JP2011113360A - Object identification apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an object identification apparatus that improves accuracy in identification. <P>SOLUTION: The object identification apparatus 1 is used for identifying objects to be identified form determined images. The object identification apparatus includes a partial feature evaluation value calculation unit 32, that evaluates feature amount of the decision images based on the partial features, calculates an evaluation value for each partial feature, and acquires weight of reliability associated with the partial features based on the evaluation value; a partial feature weight calculating unit 33 that calculates the total weight that is the total weight of acquired reliability; a unit 34 for calculating a combination pattern matching that calculates a combination matching degree that is the matching degree in the partial features whose evaluation value is determined as being a predetermined one or higher, and in the partial features related to the objects preliminarily learned to be identified; and an identified result decision unit 35 that determines the objects to be identified from the determined images, when the total weight is the predetermined value or larger, and the combined matching degree is the predetermined value or larger. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an object identification device.

  Conventionally, an apparatus that uses image information is known as an apparatus for identifying an object to be identified (see, for example, Patent Document 1). The object identification device described in Patent Literature 1 employs an AdaBoost Algorithm and evaluates a plurality of local partial features of a determination image, thereby identifying an identification target object in the determination image. Specifically, a plurality of weak classifiers suitable for object detection are generated while learning based on the partial features of the objects in the learning data, and each of the generated weak classifiers is represented with a reliability (degree of identification target object-like degree). ). Then, the determination image is evaluated by the weak classifier, and the identification target object is identified by linearly combining the evaluation results of the plurality of weak classifiers.

JP 2008-140230 A

  By the way, in the conventional object identification device adopting the AdaBoost algorithm, it is determined only whether the sum of the weights assigned to the weak classifier that has detected the partial feature exceeds the threshold, thereby identifying the object to be identified. Has been identified. For this reason, in the conventional object identification device, for example, another object having only similar local partial features is identified as an identification target object, or the identification target object whose sum of weights is slightly less than the threshold value May not be identified.

  Accordingly, the present invention has been made to solve such a technical problem, and an object thereof is to provide an object identification device capable of improving identification accuracy.

  In other words, the object identification device according to the present invention can identify an object to be identified from a determination image based on feature amounts of a plurality of partial features learned in advance using a learning image and reliability weights associated with the partial features. An evaluation value calculating means for evaluating a feature amount of the determination image based on the feature amount of the partial feature and calculating an evaluation value for each partial feature, and based on the evaluation value Acquisition means for acquiring reliability weights associated with the partial features, weight sum calculation means for calculating a weight sum that is a sum of the reliability weights acquired by the acquisition means, and the evaluation value A degree of coincidence calculating means for calculating a degree of coincidence of matching between the combination of the partial features determined to be equal to or greater than a predetermined value and the combination of the partial features related to the identification target object learned in advance , The weight sum is equal to or higher than a predetermined value, and wherein when the combination coincidence degree is equal to or higher than the predetermined value, configured and a determining identification result determining means and identified the identification object from the determination image.

  In the object identification device according to the present invention, the evaluation value calculation unit evaluates the feature amount of the determination image based on the feature amount of the partial feature to calculate the evaluation value for each partial feature, and the acquisition unit calculates the evaluation value based on the evaluation value. Thus, the reliability weight associated with the partial feature is acquired, and the weight sum calculation means calculates a weight sum that is the sum of the reliability weights acquired by the acquisition means. Further, the degree of coincidence of the partial features determined to have an evaluation value equal to or greater than a predetermined value and the degree of coincidence between the combination of the partial features related to the identification target object learned in advance is calculated by the degree of coincidence calculation means. Is done. Then, when the sum of weights is equal to or greater than the predetermined value and the combination matching degree is equal to or greater than the predetermined value, it is determined by the identification result determination means that the identification target object has been identified from the determination image. Thus, in addition to the identification determination based on the sum of the weights of the partial features, the identification determination based on the matching degree of the partial feature combinations can be performed. For this reason, for example, even when the weight sum becomes a predetermined value or more because only the local partial feature of an object different from the identification target object is similar to the identification target object, It is possible to avoid identifying another object as an identification target object using the degree of coincidence. Therefore, identification accuracy can be improved.

  Here, the degree-of-match calculation means includes the partial features included in a combination of the partial features of the identification target object learned in advance among the partial features determined that the evaluation value is equal to or greater than a predetermined value. The number of the partial features having the same feature amount is calculated as the number of matches, the ratio of the number of matches to the number of the partial features included in the combination is calculated, and the combination matching degree is calculated based on the calculated ratio Is preferably calculated.

  With this configuration, the combination of partial features related to the identification target object learned in advance includes how many partial features are the same as the partial features determined to have an evaluation value equal to or greater than a predetermined value. It is possible to calculate a ratio indicating whether or not. Accordingly, it is possible to evaluate the degree of coincidence between the combination of the partial features of the determination image evaluated to match the learned partial features and the combination of the partial features of the learned identification target object.

  Alternatively, the degree-of-match calculation means is included in a combination of the partial features of the identification target object learned in advance among the partial features of the identification target object determined that the evaluation value is equal to or greater than a predetermined value. The number of the partial features having the same or similar feature quantity as the partial feature is calculated as the number of matches, the ratio of the number of matches to the number of the partial features included in the combination is calculated, and based on the calculated ratio It is preferable to calculate the combination matching degree.

  With this configuration, the combination of partial features related to the identification target object learned in advance includes how much partial features are the same or similar to the partial features determined to have an evaluation value equal to or greater than a predetermined value. It is possible to calculate a ratio indicating whether or not As a result, when the partial feature determined to have an evaluation value equal to or greater than the predetermined value is not the same as the partial feature included in the combination but is similar, the partial feature related to the identification target object learned in advance Can be included in the combination. The degree of coincidence between the combination of the partial feature of the determination image evaluated to match the learned partial feature and the combination of the partial feature of the learned identification target object by providing a range for determining whether or not they match Can be calculated including an error. Therefore, it is possible to avoid making the identification target object unidentified by strict determination.

  In addition, the object identification device includes weight sum correction means for adding and subtracting the weight sum based on the combination matching degree, and the identification result determination means is configured when the weight sum is corrected by the weight sum correction means. Preferably, when the weight sum corrected by the weight sum correcting means is a predetermined value or more and the combination matching degree is a predetermined value or more, it is determined that the identification target object is identified from the determination image. It is.

  With this configuration, the weight sum correction means can adjust the weight sum based on the combination matching degree, and can also perform identification determination using the corrected weight sum. Thereby, for example, it is possible to avoid that an identification target object whose sum of weights is slightly less than the threshold value cannot be identified. Therefore, identification accuracy can be improved.

  Furthermore, the feature amount of the partial feature may include at least one of a shape, a position, and a size.

  According to the present invention, it is possible to improve the identification accuracy.

It is a block diagram which shows the structure of the object identification apparatus which concerns on embodiment. It is a schematic diagram for demonstrating the partial feature which the object identification device which concerns on 1st Embodiment extracts. It is a flowchart which shows operation | movement of the object identification device which concerns on 1st Embodiment. It is a schematic diagram for demonstrating operation | movement of the object identification device which concerns on 1st Embodiment. It is a flowchart which shows the combination learning operation | movement of the object identification device which concerns on 1st Embodiment. It is a schematic diagram for demonstrating learning operation | movement of the object identification device which concerns on 1st Embodiment. It is a flowchart which shows combination combination degree calculation operation | movement of the object identification device which concerns on 1st Embodiment. It is a flowchart which shows the identification determination operation | movement of the object identification device which concerns on 1st Embodiment. It is a flowchart which shows combination matching degree calculation operation | movement of the object identification device which concerns on 2nd Embodiment. It is a flowchart which shows the identification determination operation | movement of the object identification device which concerns on 3rd Embodiment.

  Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In addition, in each figure, the same code | symbol is attached | subjected to the same or an equivalent part, and the overlapping description is abbreviate | omitted.

(First embodiment)
The object identification device according to the present embodiment is a device that identifies an object included in a determination image based on a result of learning a learning image, and is suitably used for driving assistance of a vehicle, for example. is there.

  First, the configuration of the object identification device according to the present embodiment will be described. FIG. 1 is a block diagram showing the configuration of the object identification device according to the present embodiment. As shown in FIG. 1, the object identification device 1 includes a recording device 2 and an arithmetic device 3.

  The recording device 2 is a device for recording information, and is configured to be readable and writable by the arithmetic device 3. For example, an HDD (Hard Disk Drive) is used as the recording device 2. The recording device 2 includes a learning image storage unit 20, a partial feature model storage unit 21, and a combination pattern storage unit 22.

  The learning image storage unit 20 stores a plurality of learning images that the arithmetic device 3 learns to identify the identification target object. As the learning image, not only an image including an identification target object to be identified but also an image including no identification target object is used. The partial feature model storage unit 21 stores a plurality of feature amounts of partial features acquired by the arithmetic device 3 through learning. A partial feature is a partial image that constitutes a part of an image, and indicates a partial feature that exhibits distinctiveness such as a person such as a head, hand, or foot if the object to be identified is a human being, for example. It is. The feature amount is an evaluation value for evaluating the feature and is calculated based on the image information of the partial feature. For example, feature values of Haar wavelets are used, and details will be described later. The combination pattern storage unit 22 stores a plurality of partial feature combination patterns learned by the arithmetic device 3. The partial feature combination pattern is a combination of a plurality of partial features related to the identification target object, and details will be described later.

  The arithmetic device 3 is a device that performs arithmetic processing, and is connected to a recording device 2, a determination image input unit 4 that inputs a determination image, and a determination result display unit 5 that displays a determination result. For example, an ECU (Electronic Control Unit) is used as the arithmetic device 3. The ECU is a computer that is electronically controlled and includes a CPU (Central Processing Unit), a memory such as a ROM (Read Only Memory) and a RAM (Random Access Memory), an input / output interface, and the like. Note that an imaging device such as a CCD (Charge Coupled Device) camera, for example, is used as the determination image input unit 4, and a display device such as a display is used as the determination result display unit 5, for example.

  The arithmetic device 3 includes a partial feature learning unit 30, a combination pattern learning unit 31, a partial feature evaluation value calculation unit (evaluation value calculation unit, acquisition unit) 32, a partial feature weight sum calculation unit (weight sum calculation unit) 33, a combination pattern A coincidence calculation unit (coincidence calculation unit) 34 and an identification result determination unit (identification result determination unit) 35 are provided.

  The partial feature learning unit 30 has a function of learning partial features of a learning image. The partial feature learning unit 30 refers to the learning image stored in the learning image storage unit 20 and learns the partial feature of the learning image. For example, the partial feature learning unit 30 learns using the Adaboost algorithm. The Adaboost algorithm is a learning algorithm that generates different weak classifiers while sequentially changing the weights of reliability regarding examples and combines them to form a high-precision classifier. The partial feature learning unit 30 acquires the feature amount of the Haar wavelet from the learning image using several types of filters. FIG. 2A is an example of the filters F1 to F4 that extract feature amounts. As shown in FIG. 2A, the filters F1 and F2 extract features that appear in the edge portion, and the filters F3 and F4 extract features that appear in the line portion. Filters F1 to F4 are binary rectangular filters of ± 1, and a value obtained by subtracting the sum of pixel values in a black region of −1 from the sum of pixel values (luminance values) in a white region of +1 Calculated as an evaluation value. The filters F1 to F4 are changed to arbitrary positions and sizes, and appropriately extract partial features from the learning image. In other words, the extracted partial features are expressed by feature amounts including information on position, size, and shape. FIG. 2B is a schematic diagram for explaining the positions and sizes of the partial features in the image G1. As shown in FIG. 2B, the position of the partial feature is expressed by, for example, the upper left position (u1, v1) of the filter with the upper left of the learning image G1 as a reference. The size of the partial feature is expressed by the horizontal and vertical lengths (W1, H1) of the region. Further, the shape is expressed by a shape pattern indicated by the filters F1 to F4, for example. As described above, the partial feature is expressed by the position, size, and shape, and is evaluated by the evaluation value.

  As illustrated in FIG. 2C, the partial feature learning unit 30 extracts a plurality of partial features included in the learning image G1 using a plurality of filters. Note that FIG. 2C is an example of a learning image when the identification target object is a person, and shows that partial features A to H surrounded by dotted lines are detected. Then, the partial feature learning unit 30 prepares a weak classifier for each extracted partial feature. This weak classifier is a classifier used in the partial feature evaluation value calculation unit 32 described later. For example, the above-described evaluation value is calculated in the determination image region corresponding to the position, size, and shape of the partial feature. Thus, the determination image has a function of determining whether or not a feature corresponding to the partial feature of the learning image exists.

  The partial feature learning unit 30 has a function of inputting a large number of learning images and learning a weak classifier that is effective in identifying a classification target object. Specifically, the partial feature learning unit 30 selects a plurality of weak classifiers that are optimal for identifying the identification target object by learning partial features common to the identification target object. At this time, a predetermined value necessary for identification is set for each selected weak classifier, and the weight of the weak classifier is calculated based on the error value, the weight of the image, and the like, and is associated with each weak classifier. This weight indicates the reliability of identification of the weak classifier and also indicates the reliability of the partial feature corresponding to the weak classifier. That is, it can be said that this weight is associated with the partial feature. Then, the partial feature learning unit 30 stores the partial feature used for identification, the associated weight and the weak classifier in the partial feature model storage unit 21 as a partial feature model.

  The combination pattern learning unit 31 has a function of learning a combination of a plurality of partial features of the identification target object. The combination pattern learning unit 31 refers to the learning image stored in the learning image storage unit 20 and the partial feature model stored in the partial feature model storage unit 21, and learns a combination of a plurality of partial features of the identification target object. . For example, the combination pattern learning unit 31 acquires the partial feature used for identification learned by the partial feature learning unit 30 from the partial feature model storage unit 21 and applies it to the learning image including the identification target object. Then, a combination of a plurality of partial features determined to have a feature corresponding to the partial feature in the learning image is output as a partial feature combination pattern. Whether or not a feature corresponding to the partial feature exists is determined by calculating the evaluation value of the feature amount described above in the region of the learned image corresponding to the position, size, and shape of the learned partial feature. It is determined using the evaluation value. By executing this process on a plurality of learning images, a plurality of partial feature combination patterns are output. The combination pattern learning unit 31 stores the learned partial feature combination pattern in the combination pattern storage unit 22.

  The partial feature evaluation value calculation unit 32 has a function of determining whether or not an identification target object exists in the determination image using a weak classifier. The partial feature evaluation value calculation unit 32 applies all weak classifiers stored in the partial feature model storage unit 21 to the determination image input by the determination image input unit 4. Specifically, the partial feature evaluation value calculation unit 32 calculates the evaluation value of the feature amount described above in the determination image region corresponding to the learned position, size, and shape of the partial feature. Then, it is compared with a predetermined value set for each weak classifier by the partial feature learning unit 30 to determine whether or not a feature corresponding to the partial feature of the learning image exists in the determination image. The weight associated with the weak classifier used for the determination is acquired from the partial feature model storage unit 21. The weak classifier is 1 when it is determined that the feature corresponding to the partial feature of the learning image exists in the determination image, and when it is determined that the feature corresponding to the partial feature of the learning image does not exist in the determination image. It has a function of outputting 0. The partial feature evaluation value calculation unit 32 has a function of integrating the output value of the weak classifier and the weight associated with the weak classifier and outputting the integrated value.

  The partial feature weight sum calculation unit 33 has a function of adding the integrated values calculated by the partial feature evaluation value calculation unit 32 and outputting the sum as a weight sum.

  The combination pattern coincidence degree calculation unit 34 has a function of calculating the degree of coincidence between the partial feature pattern extracted using the determination image and the learned partial feature pattern of the identification target object. The combination pattern coincidence calculation unit 34 is configured so that the partial feature combination pattern stored in the combination pattern storage unit 22 and the partial feature determined by the partial feature evaluation value calculation unit 32 to be present in the determination image, that is, the evaluation value are predetermined values. Using the partial features determined as above, the matching degree of the combination pattern is calculated. For example, the combination pattern matching degree calculation unit 34 selects one of the learned partial feature combination patterns, and is extracted using all of the feature amounts of the partial features included in the selected partial feature combination pattern and the determination image. All the feature values of the partial features are compared to determine whether or not they match. Thereby, in the partial feature combination pattern, the ratio (matching ratio) occupied by the partial features extracted using the determination image is calculated. This matching ratio is the combination pattern matching degree. Then, the same processing is repeated for all of the learned partial feature combination patterns, and the maximum matching ratio among the learned partial feature combination patterns is output.

  The identification result determination unit 35 has a function of determining whether or not the identification target object can be identified from the determination image. The identification result determination unit 35 performs determination using the weight sum that is the identification result of the partial feature weight sum calculation unit 33 and the matching ratio that is the identification result of the combination pattern matching degree calculation unit 34. For example, when the weight sum is smaller than a predetermined value or when the matching ratio is smaller than the predetermined value, the identification result determination unit 35 determines that the identification target object is not identified. On the other hand, for example, when the sum of weights is equal to or greater than a predetermined value and the matching ratio is equal to or greater than a predetermined value, the identification result determination unit 35 determines that the identification target object has been identified. Then, the identification result determination unit 35 outputs the determination result to the determination result display unit 5. The determination result display unit 5 displays the determination result according to the output.

  Next, the operation of the object identification device 1 according to this embodiment will be described. FIG. 3 is a flowchart showing an operation of calculating the weight sum of the object identification device 1 according to the present embodiment. The control process shown in FIG. 3 is started at the timing when the determination image is input, for example.

  As shown in FIG. 3, first, a learned partial feature model is read (S10). In the process of S10, the partial feature evaluation value calculation unit 32 inputs a learned partial feature model. The partial feature evaluation value calculation unit 32 refers to the partial feature model storage unit 21 and inputs a learned weak classifier and a weight associated with the weak classifier. For example, weak classifiers and weights corresponding to the partial features A to H shown in FIG. When the process of S10 ends, the process proceeds to a selection process (S12).

  In the process of S12, the partial feature evaluation value calculation unit 32 selects a partial feature to be compared with the determination image. The partial feature evaluation value calculation unit 32 arbitrarily selects one weak classifier to be evaluated from the weak classifiers input in the process of S10. When the process of S12 ends, the process proceeds to an evaluation value calculation process (S14).

  In the process of S14, the partial feature evaluation value calculation unit 32 calculates the evaluation value of the determination image. The partial feature evaluation value calculation unit 32 applies the weak classifier selected in the process of S12 to the determination image, and in the determination image region corresponding to the position, size, and shape of the partial feature, the partial feature evaluation value calculation unit 32 A value obtained by subtracting the sum of the pixel values in the black region of −1 from the sum of the pixel values is calculated as the evaluation value of the feature amount. For example, it is assumed that the determination image 1 shown in FIG. The partial feature evaluation value calculation unit 32 applies the weak classifier selected in the process of S12 to the determination image 1, and calculates an evaluation value in an area corresponding to the partial feature related to the weak classifier. When the process of S14 ends, the process proceeds to a weight determination process (S16).

  In the process of S16, the weight to be added by the partial feature evaluation value calculation unit 32 is determined. The partial feature evaluation value calculation unit 32 compares the evaluation value calculated in the process of S14 with a predetermined value set in advance in the weak classifier used in the process of S14. When the evaluation value calculated in the process of S14 is larger than a predetermined value set in advance in the weak classifier used in the process of S14, 1 is output by the function of the weak classifier. On the other hand, if the evaluation value calculated in the process of S14 is less than or equal to a predetermined value set in advance in the weak classifier used in the process of S14, 0 is output by the function of the weak classifier. The partial feature evaluation value calculation unit 32 refers to the weight associated with the weak classifier input in the process of S10, integrates the output value of the weak classifier and the weight associated with the weak classifier, and performs integration. Output as a weight to add values. When the process of S16 ends, the process proceeds to a completion determination process (S18).

  In the processing of S18, it is determined whether or not the partial feature evaluation value calculation unit 32 has evaluated all the partial features. The partial feature evaluation value calculation unit 32 determines whether or not the processing of S12 to S16 has been executed for all the partial features A to H input in the processing of S10. If it is determined in the process of S18 that all partial features have not been evaluated, the process proceeds to the selection process again (S12). In this way, the processes of S12 to S18 are repeatedly executed until all partial features are evaluated. On the other hand, if it is determined in the process of S18 that the evaluation has been completed for all the partial features, the process proceeds to an addition process (S20).

  In the process of S20, the partial feature weight sum calculation unit 33 calculates the weight sum. The partial feature weight sum calculation unit 33 adds all the weights calculated in the process of S16. When the process of S20 ends, the control process shown in FIG. 3 ends. Thus, the control process shown in FIG. 3 is finished. By executing the control process shown in FIG. 3, a weight sum obtained by evaluating the determination image using all the partial features learned using the learning image is calculated.

  Next, the learning operation of the partial feature combination pattern of the object identification device 1 according to the present embodiment will be described. FIG. 5 is a flowchart showing the learning operation of the partial feature combination pattern of the object identification device 1 according to this embodiment. The control process shown in FIG. 5 is started after the end of the control process shown in FIG. 3, for example.

  As shown in FIG. 5, first, a learning image reading process is executed (S30). In the process of S30, the combination pattern learning unit 31 inputs a learning image. The combination pattern learning unit 31 inputs a plurality of learning images with reference to the learning image storage unit 20. As the learning image, an image obtained by capturing an object to be identified is used. For example, the learning image Gn (n: natural number) shown in FIG. 6 is input. When the processing of S30 ends, the process proceeds to partial feature reading processing (S32).

  In the process of S32, the combination pattern learning unit 31 inputs a learned partial feature model. The combination pattern learning unit 31 refers to the partial feature model storage unit 21 and inputs a learned weak classifier. When the process of S32 is completed, the process proceeds to a selection process (S34).

  In the process of S34, the combination pattern learning unit 31 selects a learning image. The combination pattern learning unit 31 arbitrarily selects one learning image to be learned from the learning images Gn input in the process of S30. For example, the learning image G1 is selected. When the process of S34 ends, the process proceeds to an evaluation value calculation process (S36).

  In the process of S36, the combination pattern learning unit 31 calculates the evaluation value of the learning image. The combination pattern learning unit 31 applies all weak classifiers selected in the process of S32 to the learning image G1 selected in the process of S34. Then, in the region of the learning image G1 corresponding to the position, size, and shape of the partial feature, a value obtained by subtracting the sum of the pixel values in the -1 black region from the sum of the pixel values in the +1 white region is a feature amount. Calculated as the evaluation value. When the processing of S36 ends, the process proceeds to partial feature selection processing (S38).

  In the process of S38, the combination pattern learning unit 31 selects a partial feature included in the combination pattern. The combination pattern learning unit 31 compares the evaluation value calculated in the process of S36 with a predetermined value set in advance in the weak classifier. When the evaluation value calculated in the process of S36 is larger than a predetermined value set in advance in the weak classifier, 1 is output by the function of the weak classifier. On the other hand, when the evaluation value calculated in the process of S36 is less than or equal to a predetermined value set in advance in the weak classifier, 0 is output by the function of the weak classifier. The combination pattern learning unit 31 selects all the partial features whose evaluation value is greater than a predetermined value based on the output value of the weak classifier. For example, assume that the weak classifier corresponding to partial features A, B, C, and D in the learning image G1 outputs 1 and the weak classifiers corresponding to other partial features output 0. In this case, the combination pattern learning unit 31 selects partial features A, B, C, and D. When the process of S38 is completed, the process proceeds to a storage process (S40).

  In the process of S40, the combination pattern learning unit 31 stores the partial features included in the combination pattern. The combination pattern learning unit 31 stores the partial features A, B, C, and D selected in the process of S38 as a partial feature combination pattern P1 (ABCDD). When the process of S40 ends, the process proceeds to a completion determination process (S42).

  In the process of S42, it is determined whether the combination pattern learning unit 31 has learned all the learning images. The combination pattern learning unit 31 determines whether or not the processes of S34 to S40 have been executed for all the learning images Gn input in the process of S30. If it is determined in step S42 that learning has not been performed for all learning images, the process proceeds to selection processing again (S34). As described above, the processes of S34 to S42 are repeatedly executed until learning is performed for all the learning images Gn. On the other hand, if it is determined in step S42 that learning has been completed for all learning images Gn, the control processing shown in FIG. 5 is terminated. This is the end of the control process shown in FIG. By executing the control process shown in FIG. 5, a combination of a plurality of partial features related to the identification target object is learned, and a plurality of partial feature combination patterns Pn (n: natural number) are acquired.

  Next, a comparison operation of partial feature combination patterns of the object identification device 1 according to the present embodiment will be described. FIG. 7 is a flowchart showing the comparison operation of the partial feature combination patterns of the object identification device 1 according to this embodiment. The control process shown in FIG. 7 is started after the control process shown in FIGS. In consideration of the ease of understanding the explanation, in FIG. 7, the process is executed using the determination image in FIG. 3 and the learning image in FIG.

  As shown in FIG. 7, first, a learned partial feature combination pattern reading process is executed (S50). In the process of S50, the combination pattern matching degree calculation unit 34 inputs a learned partial feature combination pattern. The combination pattern matching degree calculation unit 34 refers to the combination pattern storage unit 22 and inputs a learned partial feature combination pattern. For example, the partial feature combination pattern Pn shown in FIG. 6 is input. When the processing of S50 ends, the process proceeds to selection processing (S52).

  In the process of S52, the partial feature combination pattern Pn to be compared is selected by the combination pattern matching degree calculation unit 34. The combination pattern matching degree calculation unit 34 arbitrarily selects one partial feature combination pattern Pn input in the process of S50. When the process of S52 ends, the process proceeds to an evaluation value calculation process (S54).

例えば、Ｓ５２の処理で図６に示す部分特徴組合せパターンＰ１（Ａ−Ｂ−Ｃ−Ｄ）を入力したとし、判定画像で所定値よりも大きいと判定された部分特徴が部分特徴Ａ，Ｃ，Ｄ，Ｇであるとする。この場合、式１を用いて、一致割合Ｋ_Ａは０．７５とされる。Ｓ５４の処理が終了すると、完了判定処理へ移行する（Ｓ５６）。 In the process of S54, the combination pattern matching degree calculation unit 34 calculates the matching ratio (combination matching degree) of the partial feature combination patterns. The combination pattern coincidence degree calculation unit 34 inputs and receives the partial features determined to be larger than the predetermined value in the processes of S14 and S16 in FIG. 3, that is, the partial features determined to be included in the determination image. It is determined whether the feature amount of the partial feature matches the feature amount of the partial feature included in the partial feature combination pattern Pn selected in the process of S52, and the number of partial features determined to match is counted. For example, it is determined whether the shape, position, and size are the same. When the number of matching partial features is N _A and the number of partial features included in the partial feature combination pattern is N, the matching ratio K _A is calculated using the following Equation 1.

For example, if the partial feature combination pattern P1 (ABCD) shown in FIG. 6 is input in the process of S52, the partial features determined to be larger than a predetermined value in the determination image are the partial features A, C, Assume that D and G. In this case, using equation 1, match percentage _{K A} is set to 0.75. When the processing of S54 ends, the process proceeds to completion determination processing (S56).

  In the process of S56, it is determined whether the combination pattern matching degree calculation unit 34 has evaluated all the partial feature combination patterns. The combination pattern matching degree calculation unit 34 determines whether or not the processes of S52 and S54 have been executed for all the partial feature combination patterns Pn input in the process of S50. In the process of S56, when it is determined that all the partial feature combination patterns have not been evaluated, the process proceeds to the selection process again (S52). In this way, the processes of S52 to S56 are repeatedly executed until all partial feature combination patterns are evaluated. On the other hand, if it is determined in the process of S56 that the evaluation has been completed for all the partial features, the process proceeds to a matching ratio determination process (S58).

In the process of S58, the combination pattern matching degree calculation unit 34 selects the largest matching ratio among all the partial feature combination patterns. When the process of S58 ends, the control process shown in FIG. 7 ends. The control process shown in FIG. By executing the control process shown in FIG. 7, how many partial features determined to be similar to the characteristic portion of the identification target object in the determination image are included in the combination of partial features of the identification target object in the learning image _A matching ratio KA indicating whether or not Thereby, it is possible to evaluate to what extent the combination of partial features included in the determination image matches the learned combination of identification target objects.

  Next, the identification determination operation of the object identification device 1 according to the present embodiment will be described. FIG. 8 is a flowchart showing the identification determination operation of the object identification device 1 according to the present embodiment. The control process shown in FIG. 8 is started after the control process shown in FIGS. In consideration of the ease of understanding the explanation, in FIG. 8, it is assumed that the processing is executed using the processing result in FIG. 3 and the processing result in FIG.

  As shown in FIG. 8, first, identification determination processing using a weight sum is executed (S60). In the process of S60, the identification result determination unit 35 inputs the weight sum calculated by the control process shown in FIG. And the identification result determination part 35 determines whether a weight sum is more than predetermined value. If it is determined that the sum of weights is equal to or greater than the predetermined value, it is determined that there is a possibility that the object is an identification target object, and the process proceeds to a match ratio determination process (S62).

In the processing of S62, the identification result determining section 35 inputs the coincidence ratio K _A calculated by the control process shown in FIG. The identification result determining section 35 matches the ratio K _A is equal to or greater than a predetermined value. If a match percentage K _A is equal to or more than the predetermined value, determines that the identification object (S64). When the process of S64 ends, the control process shown in FIG. 8 ends.

On the other hand, in the process of S60, when it is determined that the weight sum is smaller than the predetermined value, it is determined that the object is not the identification target object because there is no possibility of being the identification target object (S66). Further, it is determined in the process of S62, if the matching ratio K _A is determined to be smaller than the predetermined value, as there is no possibility that the identification object, not the identification target object (S66). When the process of S66 ends, the control process shown in FIG. 8 ends.

Thus, the control process shown in FIG. 8 ends. By executing the control process shown in FIG. 8, the possibility of the identification target object is determined using the weight sum, and further, the identification target object is determined using the matching ratio K _A with the partial feature included in the partial feature combination pattern. Is determined. Thus, identification accuracy can be improved by performing two determination processes.

  For example, a case where only the partial feature of the determination image matches the identification target object will be described. If the object to be identified is a person, it is necessary to determine that the utility pole shown in FIG. 4B and the image of the house in the back are not objects to be identified. However, depending on the arrangement of the utility pole and the house in the back, only the characteristic part may match. In this case, if it is determined only by the weight sum, there is a risk of erroneous determination that there is an identification target object in the determination image 2.

On the other hand, according to the object identification device 1 according to the present embodiment, the partial feature evaluation value calculation unit 32 evaluates the feature amount of the determination image based on the feature amount of the partial feature, and the evaluation value for each partial feature. Is calculated, the reliability weight associated with the partial feature is acquired based on the evaluation value, and the weighted partial feature weight sum calculation unit 33 calculates the weight sum that is the sum of the weights. Further, the combination pattern coincidence degree calculation unit 34 is a coincidence ratio that is the degree of coincidence between the combination of partial features determined to have an evaluation value equal to or greater than a predetermined value and the combination of partial features related to the identification target object learned in advance. K _A is calculated. Then, the identification result determining section 35, the weighting sum is greater than a predetermined value, and if matching percentage K _A is a predetermined value or more is determined to have identified the identification object from the determination image. Thus, in addition to the identification judgment by the sum of the weights of the partial feature can be implemented identification determination by matching the ratio K _A combination of the partial features. For this reason, for example, even when the weight sum becomes a predetermined value or more because only the local partial feature of an object different from the identification target object is similar to the identification target object, it is possible to avoid be identified as being the identification target object to another object using the matching ratio K _a. In other words, leave learning a combination of partial features advance the learning time, when the learning-time and likely erroneous determination in different combinations, to change the determination result according to the matching ratio K _A of the combination learned beforehand it can. Therefore, identification accuracy can be improved.

Further, according to the object identification device 1 according to the present embodiment, it is possible to calculate the matching ratio K _A, in the combination of the partial features of the pre-learned identified target object, the evaluation value is equal to or higher than a predetermined value it is possible to calculate the matching rate K _a that indicates the determined partial feature identical parts characterized as being is included extent. Accordingly, it is possible to evaluate the degree of coincidence between the combination of the partial features of the determination image evaluated to match the learned partial features and the combination of the partial features of the learned identification target object.

(Second Embodiment)
The object identification device according to the second embodiment is configured in substantially the same manner as the object identification device 1 according to the first embodiment, and only the operation of the combination pattern matching degree calculation unit 34 is different. Therefore, the second embodiment will be described with a focus on differences from the first embodiment, and a duplicate description will be omitted.

  The configuration of the object identification device according to the second embodiment is the same as that of the object identification device 1 according to the first embodiment. The operation of the object identification device according to the second embodiment is the same except for the partial feature combination pattern comparison operation shown in FIG.

  A comparison operation of partial feature combination patterns of the object identification device according to the second embodiment will be described. FIG. 9 is a flowchart showing a comparison operation of partial feature combination patterns of the object identification device according to the present embodiment. The control process shown in FIG. 9 is started after the control process shown in FIGS. In consideration of the ease of understanding the description, in FIG. 9, the process is executed using the determination image in FIG. 3 and the learning image in FIG.

  As shown in FIG. 9, first, a learned partial feature combination pattern reading process is executed (S70). This process is the same as the process of S50 of FIG. When the processing of S70 ends, the process proceeds to selection processing (S72).

  In the process of S72, the combination pattern matching degree calculation unit 34 selects a partial feature combination pattern Pn to be compared. This process is the same as the process of S52 of FIG. When the processing of S72 ends, the process proceeds to evaluation value calculation processing (S74).

  In the process of S74, the combination pattern matching degree calculation unit 34 calculates the matching degree of the feature amounts of the partial feature combination patterns. The combination pattern coincidence degree calculation unit 34 inputs the partial features determined to be larger than the predetermined value in the processes of S14 and S16 in FIG. 3, that is, the partial features determined to be included in the determination image. Then, in order to determine whether the feature quantity of the input partial feature and the feature quantity of the partial feature included in the partial feature combination pattern Pn selected in the process of S52 are the same or similar, Evaluate shape, position and size.

First, the evaluation of the shape will be described. For example, as shown in FIG. 2A, partial features of four types of shapes will be described. As shown in FIG. 2A, there are four types of shapes depending on the filters F1 to F4. For this reason, the partial feature (hereinafter referred to as partial feature U1) set to 1 by the weak classifier in the process of FIG. 3 and the partial feature included in the partial feature combination pattern (hereinafter referred to as partial feature U2). When the four shapes corresponding to the filters F1 to F4 match, the matching degree K _{shape is set} to 1, and when they do not match, the matching degree K _{shape is set} to zero. Note that the four types of shapes are completely different, so 0 or 1 was selected. However, when the filters have similar shapes, for example, when the filters are just rotated, the degree of similarity is similar. The matching degree K _shape may be set to a value in the range of 0 to 1. In this way, the _shape is evaluated using the degree of coincidence K _shape .

上記式２、３を用いて算出された割合Ｒ（Ｕ１），Ｒ（Ｕ２）に基づいて、一致度Ｋ_{ｐｏｓ＆ｓｉｚｅ}を以下の式４を用いて算出する。

このように、一致度Ｋ_{ｐｏｓ＆ｓｉｚｅ}を用いて位置及び大きさについての評価を行う。以上でＳ７４の処理を終了する。Ｓ７４の処理が終了すると、一致割合算出処理へ移行する（Ｓ７６）。 Next, evaluation on the position and size will be described. For example, the area where the rectangles of the two partial features U1 and U2 overlap is calculated, and the ratios R (U1) and R (U2) of the overlapping regions in the partial features are calculated. For example, assuming that the area of the overlapping region is S _W , the area of the partial feature U 1 is S _{U 1} , and the area of the partial feature U 2 is S _{U 2} , the ratios R (U 1) and R (U 2) are expressed by the following equations 2 and 3. Can do.

Based on the ratios R (U1) and R (U2) calculated using the above formulas 2 and 3, the degree of coincidence K _{pos & size} is calculated using the following formula 4.

In this way, the position and size are evaluated using the degree of coincidence K _{pos & size} . Thus, the process of S74 is completed. When the process of S74 ends, the process proceeds to a coincidence ratio calculation process (S76).

In the process of S76, the combination pattern matching degree calculation unit 34 calculates the matching ratio of the partial feature combination patterns. The combination pattern coincidence degree calculation unit 34 inputs and receives the partial features determined to be larger than the predetermined value in the processes of S14 and S16 in FIG. 3, that is, the partial features determined to be included in the determination image. It is determined whether the feature amount of the partial feature matches the feature amount of the partial feature included in the partial feature combination pattern Pn selected in the process of S52, and the number of partial features determined to match is counted. Judgment regarding whether or not they match is performed using the matching degrees K _shape , K _{pos & size} calculated in the process of S74. For example, the integrated value of the coincidence degrees K _shape , K _{pos & size} is set as the total coincidence degree K _all, and when the total coincidence degree K _all is equal to or greater than a predetermined value, it is determined that the coincidence is achieved. In this way, it is determined that partial features that are not completely identical but have similar feature quantities match. Thereafter, the combination pattern match degree calculating section 34, similarly to the processing in S54 in FIG. 7, to calculate a coincidence rate K _A using Equation 1. When the process of S76 ends, the process proceeds to a completion determination process (S78).

  In the process of S78, it is determined whether the combination pattern matching degree calculation unit 34 has evaluated all the partial feature combination patterns. This process is the same as the process of S56 of FIG. That is, if it is determined that all partial feature combination patterns have not been evaluated, the processes of S72 to S78 are repeatedly executed until all partial feature combination patterns are evaluated. On the other hand, if it is determined in the process of S78 that the evaluation has been completed for all the partial features, the process proceeds to a match ratio determination process (S80).

  In the process of S80, the combination pattern matching degree calculation unit 34 selects the largest matching ratio among all the partial feature combination patterns. This process is the same as the process of S58 of FIG. When the process of S80 ends, the control process shown in FIG. 9 ends. Thus, the control process shown in FIG. 9 ends. By executing the control process shown in FIG. 9, how many partial features determined to be similar to the characteristic part of the identification target object in the determination image are included in the combination of the partial characteristics of the identification target object in the learning image The matching ratio indicating whether or not The matching ratio is calculated not as a completely matched partial feature but as a similar partial feature, unlike the first embodiment.

As described above, according to the object identification device according to the second embodiment, while having the same effect as the object identification device 1 according to the first embodiment, among the combinations of partial features related to the identification target object learned in advance, evaluation value can be calculated coincidence rate K _a indicating that it contains what extent the determined portion, wherein the same or similar parts, wherein a predetermined value or more. As a result, when the partial feature determined to have an evaluation value equal to or greater than the predetermined value is not the same as the partial feature included in the combination but is similar, the partial feature related to the identification target object learned in advance Can be included in the combination. Further, by providing a range for the determination of whether or not they match, the combination of the partial features of the determination image evaluated to match the learned partial features and the combination of the partial features of the learned identification target object The degree of coincidence can be calculated including an error. Therefore, it is possible to avoid making the identification target object unidentified by strict determination.

(Third embodiment)
The object identification device according to the third embodiment is configured in substantially the same manner as the object identification device 1 according to the first embodiment, and only the operation of the identification result determination unit (weight sum correction means) 35 is different. Therefore, the third embodiment will be described with a focus on differences from the first embodiment, and a duplicate description will be omitted.

  The configuration of the object identification device according to the third embodiment is the same as that of the object identification device 1 according to the first embodiment. The operation of the object identification device according to the third embodiment is the same except for the identification determination operation shown in FIG.

  An identification determination operation of the object identification device according to the third embodiment will be described. FIG. 10 is a flowchart showing an identification determination operation of the object identification device according to the present embodiment. The control process shown in FIG. 10 is started after the control process shown in FIGS. In consideration of the ease of understanding the explanation, in FIG. 10, it is assumed that the processing is executed using the processing result in FIG. 3 and the processing result in FIG.

Ｓ９０の処理が終了すると、修正重み和判定処理へ移行する（Ｓ９２）。 As shown in FIG. 10, first, a process of correcting the weight sum is executed using the matching ratio (S90). In the processing of S90, for example, the identification result determining section 35 inputs the weight sum calculated by the control processing shown in FIG. 3, and the match rate K _A calculated by the control process shown in FIG. 7 or FIG. Then, to correct the weight sum based on matching ratio K _A. The corrected weight sum Sr _SUM can be expressed by using the following Equation 5, where S _SUM is the weight sum calculated by the control process shown in FIG.

When the process of S90 ends, the process proceeds to a correction weight sum determination process (S92).

In the process of S92, the identification result determination unit 35 determines whether or not the correction weight sum Sr _SUM is greater than or equal to a predetermined value. When it is determined that the correction weight sum Sr _SUM is greater than or equal to a predetermined value, it is determined that the object is an identification target object (S94). When the process of S94 ends, the control process shown in FIG. 10 ends. On the other hand, in the process of S92, when it is determined that the correction weight sum _SrSUM is smaller than the predetermined value, it is determined that the object is not an identification target object (S96). When the process of S96 ends, the control process shown in FIG. 10 ends.

This is the end of the control process shown in FIG. By executing the control processing shown in FIG. 10, matching ratio K is calculated corrected weighting sum Sr _SUM with _A, whether the identification target object using the corrected weights sum Sr _SUM is determined. When a partial feature evaluated as “not an object to be identified” is learned by the Adaboost algorithm, a negative weight is given to the weak classifier. For this reason, when such a weak discriminator is used, the weight sum _SSUM is reduced, and it may be assumed that even if it is a discrimination target object, it cannot be discriminated.

However, according to the object identification device according to the third embodiment, the identification result by the determination unit 35, it is possible to weight the sum S _SUM is acceleration corrected based on matching ratio K _A combination, modifications weights sum Fixed Identification determination can be performed using Sr _SUM . Thus, for example, for example, albeit at identifying the object, if the sum S _SUM of weights is not slight enough to the threshold, than the weight sum S _SUM corrective weighting sum Sr _SUM by integrating the high concordance rate K _A Therefore, it is possible to avoid unidentification. Further, when the weight sum S _SUM is increased in unexpected combinations, since the correction weights sum Sr _SUM by integrating a low matching ratio K _A can be smaller than the weight sum S _SUM, misjudgment It can be avoided.

  Each of the above-described embodiments shows an example of the object identification device according to the present invention. The object identification device according to the present invention is not limited to the object identification device according to each embodiment, and the object identification device according to each embodiment is modified or changed without changing the gist described in each claim, or It may be applied to other things.

  For example, in the above-described embodiment, it has been described that the partial feature is expressed by the position, size, and shape. However, the partial feature may be expressed by at least one of the position, size, and shape. That is, in the second embodiment described above, it has been described that the shape, position, and size of the partial feature are evaluated in order to determine whether or not the feature amount of the partial feature is the same or similar. You may evaluate by one.

Further, in the second embodiment described above, the example in which the position and the size are used in Expression 2 to Expression 4 when determining whether or not the feature amounts of the partial features are the same or similar has been described. There is no limitation, and an appropriate evaluation method may be used in which the degree of coincidence increases as the two partial features U1 and U2 become similar in size at similar positions. For example, the sum of the difference between the center positions of the two partial features U1 and U2 and the difference in size may be used as the degree of matching K _{pos & size} .

  In each of the above-described embodiments, the recording device 2, the partial feature learning unit 30, and the combination pattern learning unit 31 may be provided according to required performance. For example, when the object identification device 1 includes an interface capable of acquiring information stored in the recording device 2, information output from the partial feature learning unit 30 and the combination pattern learning unit 31, the recording device 2, partial feature learning It is not necessary to include the unit 30 and the combination pattern learning unit 31.

  DESCRIPTION OF SYMBOLS 1 ... Object identification apparatus, 30 ... Partial feature learning part, 31 ... Combination pattern learning part, 32 ... Partial feature evaluation value calculation part (evaluation value calculation means, acquisition means), 33 ... Partial feature weight sum calculation part (weight sum calculation) Means), 34... Combination pattern coincidence calculating unit (matching degree calculating unit), 35... Identification result determining unit (weight sum correcting unit, identification result determining unit).

Claims (5)

An object identification device that identifies an identification target object from a determination image based on feature amounts of a plurality of partial features learned in advance using a learning image and reliability weights associated with each of the partial features,
Evaluation value calculating means for evaluating the feature amount of the determination image based on the feature amount of the partial feature and calculating an evaluation value for each partial feature;
Obtaining means for obtaining a weight of reliability associated with the partial feature based on the evaluation value;
A weight sum calculating means for calculating a weight sum that is a sum of the weights of the reliability obtained by the obtaining means;
A degree of coincidence calculation that calculates a degree of coincidence of matching that is a degree of coincidence between the combination of the partial features determined that the evaluation value is equal to or greater than a predetermined value and the combination of the partial features related to the identification target object learned in advance. Means,
An identification result determination unit that determines that the identification target object has been identified from the determination image when the weight sum is a predetermined value or more and the combination matching degree is a predetermined value or more;
An object identification device comprising:   The degree-of-match calculation means includes the partial feature and the feature amount included in the partial feature combination of the identification target object learned in advance among the partial features determined that the evaluation value is equal to or greater than a predetermined value. The number of identical partial features is calculated as the number of matches, the ratio of the number of matches to the number of partial features included in the combination is calculated, and the combination matching degree is calculated based on the calculated ratio The object identification device according to claim 1.   The degree-of-match calculation means includes the part included in the combination of the partial features of the identification target object learned in advance among the partial features of the identification target object determined that the evaluation value is equal to or greater than a predetermined value. Calculate the number of partial features that have the same or similar feature and feature quantity as the number of matches, calculate the ratio of the number of matches to the number of partial features included in the combination, and based on the calculated ratio, The object identification device according to claim 1, wherein the combination matching degree is calculated. A weight sum correction means for adjusting the weight sum based on the combination matching degree;
When the weight sum is corrected by the weight sum correction unit, the identification result determination unit is configured such that the weight sum corrected by the weight sum correction unit is equal to or greater than a predetermined value, and the combination matching degree is equal to or greater than a predetermined value. In this case, the object identification device according to claim 1, wherein it is determined that the identification target object has been identified from the determination image.   The object identification device according to any one of claims 1 to 4, wherein the feature amount of the partial feature includes at least one of a shape, a position, and a size.

Images