KR20090113553A - Semantic active object recognition - Google Patents

Abstract

PURPOSE: A semantic active object algorithm is provided to increase the recognition rate of a targeted object although prominent features of the object are difficult to be recognized. CONSTITUTION: An object recognition method through an input image from a camera comprises the following steps of: preparing many-sided images of each object; calculating/pre-storing each entropy of the many-sided images; determining a target object; determining an estimation region estimated as the target object from the input image; determining whether the estimation region is the target object; if not, determining one side of the target object corresponding to the estimation region; and moving the focus to a position where a many-sided image having lower entropy can be obtained based on the side of the target object.

Description

Active Object Recognition Algorithm Combining Knowledge {SEMANTIC ACTIVE OBJECT RECOGNITION}

The present invention relates to an object recognition algorithm using an image. More specifically, as an object recognition algorithm, in order to increase the accuracy of object recognition in a complex environment, a top-down / down-level visual focusing technique and an information amount based next viewpoint determination technique are developed, and object recognition is performed. A method of applying to an algorithm.

The present invention relates to an object game of a mobile robot, a person recognition of a mobile robot, a surveillance camera, a mobile phone camera, a video camera, a computer game, a video chat, a computer USB camera, a digital camera, a mobile robot camera, and an interactive game using a digital camera. Can be applied.

Conventional object recognition algorithms have mainly developed algorithms for extracting robust visual features. Robust visual features mean that visual features extracted when an object is standing upright are extracted from the same location even if the object is subject to various external influences such as rotation, size change, and brightness change. Recognize objects based on these powerful visual features.

 Conventionally, object recognition is performed by obtaining a visual feature from a single image and comparing it with visual features of previously stored objects. However, as illustrated in FIG. 1, there are cases where object recognition is required in various situations in which there are several objects on the screen, the objects to be searched a little, or the background is complicated. The reliability of the result of recognizing the object is inevitably deteriorated. Therefore, by taking several consecutive pictures and using the continuous data to recognize the object, the object recognition is successful and the reliability of the recognition result can be improved.

An object of the present invention is to provide a method for increasing the recognition rate of a target object and a robot using the same even when it is difficult to recognize the salient features of the object.

 In order to increase the accuracy of object recognition, it is necessary to acquire a lot of data by taking several pictures. This is the same principle that a person looks closely at to see which object fits. In order for this operation to be algorithmically possible, two more techniques must be added to the conventional object recognition technique. One is a top-down / bottom-up visual attention technique that considers the object to be searched in the input image, taking into account the knowledge of the object and its own visual characteristics, and the other is entropy to move to the point of view for better recognition of the object. It is a technology that recommends the following actions to move to the highest information level based on The addition of these techniques can increase the accuracy of object recognition.

In the present invention, to improve the accuracy of object recognition in the object recognition algorithm, an improved object recognition algorithm using the visual focusing technique and the next viewpoint determination technique is proposed. The present invention is a top-down / bottom-up visual focusing technique for obtaining a portion corresponding to an object to be recognized in visual information from a digital camera, a technique for moving a viewpoint based on an amount of information to increase the recognition rate of an object to be recognized, and an object. It consists of a recognition algorithm. Top-down / bottom-up visual focusing technology combines top-down information based on the appearance of characteristic surface information of objects and places where they were placed, and bottom-up information based on the emergence of visual information obtained from images taken with a camera. The technique of shifting the view point of an object based on the amount of information calculates the amount of information of the object and moves it to the plane of the object having a large amount of information. In addition, this change of information amount is combined with behavior to recommend the most appropriate behavior to raise the recognition rate of the object.

It can be applied to the existing object recognition algorithm and can improve the object recognition performance.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention is based on the premise of the basic configuration of the conventional robot that is basically provided with an image device for imaging the image, a driving device for movement, a memory for storing the image, a processor for analyzing the image data, and the like. The algorithm will be described based on the algorithm that can increase the recognition rate of objects.

2 is a flowchart for performing schematic object recognition according to an embodiment of the present invention.

First, the image data captured by the robot's imaging device is input.

Next, the image focuses on the image suspected of being the target object in the image data in a bottom-up and top-down manner.

Here, the bottom-up and top-down visual focusing scheme will be described with reference to FIG. 3.

To do this, the robot goes through the first process of finding an object by visually focusing on the inherent visual characteristics of the object to be found. Visual characteristics include the inherent color, contrast, dominant color, contour, orientation, and so on.

After finding the object once, it recognizes the place where the object was and the object around it and stores it together.

Now, when we need to find this object again, we use the information about the place where the object was located and the information about the object around it to obtain a top-down visual focus that predicts where the object is located, Concentrate time by combining bottom-up methods that use raised visual information.

That is, as shown in FIG. 3, the object context may be stored as meaningful knowledge information, and the place information, recognition information of surrounding objects, distance and azimuth information of surrounding objects, and the like are stored.

This top-down method can be used as a very useful information when an object is hidden by a certain object or when there are many objects mixed around it, and it is difficult to distinguish a target object into an image. Therefore, it is preferable to rely on the visual characteristics intrinsic to the object, that is, to concentrate the vision in a bottom-up method and to parallel the vision in a top-down method.

In this way, objects are recognized through extracted and focused images. If the focused image cannot confirm whether the target object is correct, a new image that can confirm this must be acquired.

In particular, the new video should be data that can lead to the right judgment. To this end, it is necessary to reasonably select a time point at which image data is acquired.

To this end, the present invention takes a point determination method based on entropy. This method will be described with reference to the drawings schematically shown in FIG.

Prior to this, the robot has previously stored the front, rear, left, right, up and down images of the object. In addition, the front, rear, left, right, up and down images of all the objects of interest are stored in advance. The entropy of each image is calculated and stored in advance for the probability that each of these images will be the decisive clue for identifying the object and the probability of misjudgment with another object.

It determines whether the image of the object included in the image data currently captured corresponds to any of the six surfaces, and moves the robot to a point in time at which an image having a lower entropy can be obtained, or determines such an operation.

That is, based on the concentrated image portion of the image data currently captured, the best robot traveling direction for determining the viewpoint of the robot for obtaining reliable image data is determined.

4 is a schematic representation of this process. Referring to FIG. 4, when it is determined that the object is the front surface through SIFT, CCH (Color Coocurrence Histgram), DC (Dominant Color) algorithm, etc., the next time point is determined in a direction of lowering entropy to determine an action. will be.

After the movement, as shown in Fig. 2, the image data is acquired again and the processes are repeated.

Finally, when it is determined that the target object matches the image data, the loop is terminated.

FIG. 5 is a screen in which images used are activated in respective windows when an image is captured and an algorithm of the present invention is applied.

As illustrated in FIG. 5, when the "camera input" image is divided based on the outline and the color, the "segmented" image may be obtained. When the "object model" is selected as the cup, a "visually focused" image can be obtained in a top-down / bottom-up manner. Now the path, or “recommended behavior,” for moving the robot in a direction that makes the cup more recognizable is shown graphically as directional.

6 is a user interface screen for selecting a bottom-up / downward function for visual focus processing. Here, it is also possible to select a feature of the cup to be used for bottom-up visual focusing, such as color, brightness, directionality, and the like.

7 is a user interface window for selecting an algorithm to apply to an object image. Referring to FIG. 7, the right screen of the object image is centered on the dominant color, and “feature 2” at the bottom represents the SIFT screen. The screen on the left shows the entropy map, which maps the entropy of the colors used to identify objects in the object list.

Although some embodiments of the present invention have been shown and described so far, it will be apparent to those skilled in the art that the present invention may be modified without departing from the principles or spirit of the invention. There will be. It is intended that the scope of the invention be defined by the claims appended hereto and their equivalents.

1 is a photograph showing a situation in which an object to be searched is mixed with various objects.

2 is a flowchart for performing schematic object recognition according to an embodiment of the present invention.

3 is a schematic diagram for explaining a bottom-up / downward manner.

FIG. 4 is a diagram illustrating a behavior determination method for obtaining a next point of view based on an appearance of an object included in currently captured image data and a corresponding entropy.

FIG. 5 is a screen in which images used are activated in respective windows when an image is captured and an algorithm of the present invention is applied.

6 is a user interface screen for selecting a bottom-up / downward function for visual focus processing.

7 is a user interface window for selecting an algorithm to apply to an object image.

Claims (5)

In the object recognition method through the input image from the camera, (a) preparing a multi-sided image of an object viewed from various angles for each object of interest, and calculating and storing each entropy of the multi-sided images in advance; (b) determining a target object; (c) determining an estimated area estimated from the input image as the target object; (d) determining whether or not the estimation region represents the target object; (e) if the estimation region is not determined as the target object, determining which side of the target object corresponds to the estimation region; (f) moving the viewpoint to a position from which a multi-sided image having a low entropy can be obtained among the multi-faceted images of the target object with respect to the plane of the target object indicated by the estimation region. Object recognition method. The method of claim 1, And the entropy is determined based on a probability of misjudging from the multi-faceted image of the object of interest to another object. The method of claim 1, (g) if it is determined in step (d) that the estimation region represents the target object, storing peripheral information of the estimation region as contextual information associated with the target object; Robot object recognition method. The method of claim 1, (c) determining the estimated area estimated as the target object from the image input from the camera, And evaluating and reflecting the degree of agreement between the contextual information associated with the target object and the surrounding information of the estimated area. The method according to claim 3 or 4, The situation information includes place information, identification information of surrounding objects, distance information with respect to surrounding objects, and orientation information.

Images