Method for detecting an object in an image and device for carrying out the method

The method addresses the challenge of recognizing complex objects in unknown poses by employing a hierarchical neural network to learn and assign reliability values based on component relationships, enhancing detection accuracy and robustness.

DE102014010982B4Active Publication Date: 2024-11-14SEW EURODRIVE GMBH & CO KG
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
DE102014010982
Authority / Receiving Office
DE · DE
Patent Type
Patents
Current Assignee / Owner
Filing Date
2014-07-29
Publication Date
2024-11-14
Estimated Expiration
2034-07-29

AI Technical Summary

Technical Problem

Existing object recognition methods struggle with recognizing complex structured objects in various poses, especially when they are unknown during the learning phase, due to lack of rotation invariance and computationally intensive scaling space representations, and fail to model global relationships effectively.

Method used

A method using a hierarchical, unconnected tree of semantic components with a neural network that learns pairwise relationships between components, allowing for reliable detection of objects in different poses by assigning reliability values based on learned weight factors and considering permissible distances and orientations.

Benefits of technology

Enables robust object recognition, including complex structured objects in unknown poses, by leveraging hierarchical modeling and neural networks to improve detection reliability through learned weight factors and pairwise component relationships.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

Method for detecting an object in an image captured by a camera, the procedure is carried out using a computer, where components are modeled in the computer C kl are stored in a hierarchical order, where I is the hierarchy level indexed from a lowest level I = 0 to a highest level 1 = L and is a natural number, and k numbers the components of a respective hierarchy level and is a natural number. where L is a natural number that denotes the highest hierarchy level, where each component C kl at least one component C of a hierarchy level 1 that is lower than the highest hierarchy level L m l+1 is assigned to the next higher hierarchy level I + 1, where m is a natural number representing the assigned components C m l+1 characterizes, where C1 0 the object is modeled, so for the value 1 = 0 the modeled component corresponds to the object, where the image is searched pairwise for components of the same hierarchy level, and from that, a reliability value for each component score C il , that is, the reliability of the detection of the respective component C il descriptive value is assigned, where i is a natural number where a weighting factor w is used in determining the reliability value l The following is taken into account, which was previously determined by a neural network during a learning phase. where a weighting factor is taken into account to determine the reliability value, which is derived from a sum of the reliability values score CI l + 1 the respective component C il assigned components of the next higher level are determined, where the reliability value of the component C il is determined according to scorecil = { wl ∑ j R s ( C il , C jl ) ∑ kscore C kl + 1 , if K ≠ ϕ ∑ j R s ( C il , C jl ) , otherwise where K is the set of subcomponents of C il is and whereby C jl ∈ J , where J is the set of coupled components of C il is and where R s a digitized measure which depends on the specific image positions of the respective component pair (C jl , C jl ) , in particular where this measure is not zero if the two components have a predetermined allowed distance to each other and a predetermined allowed relative orientation to each other.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The invention relates to a method for detecting an object in an image and a device for carrying out the method.

[0002] From US 7 239 929 B2 a method for object recognition is known in which the components of a physical object are hierarchically ordered and used as a basis.

[0003] From the publication “Articulated Human Detections with flexible mixture-of-Parts” by the authors Yi Yang and Deva Ramanan, IEEE Transactions, vol.35, no. 12, pp. 2878-2890, 2013, it is known to search for components of an object in an image and to additionally take into account whether each component is at a permissible distance from a component neighboring the object.

[0004] The method according to [3] detects components of human bodies. However, the relationships are only approximated as Gaussian distributions in the scaling space G(x1, y1, x2, y2, σ1, σ2). Therefore, the method is not applicable for extreme poses. Furthermore, the components are considered dependent on the rotation, thus, they are not considered rotationally invariant.

[0005] The method according to [4] does not represent any relationships between the parts.

[0006] Although the methods according to [5] and [6] use hierarchically ordered components, the relationships are modeled by a deformable star model. Therefore, poor results are achieved for extreme poses.

[0007] The method according to [2] uses a relational, tree-based representation with very small local intrinsic components. Relationships among the local components are learned. Additionally, a type of rectangular feature is used, as known from [7]. No rotation invariance is provided.

[0008] A representation with fully connected graphs is known from US 7 239 929 B2 [8]. The relationships between the components are taken into account using a specific similarity measure. The component representations are described using image corners or using the generalized Hough transform. During detection, one or more core components (root parts) are searched for in the entire image. The connected components are then searched for in a subset of the image space, in particular using an optimized search strategy. The fully connected graph is not hierarchical. Thus, an upper hierarchy level does not contain any components of a lower level. Distance effects can therefore only be considered using a computationally intensive scale space representation. The representation of the components is not rotation invariant.

[0009] A hierarchical, unconnected tree of semantic components is known from US Pat. Nos. 8,260,059 B2 [9] and 7,725,484 B2

[10] . Relationships among local components, i.e., components within the same level, are learned separately. In contrast to [5], [6], and [2], global relationships between components are not modeled. This is because objects, such as human bodies or the like, can perform poses that would be complex to represent and therefore lead to an extremely large number of configurations, making it impossible to learn all possible poses with a limited amount of data in the training phase. From the publication

[0010]

[11] DUAN, K. [et al.]: A Multi-layer Composite Model for Human Pose Estimation. Proceedings of the British Machine Vision Conference, Vol. 2, 2012, Paper 116, pages 1 to 11 A method for object recognition is known. From the publication

[0011]

[12] Felzenszwalb, PF [et al.]: Object detection with discriminatively trained part-based models. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 32, NO. 9, 2010, pages 1627 - 1645 A multi-layered model for estimating human poses is known. From the publication

[0012]

[13] Wang, Y. [et al.]: Learning Hierarchical Poselets for Human Parsing, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2011, pages 1705-1712 A learning method for recognizing human poses is known.

[0013] The invention is therefore based on the object of carrying out object recognition with a learning algorithm, whereby the object should be recognizable even if the pose of the object was still unknown in the learning phase.

[0014] According to the invention, the object is achieved in the method for detecting an object in an image according to the features specified in claim 1 and in the device according to the features specified in claim 2.

[0015] Important features of the invention in the method for detecting an object in an image, in particular in an image taken by a camera, are that the method is carried out with a computer, where computer-modeled components Ckl are stored in a hierarchical order, where I is the level, in particular hierarchy level, indexed from a lowest level I = 0 to a highest level I = L and is a natural number and k numbers the components of a respective hierarchy level and is a natural number, where L is a natural number indicating the highest level, in particular, each component Ckl a level 1 which is lower than the highest level L, at least one component C m l+1 is assigned to the next higher level I + 1, where m is a natural number which represents the assigned components C m l+1 indicates, where C1 0 the object is modelled, in particular for the value I = 0 the modelled component corresponds to the object, where the image is searched pairwise for components of the same level, in particular if more than one component is classified in the level, and from this each component a reliability value scoreCil in particular the reliability of the detection of the respective component Cil descriptive value, where i is a natural number.

[0016] Advantageously, the invention enables the recognition of complexly structured objects that can assume various poses. Such object recognition can be used in various applications. In particular, robots equipped with such object recognition are capable of recognizing objects, such as humans, even if the respective objects assume previously unknown poses. In an upstream initial method step, at least one object and its components are modeled, and in a further upstream step, a computer, in particular comprising a neural network, is trained using a learning algorithm.

[0017] The computer is trained in an offline training and then an online detection of the object is carried out, i.e. the recognition of an object is carried out.

[0018] In the aforementioned process steps, dense feature vectors are provided over the detection and / or training window. These feature vectors are intended for training the neural network and for detecting an object. Advantageously, this enables simple reconfiguration of the object's components, and object detection can be performed independently of the rotation of the respective component, distance effects, scaling, or affine transformations.

[0019] Advantageously, according to the invention, objects in different poses can be recognized, whereby the invention makes use of semantic components of the object.

[0020] According to the invention, human poses are characterized by locally paired component configurations learned from a limited number of training images. Components that are located at the same hierarchy level are connected by links.

[0021] In an advantageous embodiment, a weighting factor w is used to determine the reliability value. l The weight factor is taken into account, which was previously determined by a neural network in a learning phase. The advantage here is that in the training phase, a limited number of images are sufficient to determine the weight factor.

[0022] According to the invention, a weighting factor is taken into account to determine the reliability value, which is a sum of the reliability values scoreCkl+1 of the respective component Cil assigned components of the next higher level, where k numbers these assigned components. The advantage here is that the more and more reliably subcomponents of the component are recognized, the more reliably a component is recognized.

[0023] According to the invention, the reliability value of the component Ckl determined according to scoreCil={wl∑jRs(Cil,Cjl)∑kscoreCkl+1,if K≠ϕ∑jRs(Cil,Cjl),otherwise where K is the set of subcomponents of Cil; so C k l+1 ∈ K, where Cjl∈J, where J is the set of coupled components of Cil is.

[0024] The advantage of this is that when subcomponents of a component are detected, the reliability of their detection is improved. In this case, k assigned components of the next higher level are numbered, so that the sum is carried out over k. Since no weighting factors are taken into account in this sum, a simple and fast determination is possible. R s is a digitized measure which depends on the specific image positions of the respective component pair (Cil,Cjl) This dimension is zero if the two components are not within the permitted distance from each other and / or if the two components are not in the permitted relative orientation. The dimension is not zero if the two components are within the specified permitted distance from each other and within the specified permitted relative orientation. This is specified in the initial process step, where the respective relative distance and the respective relative orientation of each pair of components connected by connections are specified. The weighting factor w l is learned in the training phase by supplying a limited number of stored images.

[0025] Important features of the device for carrying out the aforementioned method are that a computer is connected to a camera in such a way that the images captured by the camera can be evaluated by the computer. This is advantageous because it enables rapid recognition.

[0026] Further advantages emerge from the dependent claims. The invention is not limited to the combination of features in the claims. Further possible combinations of claims and / or individual claim features and / or features of the description and / or the figures will become apparent to those skilled in the art, particularly from the problem and / or the problem posed by comparison with the prior art.

[0027] The invention will now be explained in more detail with the aid of illustrations: In the Fig. 1 shows a physical object with its semantic components. In the Fig. Figure 2 shows various scales SZ1 to SZ3, with the search windows 20 being correspondingly large. For each scale, the image is defined in different resolutions, so that each hierarchy level can be assigned a correspondingly different resolution image. In the Fig. 3 shows search windows 20 with components assigned to the levels as examples. In the Fig. Figure 4 shows an example of the training phase sequence, wherein high-resolution training images 3 are fed in, and scaled images 8 are considered during training 4 of the recognition device, which also takes into account a spatial definition 2 of the respective component. Training 5 is then performed for the relationship between the components, and offline model values ​​7, in particular weight factors, are determined. During training 5, definitions 1 of the components and their relationship to one another are also considered. When the decision 6 is made that the lowest scaling has been reached, the training phase is terminated 9. Otherwise, the images are scaled differently 8, and training 4 continues.

[0028] As in Fig. 1, different levels are shown, with the lowest level (Level 0) representing projections of the object as a whole in different spatial orientations, specifically poses, where the respective set of projections is referred to as an assembly. A model of a human body is chosen as an example of such an object. The object has components C i which are shown as examples in the next level, i.e. Level 1. The components C i in Fig. 1 a torso with arms, a head and legs is shown, whereby different spatial orientations of the respective component C i the vertical projections into the image plane are shown.

[0029] Likewise, in the next higher level, i.e. Level 2, a respective representation of the respective projections of subcomponents, such as the bare torso and arm, is shown.

[0030] In the highest level, i.e. Level 3, a representation of the respective projections of even more detailed subcomponents, such as bare fingers, hand, forearm, is shown.

[0031] Thus, in Fig. 1 shows a hierarchical representation in which detailed information increases from the lowest to the highest level.

[0032] In the inventive method for object detection in an image captured with a camera, two components are searched for in each image, and whether their distance and orientation are permissible is considered. For this purpose, a neural network, in particular a perceptron, is fed a number of images in which the components and their distance and orientation are recognizable. The neural network thus learns the permissible distances and orientations and, when subsequently used for object detection in a newly captured image, is able to perform an effective evaluation. The result is a reliability value that describes the reliability of the detected object.

[0033] The special feature of the process according to the invention lies in the procedure that components C i and C j, especially at the same hierarchical level, are searched for and evaluated in pairs within the image. This also makes it possible to detect novel poses of the object, i.e., poses that have never been shown to the neural network before and are therefore unknown, and to achieve a high level of reliability in this detection.

[0034] The process is therefore carried out using a camera that is connected to a computer.

[0035] During the learning phase, images of the object are fed to the computer, allowing the object to assume various poses. With human assistance, components of the object are labeled so that the neural network learns the object during the learning phase and recognizes it in the subsequent evaluation phase, even if it is arranged in a new pose.

[0036] To fine-tune the confidence value, or reliability value, the image is also searched for the respective subcomponents, and a weighting factor is calculated from the corresponding result, which influences the reliability value. Thus, if an arm and a hand are detected that are assigned to the same hierarchy level, and fingers are also detected on the hand, which are assigned as subcomponents of the hand at the next higher level, the recognition reliability is increased.

[0037] As in Fig. As shown in Figure 2, a low-resolution Level 0 image, a higher-resolution Level 1 image, and a very high-resolution Level 2 image are determined from the respective image for detecting the Level 0 components.

[0038] The search window for component detection is moved through the respective image at various sizes (SZ1 to SZ3), and a confidence value for component detection is determined. If a threshold is exceeded, the component is considered detected.

[0039] As in Fig. As shown in Figure 3, a search window containing a human body in very low resolution is used to search for a component at the highest level, i.e., Level 0. In the next level, i.e., Level 1, a search window is used to search the associated subcomponent. Likewise, in the next level, a search window is used to search for an even more detailed subcomponent. This is shown in Fig. 4 continued up to level 3.

[0040] The inventive method for object recognition comprises several method steps, which are now presented in more detail: In the first step, the object is modeled, taking into account various poses of the object.

[0041] To model the object, the object or its components are described geometrically, using the perpendicular projections of the respective model into the image plane. Furthermore, the pairs of components connected by links are described. These links are considered as a hierarchical arrangement of the components. Thus, at the lowest hierarchical level, the object is modeled in various poses, and at the higher hierarchical levels, the object's components are modeled in different poses.

[0042] In a process step carried out as a learning phase, i.e. an offline training phase, the neural network of the computer or the learning algorithm implemented in the computer is trained by feeding a number of exemplary images to the neural network or the learning algorithm and training the appearances of the components and the pairwise relationships of the components, in particular of the same level.

[0043] In a subsequent process step, online detection is performed. An image captured by a camera is fed to the computer. The image is first downscaled in various steps, so that the next lower downscale has a lower resolution than the previous one. Each of these generated images is assigned to a level of the hierarchy. The lowest resolution corresponds to level 0, and the highest resolution to level L. The levels are numbered with the variable I, which can assume integer values ​​from 0 to L, since there are L+1 levels.

[0044] In each image assigned to Level I, the components assigned to Level I are searched for pairwise, and the relationship between the two components is detected. The weighting factors determined in the offline training phase are used for weighting.

[0045] In this way, a reliability value can be determined that describes the reliability or confidence that can be assigned to the detected component.

[0046] The object p is composed of components C j , where j = 0, 1, 2...m, . and can be located in three-dimensional space in various orientations and at various locations. Furthermore, the projections of the components can be arranged variably relative to each other according to their poses in the image plane, and also relative to the entire object P.

[0047] The adjacent components are connected by respective flexible connections that allow translational and / or rotational relative movements, particularly with respect to the projections in the image plane. Constraints must be observed regarding the degrees of freedom. A human body is used as an example of such an object P.

[0048] The term "camera" also refers to other optical sensors capable of capturing an image of an object. The image essentially corresponds to a vertical projection of the object. Such images contain very different information content depending on the object's pose and the camera's viewing direction, although the projective relationships of the respective object's components must meet strict conditions. For example, the hand is located next to the forearm because it is connected to it.

[0049] Modeling is achieved through a hierarchically ordered set of components. The object is placed at the lowest level (Level 0) of the hierarchy, where the object's projections are stored in various poses and viewing directions. Components of the next higher level (Level 1) are assigned to the object. Each component of this component, in turn, is assigned subcomponents of the next higher level (Level 2). This continues up to the highest level (Level 3).

[0050] The arrangement of components provided in different poses is also called a component assembly. Each such component assembly is strictly defined as a projective region containing subcomponents. All components and component arrangements are semantic.

[0051] A respective component C j has two defining characteristics: 1) C jis a part of the component assembly. An example is a human body and its parts, such as the torso and extremities. Another example is a human face and its parts, such as eyes, ears, etc. 2) C j is connected to at least one other component by means of flexible or fixed connections. The flexible connections allow rotational and translational degrees of freedom, subject to constraints that restrict the degrees of freedom of movement. The fixed connections do not allow any degrees of freedom of movement.

[0052] Accordingly, a semantic composite hierarchy of objects in different poses can be expressed as a directed acyclic graph, as in Fig. 1. The nodes of this graph are individual components or arrangements. Each node of the graph contains the projections of different poses, especially from different viewing directions.

[0053] The components of the next higher level are displayed at a higher resolution than the components of the lower level. Components of the same level are connected to each other by links.

[0054] To eliminate the influences of global illumination changes and local illumination changes, such as shadows, global normalization, particularly gamma normalization, and overcomplete feature matching are used. In particular, 2D filters, i.e., two-dimensional spatial filters based on complex integrals of the Fourier field of gradients, are used, as shown in [1]. The use of integrated Fourier fields achieves rotation invariance.

[0055] By using multiple projections of each component, with different viewing directions and poses, object detection is enabled despite the many possible rotational positions and pose-dependent orientations of the component. Scaling effects are mitigated by a novel integration of hierarchy in maps of different resolutions, especially feature maps.

[0056] The following steps are carried out during the training phase: For a hierarchical model with L levels (Level 0 to Level L), where L ∈ Z is defined for each level (Level I) and for each component A) The circular perimeter of the discrete pixels is minimized, which is necessary for an overcomplete description of the component. Globally, the number N SZ various forms minimized. B) Then the component is defined by: Ci=[S;r]

[0057] Where S is a set of spatial one-to-one relationships between components of the same plane. r is the radius of the circular perimeter of the component or the largest extent of the component's image in the image plane. A spatial relationship s from S refers to the pairing to a component C j the same level.

[0058] The spatial relationships between the component projections are determined in the learning phase, initially assuming a normal distribution, especially in the initial offline training phase, as known from [3]. In a preferred embodiment, the spatial relationships are nonlinear functions that are learned in the training phase using a neural network, such as a multilevel perceptron (MLP) or a support vector machine (SVM). For a positive example image x p with connected components C i and C jwhich are localized at p i and p j , where p = (r, θ) is score1=Fs(pi,pj) Rs(Ci,Cj)={Fs, ifscore1−t1>00,otherwise

[0059] Finally, for each component C i a filter F is defined. Filter F is a function learned in the training phase: score2=F(h(x→)) y2={1,if score2−t2>0,amotherwise h(x→) is a subwindow at position p(x, y, s) of the Fig. 3 shown map with different resolutions.

[0060] As in Fig. As shown in Figure 3, images with lower resolutions are generated from the image captured by the camera. The search begins in the plane with the lowest resolution. The search continues only in the lowest H planes, i.e., the identification planes. In this process, dense feature maps with different resolutions are calculated. Such a feature map with different resolutions is defined as h(X→) where h(X→)=[xys sz]T,(x, y, s) is the position in the scaling space whose number of levels, to which different resolutions are assigned, is sz. The number of maps according to Fig. 3 is preferably kept to a minimum.

[0061] In the offline training phase, two instances of the learning algorithm or neural network must be trained: the filters F and the spatial relationships between the components of the same level. In particular, the aforementioned instances of the learning algorithm or neural network, i.e., support vector machines or multilevel perceptrons, can be used for this purpose. For each component C i a filter F i learned. For each spatial relationship, a relationship R s learned.

[0062] When evaluating the captured image, the following steps are carried out recursively, especially online, whereby the hierarchical model has L+1 levels, i.e. level 0 to level L and H identification levels, where 0 is the level that has the sparsest, i.e. coarsest resolution: 1) In a first step, the cards h fdetermined with L+1 hierarchy levels, i.e. levels or pyramid levels, and a number of magnifications SZ. 2) Then for each level, i.e. for each Level I, in h f from the lowest resolution level s = 0 to H-1, a) for each component C i a level, i) The value score2 = F i (h f ) is determined for each position on the respective image. Thus, a detection of component C i executable in the identification levels. 3) For the further levels, i.e. levels H to L, when a threshold value t2 is exceeded, i.e. when the term is positive (score2 - t2), the positions of the component C i assigned component CiH−1 the next level, i.e. P(x, y, H-1, sz) for the components CiH−1, and the previous steps are executed recursively. 4) For each level / from 0 to L, a) if the threshold value t2 is exceeded, i.e. if the term is positive (score2 - t2), for the component Cil the value scoreCil determined according to (6), where ckl+1∈K, and where K is the set of subcomponents of Cil and where Cjl∈J, where J is the amount of components Cil paired components and w l a level-specific weighting factor. 5) The objects are then evaluated by determining the reliability value scoreCil detected. This is determined according to: scoreCil={Wl∑jRs(Cil,Cjl)∑kscoreCki+1if K≠ϕ∑jRs(Cil,Cjl),otherwise

[0063] Various hardware architectures can be used to implement the method, such as a microprocessor, a digital signal processor, a graphics processing unit, or the like. Parallelization improves performance. A cascaded implementation is also possible in the case of a microprocessor-based architecture. In a preferred embodiment, logic circuits are used in an FPGA-based processor system. REFERENCES [1] K. Liu, H. Skibbe, T. Schmidt, T. Blein, K. Palme, T. Brox, and O. Ronneberger, “Rotation invariant hog descriptors using fourier analysis in polar and spherical coordinates,” International Journal of Computer Vision, vol. 106, no. 3, pp. 342-364, 2014. [2] Y. Yang and D. Ramanan, “Articulated human detection with flexible mixtures of parts,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 35, no. 12, pp. 2878-2890, 2013. [3] K. Mikolajczyk, C. Schmid, and A. Zisserman, „Human detection based on a probabilistic assembly of robust part detectors,“ in Computer Vision-ECCV 2004. Springer, 2004, pp. 69-82. [4] B. Wu and R. Nevatia, „Detection of multiple, partially occluded humans in a single image by bayesian combination of edgelet part detectors,“ in Computer Vision, 2005. ICCV 2005. Tenth IEEE Inter-national Conference on, vol. 1. IEEE, 2005, pp. 90-97. [5] P. Felzenszwalb, D. McAllester, and D. Ramanan, „A discriminatively trained, multiscale, deformable part model,“ in Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on. IEEE, 2008, pp. 1-8. [6] P. F. Felzenszwalb, R. B. Girshick, and D. McAllester, „Cascade object detection with deformable part models,“ in Computer vision and pattern recognition (CVPR), 2010 IEEE conference on. IEEE, 2010, pp. 2241- 2248. [7] N. Dalal and B. Triggs, „Histograms of oriented gradients for human detection,“ in Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on, vol. 1. IEEE, 2005, pp. 886-893. [8] C. Steger and M. Ulrich, „Hierarchical component based object recognition,“ Jul. 3 2007, US Patent 7 239 929 B2. [9] A. Hofhauser and C. Steger, „System and method for deformable object recognition,“ Sep. 4 2012, US Patent 8 260 059 B2.

[10] D. Nistér and H. Stewénius, „Scalable object recognition using hierarchical quantization with a vocabulary tree,“ May 25 2010, US Patent 7 725 484 B2.

[11] DUAN, K.: A Multilayer Composite Model for Human Pose Estimation, Proceedings of the British Machine Vision Conference, Vol. 2 , 2012, Paper 116, Seite 1 bis 11.

[12] Felzenszwalb, PF: Object detection with discriminatively trained part-based models, IEEE Transactions on Pattern Analysis and machine intelligence, Vol. 32, No. 9, 2010, pages 1627 - 1645.

[13] Wang, Y.: Learning Hierarchical Poselets for Human Parsing, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2011, pages 1705-1712. List of reference symbols Level 0 top level Level 1 top level Level 2 top level Level 3 top level 1 Definition of the components and their relationship to each other 2 Spatial definition of the respective component 3 images for the learning phase, especially high-resolution images for the learning phase 4 Training phase for the detection of components 5 Training phase for the relationship between the components 6 Decision whether the lowest scale has been reached 7 Offline model values, especially weight factors 8 Changing the image magnification or scaling 9 End 20 search windows SZ1 large scale SZ2 medium scale SZ3 small scale

Claims

[1] Method for detecting an object in an image taken by a camera, wherein the method is carried out using a computer, where computer-modeled components Ckl are stored hierarchically, where I is the hierarchy level indexed from a lowest level I = 0 to a highest level 1 = L and is a natural number and k numbers the components of a respective hierarchy level and is a natural number, where L is a natural number that indicates the highest hierarchy level, where each component Ckl a hierarchy level 1 which is lower than the highest hierarchy level L, at least one component C m l+1 is assigned to the next higher hierarchy level I + 1, where m is a natural number which represents the associated components C m l+1 indicates, where C1 0 the object is modelled, so for the value 1 = 0 the modelled component corresponds to the object, where the image is searched pairwise for components of the same hierarchy level, and from this each component a reliability value scoreCil, so the reliability of the detection of the respective component Cil descriptive value, is assigned where i is a natural number, where a weighting factor w l which was previously determined by a neural network in a learning phase, To determine the reliability value, a weighting factor is taken into account, which is calculated from a sum of the reliability values scoreCIl+1 of the respective component Cil assigned components of the next higher level, where the reliability value of the component Cil is determined according to scorecil={wl∑jRs(Cil,Cjl)∑kscoreCkl+1,if K≠ϕ ∑jRs(Cil,Cjl),otherwise where K is the set of subcomponents of Cil is and where Cjl∈J, where J is the set of coupled components of Cil is and where R s is a digitized measure which depends on the specific image positions of the respective component pair (Cjl,Cjl), in particular, this measure is not zero if the two components have a predetermined permissible distance from each other and a predetermined permissible relative orientation to each other. [2] Device for carrying out a method according to claim 1, characterized bythat a computer is connected to a camera in such a way that the images taken by the camera can be evaluated by the computer.

Citation Information

Patent Citations

  • Hierarchical component based object recognition

    US7239929B2

  • Scalable object recognition using hierarchical quantization with a vocabulary tree

    US7725484B2

  • System and method for deformable object recognition

    US8260059B2