JP6493102B2 - Object recognition method, object recognition apparatus, and program - Google Patents

Description

  The present invention relates to an object recognition method, an object recognition apparatus, and a program for recognizing an object.

  An object recognition method is known in which a search space for a tree structure database is generated, a condition is determined at each node of the tree structure, a search is performed to reach each table of the tree structure database, and matching is performed ( For example, refer nonpatent literature 1).

Akiyama Mizuki and Yanai Keiji, general object recognition using a specific object recognition method using a large number of images on the Web, IPSJ SIG Technical Report

In the above method, since it is necessary to make a condition determination at each node of the tree structure, it takes a long time to search and there is a possibility that the recognition speed is lowered.
The present invention has been made in view of such problems, and mainly provides an object recognition method, an object recognition apparatus, and a program that enable high-speed object recognition while maintaining good recognition accuracy. Objective.

但し、Ｎ_ｉは前記探索空間のｉ番目（ｉ＝１〜ｎ：ｎは予め設定された木構造の階層数）のノードにおける前記特徴量テンプレートの数であり、｜Ｓ^Ｂ _Ｌ(Ｎ_ｉ)｜は、ｉ番目のノードから分岐し、左側に分配される前記特徴量テンプレートの数であり、｜Ｓ^Ｂ _Ｒ(Ｎ_ｉ)｜は、ｉ番目のノードから分岐し、右側に分配される前記特徴量テンプレートの数である。
この一態様において、下記評価関数の評価値が最小となるように、前記特徴量テンプレートにおけるビット位置を算出してもよい。

但し、ｘ、ｙは同一物体の特徴量を示し、ｑ_ｘ、ｑ_ｙは、その姿勢を示すクォータニオンであり、τは、予め設定される閾値であり、上記関数Ｐ（Ｎ）は、同一の物体で姿勢が近い場合に、ペナルティとして１を出力する関数である。
上記目的を達成するための本発明の一態様は、複数の物体の画像を取得する画像取得手段と、前記画像取得手段により取得された物体の画像の特徴量を抽出した特徴量テンプレートを生成し、該特徴量テンプレートをバイナリ化する特徴量抽出手段と、木構造データベースにハッシュテーブルを配置した探索空間を生成し、前記特徴量抽出手段により前記バイナリ化された特徴量テンプレートに基づいて、下記評価関数の評価値が最小となるように、該特徴量テンプレートにおけるビット位置を算出し、該算出したビット位置における前記特徴量テンプレートの値に基づいて前記ハッシュテーブルのハッシュ値を算出し、前記特徴量テンプレートのハッシュ値に基づいて、前記特徴量テンプレートを前記木構造データベースのハッシュテーブルに分配するテーブル分配手段と、前記画像取得手段により認識対象物の画像を取得しており、前記特徴量抽出手段により該取得された認識画像の特徴量を抽出した特徴量テンプレートを生成し該特徴量テンプレートをバイナリ化しており、該バイナリ化した特徴量テンプレートの前記ビット位置の値に基づいて、前記認識対象物の特徴量テンプレートのハッシュ値を算出し、該算出したハッシュ値に対応する前記木構造データベースのハッシュテーブルを選択し、該選択したハッシュテーブルに属する特徴量テンプレートと、前記認識対象物の特徴量テンプレートとのマッチングを行うマッチング手段と、を備える、ことを特徴とする物体認識装置であってもよい。

但し、Ｎ_ｉは前記探索空間のｉ番目（ｉ＝１〜ｎ：ｎは予め設定された木構造の階層数）のノードにおける前記特徴量テンプレートの数であり、｜Ｓ^Ｂ _Ｌ(Ｎ_ｉ)｜は、ｉ番目のノードから分岐し、左側に分配される前記特徴量テンプレートの数であり、｜Ｓ^Ｂ _Ｒ(Ｎ_ｉ)｜は、ｉ番目のノードから分岐し、右側に分配される前記特徴量テンプレートの数である。
上記目的を達成するための本発明の一態様は、複数の物体の画像を取得する処理と、前記物体の画像の特徴量を抽出した特徴量テンプレートを生成し、該特徴量テンプレートをバイナリ化する処理と、木構造データベースにハッシュテーブルを配置した探索空間を生成し、前記バイナリ化した特徴量テンプレートに基づいて、下記評価関数の評価値が最小となるように、該特徴量テンプレートにおけるビット位置を算出し、該算出したビット位置における前記特徴量テンプレートの値に基づいて前記ハッシュテーブルのハッシュ値を算出し、前記特徴量テンプレートのハッシュ値に基づいて、前記特徴量テンプレートを前記木構造データベースのハッシュテーブルに分配する処理と、認識対象物の画像を取得する処理と、前記認識対象物の画像の特徴量を抽出した特徴量テンプレートを生成し、該特徴量テンプレートをバイナリ化する処理と、前記バイナリ化した特徴量テンプレートの前記ビット位置の値に基づいて、前記認識対象物の特徴量テンプレートのハッシュ値を算出する処理と、前記算出したハッシュ値に対応する前記木構造データベースのハッシュテーブルを選択し、該選択したハッシュテーブルに属する特徴量テンプレートと、前記認識対象物の特徴量テンプレートとのマッチングを行う処理と、をコンピュータに実行させる、ことを特徴とするプログラムであってもよい。

但し、Ｎ_ｉは前記探索空間のｉ番目（ｉ＝１〜ｎ：ｎは予め設定された木構造の階層数）のノードにおける前記特徴量テンプレートの数であり、｜Ｓ^Ｂ _Ｌ(Ｎ_ｉ)｜は、ｉ番目のノードから分岐し、左側に分配される前記特徴量テンプレートの数であり、｜Ｓ^Ｂ _Ｒ(Ｎ_ｉ)｜は、ｉ番目のノードから分岐し、右側に分配される前記特徴量テンプレートの数である。 In order to achieve the above object, according to one aspect of the present invention, a step of acquiring images of a plurality of objects, a feature amount template obtained by extracting feature amounts of the object images are generated, and the feature amount template is binarized. A search space in which a hash table is arranged in a tree structure database, and a bit position in the feature value template so that an evaluation value of the following evaluation function is minimized based on the binary feature value template And calculating a hash value of the hash table based on the value of the feature amount template at the calculated bit position, and using the hash value of the feature amount template to convert the feature amount template into a hash of the tree structure database Distributing to the table, obtaining an image of the recognition object, Generating a feature amount template obtained by extracting the feature amount of the image of the recognition target object, binarizing the feature amount template, and based on the value of the bit position of the binarized feature amount template, the recognition target object Calculating a hash value of the feature amount template of the image, selecting a hash table of the tree-structure database corresponding to the calculated hash value, a feature amount template belonging to the selected hash table, and a feature of the recognition target object And a step of performing matching with a quantity template.

Here, N _i is the number of the feature amount templates in the i-th node (i = 1 to n: n is a preset number of hierarchies of the tree structure) in the search space, and | S ^B _L (N _i ) | branches from the i-th node, the number of the feature amount template is distributed to the ^{_{left, | S B R (N i}} ) | , the which branched from i-th node, is distributed to the right This is the number of feature quantity templates.
In this aspect, the bit position in the feature amount template may be calculated so that the evaluation value of the following evaluation function is minimized.

However, x and y indicate feature quantities of the same object, q _x and q _y are quaternions indicating their postures, τ is a preset threshold value, and the function P (N) is the same This is a function that outputs 1 as a penalty when the posture is close to an object.
In order to achieve the above object, one aspect of the present invention is to generate an image acquisition unit that acquires images of a plurality of objects, and a feature amount template that extracts the feature amounts of the image of the object acquired by the image acquisition unit. A feature amount extracting unit that binarizes the feature amount template, and a search space in which a hash table is arranged in a tree structure database, and the following evaluation is performed based on the binary feature amount template by the feature amount extracting unit: A bit position in the feature quantity template is calculated so that the evaluation value of the function is minimized, a hash value of the hash table is calculated based on a value of the feature quantity template at the calculated bit position, and the feature quantity Based on the hash value of the template, the feature amount template is stored in the hash table of the tree structure database. The image of the recognition object is acquired by the table distribution unit that distributes to the image and the image acquisition unit, and the feature amount template is generated by extracting the feature amount of the acquired recognition image by the feature amount extraction unit. A quantity template is binarized, a hash value of the feature quantity template of the recognition target is calculated based on the value of the bit position of the binarized feature quantity template, and the tree corresponding to the calculated hash value An object recognition apparatus comprising: a hash table of a structure database; and a matching unit that matches a feature amount template belonging to the selected hash table and a feature amount template of the recognition target object. There may be.

Here, N _i is the number of the feature amount templates in the i-th node (i = 1 to n: n is a preset number of hierarchies of the tree structure) in the search space, and | S ^B _L (N _i ) | branches from the i-th node, the number of the feature amount template is distributed to the ^{_{left, | S B R (N i}} ) | , the which branched from i-th node, is distributed to the right This is the number of feature quantity templates.
In order to achieve the above object, according to one aspect of the present invention, a process for acquiring images of a plurality of objects, a feature amount template obtained by extracting feature amounts of the object images are generated, and the feature amount template is binarized. A search space in which a hash table is arranged in the tree structure database is generated, and the bit position in the feature value template is set based on the binary feature value template so that the evaluation value of the following evaluation function is minimized. And calculating a hash value of the hash table based on the value of the feature amount template at the calculated bit position, and using the hash value of the feature amount template to convert the feature amount template into a hash of the tree structure database A process of distributing to a table, a process of acquiring an image of a recognition object, and an image of the recognition object A feature amount template extracted from the feature amount template, and the binarization of the feature amount template, and the value of the bit position of the binarized feature amount template, the feature amount template of the recognition target object A process of calculating a hash value, a hash table of the tree structure database corresponding to the calculated hash value is selected, and a feature amount template belonging to the selected hash table is matched with a feature amount template of the recognition target object A program characterized by causing a computer to execute the process of performing the above.

Here, N _i is the number of the feature amount templates in the i-th node (i = 1 to n: n is a preset number of hierarchies of the tree structure) in the search space, and | S ^B _L (N _i ) | branches from the i-th node, the number of the feature amount template is distributed to the ^{_{left, | S B R (N i}} ) | , the which branched from i-th node, is distributed to the right This is the number of feature quantity templates.

  The present invention has been made in view of such problems, and can provide an object recognition method, an object recognition apparatus, and a program that enable high-speed object recognition while maintaining good recognition accuracy. .

1 is a block diagram showing a schematic system configuration of an object recognition device according to Embodiment 1 of the present invention. It is a figure for demonstrating the offline process of an object recognition apparatus. It is a figure which shows an example of the feature-value template binarized. It is a figure which shows an example of the database of a tree structure. It is a figure for demonstrating the online process of an object recognition apparatus. It is a flowchart which shows the processing flow of the object recognition method which concerns on Embodiment 1 of this invention. It is a figure which shows the relationship between the recognition accuracy by RBS, PBS, TBS, and TBV and the number of recognition objects. It is a figure which shows the relationship between the recognition accuracy by RBS, PBS, TBS, and TBV and the number of recognition objects. It is a figure which shows the relationship between the recognition time by RBS, PBS, TBS, and TBV and the number of recognition objects. It is a figure which shows the relationship between the recognition time by RBS, PBS, TBS, and TBV and the number of recognition objects. It is a figure for demonstrating the structure which changes a hash function according to the size of a feature-value template.

Embodiment 1
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
The object recognition apparatus according to the first embodiment of the present invention recognizes a two-dimensional or three-dimensional object (item and posture) by performing template matching.

  Here, when matching is simply performed between the image of all models and the image of the object to be recognized, the search time becomes long and the real-time property is lacking. Therefore, the object recognition apparatus according to the first embodiment designs a tree-structured search space that achieves both high speed and high accuracy, and recognizes a feature amount template of a model image (hereinafter, model image) in this search space. Matching is performed with a feature amount template of an object image (hereinafter, a recognized object image). Thereby, high-speed object recognition becomes possible.

  The object recognition device includes, for example, a CPU (Central Processing Unit) that performs arithmetic processing, a ROM (Read Only Memory) and a RAM (Random Access Memory) that store arithmetic programs executed by the CPU, The hardware configuration is centered on a microcomputer including an interface unit (I / F) for inputting and outputting signals. The CPU, memory, and interface unit are connected to each other via a data bus or the like.

  For example, in the offline process, the object recognition apparatus according to the first embodiment generates a feature amount template from a model image, distributes the generated feature amount template to a hash table of a tree-structured database, and associates them. Then, in the online process, the object recognition device generates a feature amount template from the recognized object image, matches the feature amount template with the feature amount template of the hash table of the tree structure database, and recognizes the object. To do.

First, the online process described above will be described.
FIG. 1 is a block diagram illustrating a schematic system configuration of the object recognition apparatus according to the first embodiment. The object recognition device 1 according to the first embodiment includes an image acquisition unit 2, a feature amount extraction unit 3, a template distribution unit 4, and a matching unit 5.

  The image acquisition unit 2 is a specific example of an image acquisition unit. As shown in FIG. 2, the image acquisition unit 2 acquires model images of a plurality of types of objects (items) that are models using a camera (a). The image acquisition unit 2 uses a camera to acquire model images of different postures obtained by photographing each model from a plurality of different directions.

  The feature quantity extraction unit 3 is a specific example of feature quantity extraction means. The feature amount extraction unit 3 extracts a feature amount from the model image acquired by the image acquisition unit 2, and generates a feature amount template (b). Further, the feature quantity extraction unit 3 binarizes the generated feature quantity template.

  The feature amount extraction unit 3 generates a feature amount template from which the LINEMOD feature amount is extracted from the model image acquired by the image acquisition unit 2, for example, and binarizes the feature amount template.

  The feature amount extraction unit 3 extracts a LINEMOD feature amount as an example, but is not limited thereto. The feature quantity extraction unit 3 may extract, for example, HOG (Histograms of Oriented Gradients) feature quantity, SIFT feature quantity (Scale-Invariant Feature Transform), SURF feature quantity (Speeded Up Robust Features), and any feature quantity May be extracted.

  The template distribution unit 4 is a specific example of a template distribution unit. The template distribution unit 4 generates a search space in which a hash table is arranged in the tree structure database. For example, the template distribution unit 4 generates a three-level tree structure database as shown in FIG. 2 (c). Note that the number of hierarchical levels of the tree-structured database is set in advance by the user. For example, when the number of hierarchies is increased, as will be described later, the number of feature amount templates in the hash table is narrowed down, and the number of matching is reduced, so that the recognition speed is improved.

  Based on the binarized feature value template, the template distribution unit 4 calculates the bit position (for example, (A), (I)) in the feature value template so that the evaluation value of the evaluation function becomes the minimum ( b).

  Note that the number of calculated bit positions depends on the set number of layers. For example, comparing the object recognition speed, the accuracy of the recognized object image (input noise), the required object recognition accuracy, etc., and the number of feature amount templates to be matched (to be described later) Set the number of templates) experimentally. In order to increase the recognition speed, the number of feature amount templates to be matched is reduced. In addition, when the accuracy of the recognized object image is low (the noise is large) or when it is desired to increase the object recognition accuracy, the number of feature amount templates to be matched is increased. Then, the number of hierarchies of the tree structure is determined so that the set number of feature amount templates are distributed to the hash table. If the number of hierarchies of the tree structure is determined, the number of bit positions is determined according to the number of hierarchies. The number of bit positions to be calculated is obtained by subtracting 1 from the determined number of hierarchies. In the case of (c) in FIG. 2, since there are three layers, the number of bit positions is two.

なお、上記式において、Ｎ_ｉは探索空間のｉ番目（ｉ＝１〜ｎ：ｎは予め設定された木構造の階層数）のノードにおける特徴量テンプレートの数であり、｜Ｓ^Ｂ _Ｌ(Ｎ_ｉ)｜は、ｉ番目のノードから分岐し、左側に分配される特徴量の数であり、｜Ｓ^Ｂ _Ｒ(Ｎ_ｉ)｜は、ｉ番目のノードから分岐し、右側に分配される特徴量の数である。 For example, the template distribution unit 4 calculates the bit position in the feature amount template using the following evaluation function for distributing the feature amount template so that the number of feature amount templates is uniformly closest in each hash table. To do.

In the above equation, N _i is the number of feature quantity templates at the i-th node (i = 1 to n: n is a preset number of tree structure hierarchies) in the search space, and | S ^B _L (N _i) | is branched from the i-th node, the number of feature amounts to be distributed to the ^{_{left, | S B R (N i}} ) | is branched from the i-th node, the features are distributed to the right Is the number of quantities.

  Minimizing the evaluation value of the evaluation function means that the difference between the number of feature amount templates distributed on the left side and the number of feature amount templates distributed on the right side from each node is minimum (for example, approximately 0). Is to do. That is, by minimizing the evaluation value of the evaluation function, the feature amount templates can be distributed so that the number of feature amount templates is uniformly closest in each hash table. In the case shown in FIG. 2C, 1000 feature value templates are distributed to 500 feature value templates on the left side and the right side in the first layer node. Further, at the second layer node, the feature amount templates are distributed to 250 feature amount templates on the left side and the right side. In each hash table, 250 feature amount templates are finally uniformly distributed.

Next, a method for calculating the bit position in the feature amount template so that the evaluation value of the evaluation function is minimized will be described.
As shown in FIG. 3, the template distribution unit 4 includes a cell column position (r), a cell row position (c), and a position (b) in each cell of the binarized feature amount template (consisting of a plurality of cells). ) (Hereinafter, the positions specified by r, c, and b are referred to as bit positions), the evaluation value of the evaluation function is calculated, and the combination of the bit positions that minimizes the evaluation value is calculated. Note that the number of bit positions to be combined is set in advance according to the number of hierarchies in the tree-structured database as described above.
The template distribution unit searches for a bit position on the feature amount template by changing r, c, and b, and calculates a combination of bit positions that minimizes the evaluation value of the evaluation function.

  For example, as shown in FIG. 4, when the number of hierarchies in a tree-structured database is four, for example, the number of bit positions to be combined is 3 bits, and three bit positions are calculated. When there are 1000 feature value templates in the first layer node, the template distribution unit 4 searches for bit positions divided into 500 feature value templates in the first layer node using the evaluation function. More specifically, the template distribution unit 4 has 500 feature value templates whose bit position values are 0 (branch to the left) and 1 bit positions (branch to the right). The number of feature quantity templates is 500, and search is made for bit positions that are equal.

  Hereinafter, similarly, the template distribution unit 4 searches for bit positions divided into 250 feature amount templates at each node in the second hierarchy using the evaluation function. Further, the template distribution unit 4 searches for bit positions that are divided into 125 feature value templates at each node in the third hierarchy using the evaluation function.

  As described above, each node of the tree structure branches to the right side or the left side depending on the value of the bit position (0 or 1). Therefore, the 1000 feature value templates are finally divided into 125 feature value templates, which are distributed and associated with each hash table.

  The template distribution unit 4 calculates a hash value of the hash table based on the feature value template value at the calculated bit position. The hash value is, for example, a combination of values at each bit position of the feature amount template (arranged in order). In the example of FIG. 4, the hash value of the fourth hash table from the left is 011. Note that the template distribution unit 4 may calculate a hash value using a predetermined hash function based on the value of each bit position of the feature amount template.

  The template distribution unit 4 distributes each feature value template to the hash table of the tree structure database based on the hash value of the feature value template of the model image. The template distribution unit 4 calculates a hash value of the feature amount template from the value of the bit position in the feature amount template of each model image. Then, the template distribution unit 4 distributes and associates the feature amount template of each model image with the hash table of the tree structure database that matches the calculated hash value.

  As described above, in the object recognition apparatus 1 according to the first embodiment, the bit position in the feature amount template is calculated so that the evaluation value of the evaluation function is minimized, and the value of the feature amount template at the calculated bit position is calculated. The hash value of the hash table is calculated based on the feature value, and the feature value template is distributed to the hash table of the tree structure database based on the hash value of the feature value template. Thereby, each feature amount template can be distributed so that the number of feature amount templates of the model image is uniformly closest in each hash table. That is, the search range of the feature amount template of the model image can be narrowed uniformly in each hash table. Therefore, the feature amount template of the recognition object in any posture can be matched with the feature amount template of the model image of the hash table in which the number is uniformly narrowed down. Therefore, high-speed object recognition is possible while maintaining good recognition accuracy.

Then, the process of the offline process mentioned above is demonstrated.
The image acquisition unit 2 acquires a recognition object image using a camera. For example, as shown in FIG. 5A, the image acquisition unit 2 cuts out an image of a predetermined area including the object from the acquired recognized object image, and the recognized object image of the predetermined area is transferred to the feature amount extraction unit 3. Output.

  The feature quantity extraction unit 3 extracts a feature quantity from the recognized object image output from the image acquisition unit 2 and generates a feature quantity template. Further, the feature quantity extraction unit 3 binarizes the generated feature quantity template (b).

  The matching unit 5 is a specific example of matching means. For example, the matching unit 5 arranges the values (10, 00, 11, 11) of the bit positions (mark portions of (c)) of the calculated feature amount template and calculates the hash value (100001111) (d ). The matching unit 5 selects a hash table of the tree structure database corresponding to the calculated hash value (e).

  In this way, the matching unit 5 calculates a hash value from the bit position of the feature amount template of the recognized object image, and can immediately select a hash table of a tree-structured database using this hash value. Therefore, it is not necessary to determine the condition of each node of the tree structure, which leads to a significant reduction in search time and enables high-speed object recognition.

  Further, a feature amount template whose number is previously narrowed down is associated with the hash table. For this reason, the matching unit 5 only needs to perform matching on the narrowed feature amount template. Therefore, the matching time is greatly shortened, and higher-speed object recognition is possible. For example, in the case of an 8-level tree-structured database (7-bit hash), the number of feature amount templates can be narrowed down from 5000 to 30 and associated with each hash table. In this case, since it is only necessary to perform matching for 30 feature amount templates, this significantly reduces the matching time.

  The matching unit 5 performs matching between the feature amount template of the model image belonging to the selected hash table and the feature amount template of the recognized object image (f). For example, the matching unit 5 calculates a correlation value between the feature value template of each model image and the feature value template of the recognized object image. Then, the matching unit 5 recognizes the feature amount template of the model image having the highest calculated correlation value as a recognition object, and outputs the object and posture.

  Note that the matching unit 5 may end the recognition process when the calculated correlation value is equal to or less than a predetermined value and the correlation is low (no matching). Alternatively, when the calculated correlation value is equal to or lower than the predetermined value and the correlation is low, the matching unit 5 may perform matching between the feature amount template of the model image belonging to another hash table and the feature amount template of the recognized object image again. Good.

Next, the object recognition method according to the first embodiment will be described. FIG. 6 is a flowchart showing a processing flow of the object recognition method according to the present embodiment.
In the offline process, the image acquisition unit 2 acquires a model image of an object using a camera (step S101).

  The feature amount extraction unit 3 extracts a feature amount from the model image acquired by the image acquisition unit 2, and generates a feature amount template (step S102). The feature quantity extraction unit 3 binarizes the generated feature quantity template (step S103).

The template distribution unit 4 generates a search space in which hash tables are arranged in a tree structure database set in advance (step S104).
The template distribution unit 4 calculates the bit position in the feature value template based on the feature value template binarized by the feature value extraction unit 3 so that the evaluation value of the evaluation function is minimized (step S105). .

  The template distribution unit 4 calculates the hash value of each feature value template, and distributes and associates each feature value template with the hash table of the tree structure database that matches the calculated hash value (step S106).

In the online process, the image acquisition unit 2 acquires a recognized object image using a camera.
The feature amount extraction unit 3 extracts a feature amount from the recognized object image acquired by the image acquisition unit 2 (step S107), and generates a feature amount template (step S108). Further, the feature quantity extraction unit 3 binarizes the generated feature quantity template (step S109). The matching unit 5 arranges the calculated bit position values of the feature amount template to calculate a hash value (step S110). The matching unit 5 selects a hash table of the tree structure database corresponding to the calculated hash value (step S111).

The matching unit 5 performs matching between the feature amount template of the model image belonging to the selected hash table and the feature amount template of the recognized object image (step S112).
The matching unit 5 recognizes the feature amount template of the model image having the highest correlation value with the feature amount template of the recognized object image as a recognition object, and outputs the object and posture (step S113).

As described above, in the first embodiment, the search space in which the hash table is arranged in the tree structure database is generated, and the evaluation value of the evaluation function is minimized based on the feature amount template of the model image binarized. A combination of bit positions in the feature amount template is calculated, a hash value of the hash table is calculated based on a value of the feature amount template in the calculated combination of bit positions, and a feature amount template is calculated based on the hash value of the feature amount template Is distributed to the hash table of the tree structure database. As a result, the number of feature amount templates is uniformly closest in each hash table, and the hash value of each hash table is calculated based on the value of the feature amount template in the calculated bit position combination. Furthermore, the hash value of the feature amount template of the recognition target object is calculated based on the bit position value of the feature amount template of the recognized recognition object image, and the hash table of the tree structure database corresponding to the calculated hash value is selected. Then, the feature amount template belonging to the selected hash table is matched with the feature amount template of the recognition target object. Thereby, a hash value is calculated from the bit position of the feature amount template of the recognized object image, and a hash table of a tree-structured database can be immediately selected using this hash value. Therefore, it is not necessary to determine the condition of each node in the tree structure, which leads to reduction in search time and enables high-speed object recognition. Further, a feature amount template whose number is previously narrowed down is associated with the hash table. For this reason, the matching unit only needs to perform matching on the narrowed-down feature amount template. Therefore, matching time is shortened, and high-speed object recognition is possible.
Furthermore, each feature quantity template can be distributed so that the number of feature quantity templates of the model image is uniformly closest in each hash table. That is, the search range of the feature amount template of the model image can be narrowed uniformly in each hash table. Therefore, the feature amount template of the recognition object in any posture can be matched with the feature amount template of the model image of the hash table in which the number is uniformly narrowed down. Therefore, high-speed object recognition is possible while maintaining good recognition accuracy.

但し、ｘ、ｙは同一物体の特徴量を示し、ｑ_ｘ、ｑ_ｙは、その姿勢を示すクォータニオンである。τは、予め実験的に求め設定される閾値であり、例えば、０．３が設定されている。上記関数Ｐ（Ｎ）は、同一モデルで姿勢が類似している場合に、ペナルティとして１を出力する関数である。 Embodiment 2
In Embodiment 2 of the present invention, for example, the template distribution unit 4 distributes feature amount templates so that the number of feature amount templates is uniformly closest in each hash table, and the feature amount of a similar posture in the same model. The bit position in the feature amount template is calculated using the following evaluation function that does not distribute the template to the same hash table.

However, x, y denotes the characteristic quantity of the same _object, q x, _{q y} is a quaternion representing the posture. τ is a threshold that is experimentally determined and set in advance, and is set to, for example, 0.3. The function P (N) is a function that outputs 1 as a penalty when the posture is similar in the same model.

  In the evaluation function, the front term is the same as the evaluation function of the first embodiment, and is a term for distributing feature amount templates so that the number of feature amount templates is uniformly closest in each hash table. . Further, the backward term (term including the function P (N)) is a penalty term and prevents feature amount templates of similar poses in the same model from being distributed to the same hash table (so that model types can be scattered). It is a term for). With this term, feature amount templates of the same model and similar posture are distributed to different hash tables.

  For example, when the same model is in the same hash table, if only the bit positions referred to in the calculation of the hash value are viewed, those with the same model and similar attitude are likely to be distributed to the same hash table. On the other hand, in general, a better search result can be obtained by performing a global search from as many different players as possible than from a local search from similar ones. In the present embodiment, first, only the bit position referred to in the calculation of the hash value is viewed, the corresponding hash table is selected, and then all the feature amount templates belonging to the selected hash table and the features of the recognition target object are selected. Match with quantity template. For this reason, when the feature amount templates of similar postures in the same model are distributed to different hash tables and the model types are scattered, a better search result can be obtained and the object recognition accuracy is improved.

As described above, according to the second embodiment, by using the above evaluation function, the number of feature amount templates of the model image is uniformly closest in each hash table, and the feature amount template having the similar posture in the same model is used. The feature amount templates can be distributed so that they do not become the same hash table.
By distributing the feature amount templates so that the number of feature amount templates of the model image is uniformly closest in each hash table, high-speed object recognition is possible as in the first embodiment. Further, by distributing the feature amount templates so that the feature amount templates of similar postures in the same model do not become the same hash table, an object can be recognized with higher accuracy than in the first embodiment. That is, high-speed object recognition is possible while maintaining better object recognition accuracy.

Next, a simulation result obtained by recognizing an object using the object recognition apparatus according to the first and second embodiments described above will be described.
7 and 8 are diagrams showing the relationship between the recognition accuracy by RBS, PBS, TBS, and TBV and the number of recognized objects. 9 and 10 are diagrams showing the relationship between the recognition time and the number of recognition objects by RBS, PBS, TBS, and TBV.

  7 and 8, a monkey figurine (ape) is used as an example of a recognition object, and a lamp figurine (lamp) is used as an example of a recognition object in FIGS. 9 and 10. In RBS (Randomness-based selection), the bit position of the feature amount template is set at random. In PBS (Probability-based selection), the bit position of the feature amount template is set based on the entropy of the recognized object. In TBS (Tree-based selection), the bit position of the feature amount template is set using the evaluation function (1) according to the first embodiment. In TBV (Tree-based selection with view scattering), the bit position of the feature amount template is set using the evaluation function (2) according to the second embodiment. In this simulation, the bit position of the feature amount template is set as described above for RBS, PBS, TBS, and TBV, and matching is performed.

  As shown in FIGS. 7 and 8, both TBS (evaluation function (1)) and TBV (evaluation function (2)) according to the present embodiment are better recognized than other PBSs and RBSs. It can be seen that accuracy is obtained. It can also be seen that the TBV (evaluation function (2)) according to the second embodiment has higher recognition accuracy than the TBS (evaluation function (1)) according to the first embodiment.

  Furthermore, as shown in FIGS. 9 and 10, both TBS (Evaluation Function (1)) and TBV (Evaluation Function (2)) according to the present embodiment are significantly larger than other PBS and RBS. It can be seen that the recognition time is shortened and high-speed object recognition is possible. In particular, when the number (type) of recognition objects increases, in other PBSs and RBSs, the recognition time increases in proportion to the number, but the TBS (Evaluation Function (1)) and TBV (evaluation function (2)) is excellent in that the recognition time hardly changes even when the number of recognized objects increases.

  That is, from the simulation results shown in FIGS. 7 to 10, the object recognition apparatus 1 according to the first and second embodiments using TBS (evaluation function (1)) and TBV (evaluation function (2)) It can be seen that high-speed object recognition is possible while maintaining good recognition accuracy.

Note that the present invention is not limited to the above-described embodiment, and can be changed as appropriate without departing from the spirit of the present invention.
For example, in the above-described embodiment, the feature amount extraction unit 3 may extract a feature amount from the model image acquired by the image acquisition unit 2 and generate a feature amount template that is changed to a multistage size. As a result, even when the shooting range is changed, for example, when an object or model is viewed at a distance or close, the detection accuracy is improved by changing the size of the feature amount template corresponding to the shooting range. be able to.

  In this case, the template distribution unit 4 calculates a hash value using a hash function corresponding to the size of the feature amount template based on the value of each bit position of the feature amount template. For example, as shown in FIG. 11, the template distribution unit 4 performs the hash function A corresponding to the size 32 * 32, 48 * 48, 64 * 64 of the feature amount template based on the value of each bit position of the feature amount template. , B, and C are used to calculate hash values.

The present invention can also be realized, for example, by causing a CPU or a GPU (Graphics Processing Unit) to execute a computer program as shown in FIG.
The program may be stored using various types of non-transitory computer readable media and supplied to a computer. Non-transitory computer readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (for example, flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (for example, magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / W and semiconductor memory (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (random access memory)) are included.

  The program may also be supplied to the computer by various types of transitory computer readable media. Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves. The temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

DESCRIPTION OF SYMBOLS 1 Object recognition apparatus, 2 Image acquisition part, 3 Feature value extraction part, 4 Template distribution part, 5 Matching part

Claims (4)

Acquiring images of a plurality of objects;
Generating a feature amount template obtained by extracting the feature amount of the image of the object, and binarizing the feature amount template;
Generate a search space in which a hash table is arranged in a tree structure database, and based on the binary feature quantity template, calculate a bit position in the feature quantity template so that an evaluation value of the following evaluation function is minimized, A hash value of the hash table is calculated based on the value of the feature value template at the calculated bit position, and the feature value template is distributed to the hash table of the tree structure database based on the hash value of the feature value template And steps to
Obtaining an image of the recognition object;
Generating a feature amount template obtained by extracting the feature amount of the image of the recognition object, and binarizing the feature amount template;
Calculating a hash value of the feature amount template of the recognition target based on the value of the bit position of the binary feature amount template;
Selecting a hash table of the tree structure database corresponding to the calculated hash value, and matching the feature amount template belonging to the selected hash table with the feature amount template of the recognition target;
An object recognition method comprising:

Here, N _i is the number of the feature amount templates in the i-th node (i = 1 to n: n is a preset number of hierarchies of the tree structure) in the search space, and | S ^B _L (N _i ) | branches from the i-th node, the number of the feature amount template is distributed to the ^{_{left, | S B R (N i}} ) | , the which branched from i-th node, is distributed to the right This is the number of feature quantity templates. The object recognition method according to claim 1,
An object recognition method characterized by calculating a bit position in the feature amount template so that an evaluation value of the following evaluation function is minimized.

However, x and y indicate feature quantities of the same object, q _x and q _y are quaternions indicating their postures, τ is a preset threshold value, and the function P (N) is the same This is a function that outputs 1 as a penalty when the posture is close to an object. Image acquisition means for acquiring images of a plurality of objects;
A feature amount extraction unit that generates a feature amount template obtained by extracting the feature amount of the image of the object acquired by the image acquisition unit, and binarizes the feature amount template;
Generate a search space in which a hash table is arranged in a tree structure database, and based on the binary feature value template by the feature value extraction means, the feature value template so that the evaluation value of the following evaluation function is minimized The hash value of the hash table is calculated based on the value of the feature value template at the calculated bit position, and the feature value template is calculated based on the hash value of the feature value template. Table distribution means for distributing to the hash table of the structure database;
An image of a recognition object is acquired by the image acquisition unit, a feature amount template is generated by extracting the feature amount of the acquired recognition image by the feature amount extraction unit, and the feature amount template is binarized; Based on the value of the bit position of the binarized feature value template, the hash value of the feature value template of the recognition target is calculated, and the hash table of the tree structure database corresponding to the calculated hash value is selected. Matching means for matching the feature amount template belonging to the selected hash table and the feature amount template of the recognition target object;
An object recognition apparatus comprising:

Here, N _i is the number of the feature amount templates in the i-th node (i = 1 to n: n is a preset number of hierarchies of the tree structure) in the search space, and | S ^B _L (N _i ) | branches from the i-th node, the number of the feature amount template is distributed to the ^{_{left, | S B R (N i}} ) | , the which branched from i-th node, is distributed to the right This is the number of feature quantity templates. Processing to acquire images of a plurality of objects;
Generating a feature amount template obtained by extracting the feature amount of the image of the object, and binarizing the feature amount template;
Generate a search space in which a hash table is arranged in a tree structure database, and based on the binary feature quantity template, calculate a bit position in the feature quantity template so that an evaluation value of the following evaluation function is minimized, A hash value of the hash table is calculated based on the value of the feature value template at the calculated bit position, and the feature value template is distributed to the hash table of the tree structure database based on the hash value of the feature value template Processing to
Processing to acquire an image of the recognition object;
Generating a feature amount template obtained by extracting the feature amount of the image of the recognition target object, and binarizing the feature amount template;
A process of calculating a hash value of the feature amount template of the recognition target based on the value of the bit position of the binary feature amount template;
A process of selecting a hash table of the tree structure database corresponding to the calculated hash value, and performing a matching between a feature value template belonging to the selected hash table and a feature value template of the recognition target object;
A program characterized by causing a computer to execute.

Here, N _i is the number of the feature amount templates in the i-th node (i = 1 to n: n is a preset number of hierarchies of the tree structure) in the search space, and | S ^B _L (N _i ) | branches from the i-th node, the number of the feature amount template is distributed to the ^{_{left, | S B R (N i}} ) | , the which branched from i-th node, is distributed to the right This is the number of feature quantity templates.

Images