifier and post processing are changed. In this case, unlike the prior art, the method based on the context analysis of the image according to the present invention is characterized in that it performs better than the existing methods in terms of the accuracy of object detection and tracking for various changing environments.

The application of the Bayesian classifier 450 is performed by classifying feature points using a Bayesian algorithm as a process of applying a Bayesian method to an image of an object classified in the above step, . This can be performed by a method of classifying an object by assigning a threshold suitable for each situation of the object. In this case, unlike the prior art, a Bayesian classifier having an adaptive threshold value is used to detect and track other objects in the fusion security field not only in object detection.

In other words, a general Bayesian classifier uses a fixed threshold value, so that the probability of object detection failure increases by using a uniform reference when detecting an object in the input image. However, when an object is detected using the Bayesian method, which is another optimized threshold value proposed in the present invention, it is possible to effectively detect an object even when there is a noise image around the object or overlapping images with other objects Feature.

The merging and tuning strategy application step 460 is performed by performing the 2D Haar wavelet transform process as a process of characterizing the detected object candidate regions into several definite objects. In this case, unlike the conventional technique, it is difficult to detect the object of the various operation patterns and to make the tracking algorithm difficult. In this case, And the problem of the processing speed of the object detection system can be solved.

FIG. 5 is a view for explaining a post-processing process by merger and adjustment according to an embodiment of the present invention. Referring to FIG. 5, the merging and reconciliation strategy application steps will be described in more detail. In FIG. 5, the merger and reconciliation strategies are applied and the center of the object is determined. In this process, the merging strategy is devised to improve the reliability of object detection by removing false detected objects for candidate regions detected as objects. As shown in FIG. 5, it can be seen that a region in which multiple candidate windows of an object are overlapped is found near real objects having a high frequency, whereas a false detected object usually occurs in a place having a low frequency.

This method gives reasonable decision rules that can eliminate false detection of objects and allows real object detection with a single merged detection. At this time, the number of candidate windows detected at that position with respect to the center of each object is calculated. Therefore, if the number is greater than a certain threshold value, the position is classified as the area of the object. The location of the object searched for by the center of the neighboring detections found next to it can be eliminated by duplicate object detection. If a particular location is correctly classified as an object, all other detected locations that are overlapped are considered as errors and these can be removed from the object candidate. A large number of detected object candidates are preserved and less detected locations are removed. The merging and reconciliation strategy can solve the problem of the correction of the difference of brightness per detection area and the processing speed of the object detection system which makes the object detection and tracking algorithm of various motion patterns which have been raised recently difficult.

The feature of the technology proposed in the present invention is to improve the problem of degradation of the object image recognition performance due to the illumination change of the module which is the most problematic in the current image security system by the situation analysis method of the object which gives the self- It is a sufficient way to apply to mobile systems in real time. In addition, by using image information of a reliable object through environment-adaptive illumination and noise cancellation, a more stable detection performance can be exhibited against changes in internal and external environments using a surveillance system.

The present invention can be applied to an intrusion detection system requiring high reliability by minimizing the influence of a change in illumination environment (illumination direction, brightness contrast, contrast, etc.), which is a fatal problem of existing object detection and tracking technology. It is possible to provide an active vision-based high-reliability object tracking system by developing a feature point extraction technique of an object having a certain performance and a tracking program by developing an algorithm and system that minimizes the influence of external illumination and noise.

8 to 10 are views for explaining the implementation of the feature region setting and the effect thereof by applying the Fourier transform in the embodiment of the present invention. As described above, FIG. 8 is an image used in the experiment, FIG. 10 is a Fourier transformed spectrum for a region located about 4 pixels to the right from the characteristic region, and FIG. 11 is a Fourier transform time relationship measurement result of the number of comparison blocks.

There are three methods to use real-time image. There are "frame differencing, environment map" and "optical flow" methodologies. At this time, there are pros and cons. "Frame differencing" means that each frame is individually obtained, so it does not reflect the noise effect of the previous image. The disadvantage is that it is affected by the noise effect. On the other hand, the spectral methodology using fast Fourier calculations is more stable than the other three methods. A fast Fourier operation is a system that continuously updates a spectrum and detects a portion that is different from the generated spectrum. This method has a merit that it exhibits a stable effect on noise and is hardly affected by an image with a large change in illumination. Therefore, it is preferable to use a fast Fourier transform in the input image to reduce the noise effect.

In addition, in the present invention, a spectral image through a fast Fourier operation is used in parallel with a general Fourier operation. That is, in the present invention, first, the difference of pixels is obtained by using a spectral image of a general Fourier operation, then the sum of the number of pixels is summed, and if the value does not exceed a threshold value, Apply the spectral image from the Fourier calculations. The number of pixels with more differences than the spectral image of a normal Fourier operation can then be obtained. If the number of pixels changed from the above value is small after the application, the image is considered to have no change, and it is considered to be the same as the previous position. If it is not, motion is detected, so it removes basic noise through preprocessing. In this case, the characteristic of the object is restored by applying an algorithm for restoring the characteristics of the lost object in the portion for removing the noise.

The detection method of the moving object by the low resolution spectral image difference shows fast detection performance for the real time processing in the motion detection step by detecting the moving object using the image difference after noise reduction using the low resolution representation and filtering of the image, It is more effective than conventional methods in terms of accuracy. 20 shows a spectral image acquisition process through preprocessing and fast Fourier operations.

13 to 15 are views for explaining an object detection and tracking process and an effect thereof according to an embodiment of the present invention. FIG. 13 shows a structure of an implemented learning system. FIG. FIG. 15 shows the result of object detection rate experiment per sample. FIG.

The clustering-based object detection method refers to a cluster-based object detection method for determining whether or not an object region is not a pixel-based value of an object. A general method of the cluster-based object detection system is shown in FIG.

In the cluster-based object detection method, it is necessary to define the input pattern of the underlying object search window, and the defined input patterns are combined into several clusters using the k-means method. Accordingly, as shown in FIG. 13, the object to be searched is determined and a tracking window for tracking is set. At this time, the object is detected by performing Bayesian algorithm to determine whether the object is included or not in the area within the tracking window and whether the area is an object or a non-object. This method is referred to as a Bayesian classification method based on the situation analysis of an image in the present invention.

In an object-like nonobject detector modeling method similar to an object, feature vectors extracted in the previous step are classified into objects and non-objects through a Bayesian classifier. It classifies the input pattern into objects and non-objects by analyzing the statistical characteristics of the learning patterns of the input image. Unlike objects, objects that are difficult to model unlike objects can not be characterized. Since everything except objects is non - objects, it is difficult to model non - objects by extracting all of these features. Therefore, in the present invention, a non-object similar to an object is modeled based on the "Mahalanobis" distance of the object sample. The method of selecting these non-objects consists of the following steps.

1. Create a set of several object samples called Tf and f, the covariance matrix of Tf.

2. Calculate "Mahalanobis" distance df, i for f for all fiTfs.

3. Set the output value to max {df, i | 1imf}.

4. Compute several non-object sample sets as Tn and calculate df, j for each sample, njTn.

5. Store the argument nj as a non-object sample set Tl similar to an object.

6. Let the covariance matrix of the non-object be l and generate it from Tl.

As shown in FIG. 13, the object images are classified according to the state of the image in the learning step, and the "Bayesian" theory is applied as thresholds optimized in each cluster in the experience stage. This method is called a context-based method. Although the general "Bayesian" method applies the same threshold value for the entire test image, the context-based method according to the present invention applies several threshold values depending on the number of clusters.

This method is a method of learning a network in two distinct stages as shown in FIG. In the first step, we perform k-means. Next, in the second step, various thresholds are learned using a regular supervision method.

At this time, the input pattern is vectorized to a gray scale image size of 16 x 16 pixels, for example. Deformation from input space to hidden unit space is nonlinear. On the other hand, the transformation from the hidden unit space to the output space is linear. The RBF classifier expands the input vector into a high dimensional space. The RBF network has a traditional 3-tier back-propagation structure. In this process, hidden units are learned using k-means.

The network input consists of n normalized and resized object images into a network, such as a one-dimensional vector. At this time, the input unit has a real value. The vector values are then normalized.

Currently, there is an increasing need for security using object image information, and there is an increasing trend in domestic and overseas demand. The biometric information such as fingerprints and long texts is information input by a pressure type or a contact type. The iris information requires a close-up photograph to measure biometric information and has a disadvantage of using an expensive camera system.

The method proposed in the present invention performs access security through object detection and tracking using image information of an object in order to measure the non-depressed image information. In addition, the object detection and tracking technology in the input image according to the present invention can create effects such as human-centered user interface, automation of factory inspection, operation of household appliances through gesture recognition, . In addition, the proposed method is effective for human access control using monitoring function, it can generate high efficiency with low cost, and has an intrusion detection function, which makes it possible to perform reliable system management.

In recent years, the market requirements for security facilities and systems in buildings and certain authorized areas have been increasing, and by developing the system, it will be possible to use them in many fields such as hospitals, welfare facilities, . It is possible to add and configure functions in accordance with the purpose and environment through large discount marts or distribution companies through customer flow through customer flow analysis and purchasing behavior analysis.

Meanwhile, the object detection, recognition and tracking method based on context recognition in a smart home according to an embodiment of the present invention can be implemented as a computer-readable code on a computer-readable recording medium. A computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer system is stored. Examples of the computer-readable recording medium include a ROM, a RAM, a CD-ROM, a magnetic tape, a hard disk, a floppy disk, an optical data storage device, and the like in the form of a carrier wave (for example, .

The various operations and functions as described herein with respect to the various methods may be performed by any of a number of types of functions, such as a particular processing function and / or a processing function implemented therein, and / &Lt; / RTI &gt; For example, such functions may be performed by various operations and processes as described herein, or by performing any other operations and functions as described herein, or the like, as well as performing each of the equivalents thereof, And may perform such operations, processes, and the like.

The invention has been described above with the aim of method steps illustrating the performance of certain functions and their relationships. The boundaries and order of these functional components and method steps have been arbitrarily defined herein for convenience of description. Alternative boundaries and sequences may be defined as long as the specific functions and relationships are properly performed. Any such alternative boundaries and sequences are therefore within the scope and spirit of the claimed invention.

As described above, the present invention has been described with reference to particular embodiments, such as specific elements, and specific embodiments and drawings. However, it should be understood that the present invention is not limited to the above- And various modifications and changes may be made thereto by those skilled in the art to which the present invention pertains.

Accordingly, the spirit of the present invention should not be construed as being limited to the embodiments described, and all of the equivalents or equivalents of the claims, as well as the following claims, belong to the scope of the present invention .

100: object feature point extraction module 110: object component detection unit
120: object component normalization unit 121: image preprocessing unit
122: ROI model generating unit 123:
124: detector generating unit 130: learning D / B storage unit
140: Fast Fourier operation unit 150: Post-
160: feature point extracting unit 200: object detecting and tracking module
210: Situation-based analysis unit of image 220: Object image classification unit
230: Situation Adaptive Bayesian Classification
240: Object candidate region detection unit 250: Object tracking unit

Claims (15)

An object feature point extraction module for extracting numerical feature vector values from characteristic element information of an object extracted from an input object image; And
And an object detection and tracking module for detecting and tracking an object by classifying the input object image according to the type of the operation pattern of the object with respect to the minutiae point of each object extracted from the object minutiae point extraction module,
Wherein the object detection and tracking module applies a different threshold value to each of the clusters for the input object image when the object is detected,
Wherein the object feature point extraction module comprises:
A region of interest model generation unit for generating a region of interest model for setting a search region of each feature point of the object; A fast Fourier calculator for performing a fast Fourier transform on the search area; And a post-processing unit for excluding, from the feature candidate region of the object, a portion of the spectral image processed through the fast Fourier arithmetic unit, the coefficient value of the frequency component of the image being lower than a predetermined threshold value,
The ROI generator sets the coordinates of the set number of feature points as a vector and normalizes the center points of the coordinates so that the distance between two adjacent feature points is 1. The normalized average coordinate vector of the feature points of each object and the feature point And sets a region vector that calculates a changeable distance at an average position as a region of interest model.
The system of claim 1, wherein the object detection and tracking module comprises:
A context - aware object detection, recognition, and tracking system that detects and tracks objects using Bayesian classifiers based on 'Haar' features.
The method of claim 1, wherein the object feature point extraction module comprises:
And a detector generating unit for generating a detector for each feature point for detecting a feature point in the search region.
delete delete delete The method of claim 1, wherein the object feature point extraction module comprises:
And an image preprocessing unit for processing the input image so that the distribution of the lightness value is made uniform by the histogram smoothing method.
An object feature point extraction step of extracting a digitized feature vector value from characteristic element information of an object extracted from an input object image; And
Detecting and tracking the object by classifying the object image input according to the shape of the operation pattern of the object with respect to the extracted feature points of the object,
The detecting and tracking of the object may include applying a different threshold value to each cluster of the entire object image when the object is detected,
The object feature point extraction step may include:
And a region of interest model generation step of generating a region of interest model for setting a search region of each feature point of the object,
The ROI model generating step may include: setting coordinates of a set number of feature points as a vector; Normalizing the center point of the coordinates so that the distance between two adjacent feature points is 1; And setting a region vector, which is calculated by a normalized average coordinate vector of feature points of each object and a changeable distance at an average position of the feature points, as a ROI model,
The object feature point extraction step may include:
Performing a fast Fourier transform on the search area; And a post-processing step of excluding, in the feature candidate region of the object, a portion of the spectral image processed through the Fast Fourier Transform having a coefficient value of a frequency component of an image less than or equal to a preset threshold value. Based object detection, recognition, and tracking methods.
9. The method of claim 8, wherein detecting and tracking the object further comprises:
An object detection, recognition, and tracking method based on context recognition, which detects and tracks objects using Bayesian classifiers based on 'Haar' features.
9. The method of claim 8,
And generating a detector for each feature point for detecting a feature point in the search region.
delete delete delete 9. The method of claim 8,
And an image preprocessing step of processing the input image so as to make the distribution of the lightness value uniform by the histogram smoothing method.
A program for executing the method according to any one of claims 8, 9, 10, and 14 is recorded.

Images