ADAPTIVE DEPTH ESTIMATION DEVICE AND ASSOCIATED AUTONOMOUS MOBILE ROBOT

By integrating an optical unit to account for camera-specific properties, the depth estimation method adapts to diverse camera types, overcoming geometric constraints and improving depth estimation for autonomous mobile robots.

FR3148311A1Pending Publication Date: 2024-11-01VISUAL BEHAVIOR
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
FR2023004264
Authority / Receiving Office
FR · FR
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-04-27
Publication Date
2024-11-01

AI Technical Summary

Technical Problem

Existing depth estimation methods for autonomous mobile robots are limited by geometric constraints on camera positioning and type, and are ineffective for transparent, reflective, or textureless surfaces, and cannot adapt to diverse vision systems like fisheye or omnidirectional cameras.

Method used

An optical unit is integrated into a recurrent neural network to account for the optical properties of various camera systems, allowing the estimation of depth using inverse depth matrices, enabling adaptation to different camera types and orientations.

Benefits of technology

The solution allows depth estimation across various vision systems, including non-standard cameras like fisheye and omnidirectional, without the need for recalculating optical properties during processing, enhancing adaptability and efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

The invention relates to a device for estimating depths from at least two images (img1, img2) captured by at least one camera, the device comprising a recurrent neural network implemented on a computer and comprising: – a module (10a-10c) for extracting features from the images (img1, img2); – a matching unit (11) configured to determine possible evolutions of position and displacements (M1, M2, Mn) between the features of the images (img1, img2); – a recurrent unit (13) enabling the determination of a correction (Δp-1) of an inverse depth matrix (P-1) as a function of the possible evolutions of position and displacements between the features of the images from the matching unit and context elements;and – an optical unit (16), integrating the optical properties (Opt1, Opt2, Pos), configured to transmit, to said recurrent unit, the possible evolutions of position and displacements by integrating said optical properties. Figure for the abbreviation: Fig.4.;
Need to check novelty before this filing date? Find Prior Art

Description

Title of the invention: ADAPTIVE DEPTH ESTIMATION DEVICE AND AUTONOMOUS MOBILE ROBOT PARTNER Field of invention

[0001] The invention relates to the field of computer vision, and more particularly, to a device for estimating the depths of an image by means of at least two images of a scene. These images may come from one or more cameras, for example the cameras of an autonomous mobile robot.

[0002] These autonomous mobile robots are used in a large number of fields, typically land vehicles, aerial vehicles or even marine or underwater vehicles. Thus, the invention can find an application in each of these fields.

[0003] The invention relates more particularly to improving the understanding of the environment for the detection and avoidance of obstacles for autonomous navigation applications.

[0004] Furthermore, the invention specifically aims at an “adaptive” depth estimation device, i.e. one that can adapt to a wide variety of autonomous mobile robots, in particular with diverse and varied vision means. State of the art

[0005] To analyze the environment of an autonomous mobile robot, stereo cameras or time-of-flight cameras are among the solutions commonly used to estimate the distance of elements close to the robot by determining the depths of these elements on the images captured by the robot. However, these solutions impose numerous geometric constraints on the position and type of camera used. In addition, stereoscopic approaches are by construction limited for the detection of transparent, reflective or textureless surfaces. On the other hand, in mobile robotics, the need to perceive the world at 360° coupled with the variable shapes and constraints of robots limits the use of cameras subject to a geometric position constraint.

[0006] To detect the distance of objects close to a robot, 2D or 3D lidar cameras are typically a frequently used alternative but remain more expensive in terms of the number of points per square meter. They can also, in some cases, miss fine obstacles or certain surfaces.

[0007] Using conventional cameras, it is also possible to estimate the depths of the elements of an image from several images, taken by two cameras or at two consecutive times by a single camera.

[0008] To do this, the publication “RAFT: Recurrent Ail-Pairs Field Transforms for Optical Flow”, Zachary Teed and Jia Deng (2020) proposes an optical flow estimation device as illustrated in [Fig.l] of the state of the art, that is to say a device for estimating the displacement between pixels from one image to another. This device uses two images imgl, img2 introduced into a recurrent neural network. This neural network comprises two modules lOa-lOb for extracting the characteristics of the first and second images imgl, img2 as well as a module 10c for extracting the context elements of the first image imgl. From the elements extracted from the modules lOa-lOb, a correspondence unit 110 determines a correspondence matrix.In the example of this publication, the matching unit 110 calculates four-dimensional matrices Ml, M2, Mn to integrate the possible position evolutions and the induced row and column displacements. In a variant of this publication, the publication “RAFT-Stereo: Multilevel Recurrent Field Transforms for Stereo Matching”, Lahav Lipson & ail (2021) proposes to use a single camera at two distinct times as well as a three-dimensional matrix to integrate the possible position evolutions and the row displacements only in order to estimate the disparity between the images.

[0009] In these two publications, from this correspondence matrix, a recurrent unit 130 is implemented to determine a disparity matrix D containing distance estimates between each pixel of the two images imgl, img2. To do this, the recurrent unit 130 calculates, at each occurrence, a disparity differential Ad to be applied to the next occurrence to obtain the most coherent disparity matrix D possible as a function of the displacements measured between the two images imgl, img2. To do this, the internal state Ei of the reflections of the recurrent unit 130 can be propagated to each occurrence. Thus, the disparity differential Ad is introduced into a summer 14 causing, at each occurrence, the disparity matrix D initially set to a first value, typically 0, to evolve.

[0010] For each occurrence, a hypothesis generator 120 uses the disparity matrix D previously fixed to search, in the correspondence unit 110, for the possible evolutions of position and corresponding displacements in the matrices M1, M2, Mn.

[0011] More precisely, as illustrated in [Fig.2], the hypothesis generator comprises a first module 120a which determines a set of displacement hypotheses between the two images imgl, img2.

[0012] Based on these hypotheses, a second module 120b extracts the possible corresponding evolutions from the matrices M1, M2, Mn calculated by the correspondence unit 110, in order to transmit them to the recurrent unit 130.

[0013] Thus, at the end of the different occurrences, it is possible to obtain the matrix of disparities D the most relevant so that the first image imgl can be represented as a function of the depth variations estimated from this disparity matrix D.

[0014] In addition to these elements illustrated in these publications, the images imgl, img2 introduced as input to the recurrent neural network also undergo rectification, in the modules 15a, 15b, to normalize these images with respect to intrinsic matrices Optl, Opt2 and extrinsic matrices Pos of the cameras.

[0015] The intrinsic matrices Optl, Opt2 illustrate the physical parameters of each camera: focal length, distortion... The extrinsic matrix Pos illustrates the relative position between these cameras.

[0016] Thus, the rectified images imgl, img2 are normalized at the input of the correspondence unit 110 and the module 10c.

[0017] The learning of the recurrent neural network is therefore carried out on the basis of this specific architecture with normalized images.

[0018] However, the rectifications carried out by the modules 15a, 15b are effective only for cameras whose optics have uniform correspondences between the pixels and the possible depth variation. Typically, as illustrated in [Fig.3], the rectification is conventionally used with two pinhole cameras 17a, or "pinhole" in the English literature, configured to capture the same scene Sc with a lateral offset between the two cameras 17a. Thus, this configuration makes it possible to estimate the depth in the overlap zone Zr between the two cameras 17a.

[0019] However, this overlap zone Zr can be reduced and, with the rectifications carried out, it is not possible to use cameras with spatially variable projections, such as hypergone type cameras, or "fisheye" in the Anglo-Saxon literature, or omnidirectional.

[0020] Furthermore, it is also not possible to use cameras that are offset from each other with variations other than a lateral offset. It follows that this estimation device cannot adapt to all types of vision systems.

[0021] Taking into account the state of the art, the technical problem of the invention is therefore to provide a depth estimation device which can be adapted to all types of vision systems so that it can be implemented in various autonomous mobile robots, independently of the vision system used. Statement of the invention

[0022] To address this technical problem, the invention proposes using an optical unit integrating the optical properties of the vision system so as to transmit to the recurring unit the possible evolutions of position and displacements resulting from the correspondence unit by integrating said optical properties.

[0023] To do this, the distance hypothesis generator is calculated based on the inverse of the depth and the recurrent unit is adapted to determine a correction of an inverse depth matrix.

[0024] From this inverse depth matrix, it is possible to obtain, after the different occurrences, the depth of an image by means of at least two images of a scene.

[0025] Thus, according to a first aspect, the invention relates to a device for estimating depths from at least two images captured by at least one camera, the device comprising a recurrent neural network implemented on a computer and comprising: - a module for extracting features from a first image; - a module for extracting features from a second image; - a module for extracting context elements from one of the images; and - a correspondence unit configured to determine possible evolutions of position and displacements between the characteristics of the images.

[0026] The invention is characterized in that said recurrent neural network also integrates: - a recurrent unit for determining a correction of a depth inverse matrix as a function of possible changes in position and displacements between the characteristics of the images from the correspondence unit and the context elements from the context element extraction module; at each occurrence of the recurrent unit, said depth inverse matrix being modified in a summator; - a depth inverse hypothesis generator configured to transmit, to said recurrent unit, the possible evolutions of position and displacements resulting from the correspondence unit according to hypotheses generated on the basis of said depth inverse matrix; and - an optical unit, integrating the optical properties of the at least one camera used to capture the images, configured to transmit, to said recurring unit, the possible changes in position and movements from the correspondence unit by integrating said optical properties; said recurrent neural network being trained as a function of said optical properties of the optical unit.

[0027] The invention thus makes it possible to provide a depth estimation device which can be implemented for all types of vision system. Thus, the architecture of the depth estimation device does not need to be adapted for a robot. autonomous mobile integrating cameras with spatially variable projections, such as fisheye or omnidirectional type cameras.

[0028] Unlike the publications “RAFT: Recurrent Ail-Pairs Field Transforms for Optical Flow”, Zachary Teed and Jia Deng (2020) and “RAFT-Stereo: Multilevel Recurrent Field Transforms for Stereo Matching”, Lahav Lipson & ail (2021) in which the hypothesis generator generates several matrices of possible disparities between the pixels of the images, the use of an optical unit in the recurrence loop does not allow the use of this notion of disparity since the images are not rectified.

[0029] To address this issue, the distance hypothesis generator and the recurrent unit use the inverse of the depth. The invention stems from an observation that this notion of inverse of the depth is effective with the structure of the publications “RAFT: Recurrent All-Pairs Field Transforms for Optical Flow”, Zachary Teed and Jia Deng (2020) and “RAFT-Stereo: Multilevel Recurrent Field Transforms for Stereo Matching”, Lahav Lipson & ail (2021) to integrate an optical unit upstream of the recurrent unit.

[0030] Several methods are possible for integrating this optical unit.

[0031] In a first embodiment, said optical unit is placed in said correspondence unit so that the possible evolutions of position and displacement are calculated as a function of said optical properties. In this embodiment, the depth inverse hypothesis generator can therefore recover the possible evolutions of position and displacement to be transmitted to the recurrent unit by integrating the optical properties because the possible evolutions are calculated by integrating these optical properties. Thus, when the possible evolutions are calculated and the recurrent unit is used as a function of the various depth inverse hypotheses, it is no longer necessary to recalculate the influence of the optical properties during the different occurrences, thus limiting the processing time of these optical properties during the occurrences of the recurrent unit.This implementation requires an increase in calculations before using the recurrent unit but simplifies these calculations during the different occurrences. Depending on the number of occurrences and the number of depth inverse hypotheses used, this option can therefore prove interesting.

[0032] In a second embodiment, the optical unit can be placed in the depth inverse hypothesis generator between a first module which determines a set of displacement hypotheses between the images and a second module which extracts the possible evolutions corresponding to the hypotheses of the first module. In this embodiment, the possible evolutions are searched in the correspondence unit according to the optical properties. Thus, the calculation of the cor unit Correspondence requires a computational cost similar to that of the state of the art and each occurrence requires an additional analysis cost because it is necessary to calculate the evolution of the displacement hypotheses between the images as a function of the optical properties. This embodiment can nevertheless be interesting when the computational costs of the optical properties are limited.

[0033] In another embodiment, the correspondence unit is placed in the optical unit between the first module and the second module of the depth inverse generator. In this embodiment, the possible evolutions are searched in the optical unit by the correspondence unit according to the optical properties. More precisely, instead of calculating all the possible evolutions of position and displacement before implementing the different occurrence loops, this embodiment proposes to calculate them at each occurrence and according to the different hypotheses of displacement between the images. When calculating these evolutions, the optical properties are taken into consideration.Thus, this embodiment allows the different occurrence loops to be implemented directly, but the processing time for each occurrence loop is greater than in the other embodiments because it is necessary to calculate the possible evolutions of position and displacement for each hypothesis. Once again, depending on the number of hypotheses and occurrences, this solution may be interesting in terms of computational costs.

[0034] In addition to these different possible implementations of the optical unit, it should be noted that the two images can be captured by a single camera at two consecutive times or by several cameras, typically two cameras separated by a fixed distance. Furthermore, with the invention, the orientation between the different cameras can be variable and it is no longer necessary for them to be oriented in parallel to implement the invention.

[0035] Although the invention can be implemented with at least two images, it is also possible to use more than two images to increase the overlap area in which it is possible to estimate the depth. Thus, when three images are captured by three cameras, a module for extracting features from a third image makes it possible to transmit the features of the third image to the correspondence unit or to the optical unit. In addition, the depth estimation device can also integrate a fusion module making it possible to determine which images must be used to estimate the depth in the event of overlap between several possible images.

[0036] According to a second aspect, the invention also relates to an autonomous mobile robot comprising movement means, a vision system integrating at least one camera, and a device for estimating depth from at least two images. captured by said at least one camera of the vision system according to the first aspect of the invention. Brief description of the figures

[0037] The invention will be clearly understood on reading the following description, the details of which are given by way of example, and developed in relation to the appended figures, in which identical references refer to identical elements:

[0038] [Fig.l] is a schematic representation of a device for estimating depth from two images of the state of the art;

[0039] [Fig.2] is a detailed schematic representation of the state-of-the-art depth estimation device of [Fig.l];

[0040] [Fig. 3] is a schematic sectional representation of a state-of-the-art vision system in which a depth estimation device corresponding to Figures 1 and 2 can be implemented;

[0041] [Fig.4] is a schematic representation of a device for estimating depth from two images according to a first embodiment of the invention;

[0042] [Fig.5] is a schematic representation of a device for estimating depth from two images according to a second embodiment of the invention;

[0043] [Fig.6] is a schematic representation of a device for estimating depth from two images according to a third embodiment of the invention;

[0044] [Fig.7] is a schematic representation of a vision system according to a first embodiment of the invention making it possible to implement a depth estimation device according to one of figures 4 to 6;

[0045] [Fig.8] is a vision system according to a second embodiment of the invention making it possible to implement a depth estimation device according to one of figures 4 to 6;

[0046] [Fig.9] is a schematic representation of a vision system according to an embodiment of the invention with three cameras; and

[0047] [Fig. 10] is a schematic representation of a device for estimating depth from three images according to one embodiment of the invention. Detailed description of the invention

[0048] [Fig.4] illustrates a depth estimation device from two images imgl and img2 introduced into a recurrent neural network. This recurrent neural network comprises two modules 10a and 10b for extracting the characteristics of the first and second images imgl img2 as well as a module 10c for extracting the context elements of the first image imgl. These modules 10a-10c can cor meet those of the state of the art and, for example, the lOa-lOc modules can be implemented by Resnet 50 type networks.

[0049] From the characteristics extracted from the modules 10a and 10b, a correspondence unit 11 determines a correspondence matrix. More precisely, the correspondence unit 11 determines in this correspondence matrix a set of possible evolutions of position and displacement Ml, M2, Mn between the characteristics of the images imgl and img2.

[0050] In the embodiment of [Fig.4], these possible evolutions of position and displacement Ml, M2, Mn are calculated as a function of the optical properties because the optical unit 16 is integrated into the correspondence unit 11. Thus, for each possible evolution of position or displacement, the characteristics of the images imgl and img2 are transformed as a function of the optical properties Optl, Opt2, Pos transmitted to the optical unit 16.

[0051] More precisely, the optical properties integrate the intrinsic matrices Optl, Opt2 which illustrate the physical parameters of each camera such as the focal length or the distortion, as well as the extrinsic matrix Pos which illustrates the relative position between these cameras. It should be noted that the relative position between the cameras can integrate several dimensions so that it is possible, in the extrinsic matrix Pos , to integrate both distance and orientation offsets.

[0052] When the possible evolutions of position and displacement Ml, M2, Mn are calculated by integrating the optical properties, the recurrent neural network can implement a recurrent unit 13 making it possible to determine a correction Ap-1 of a depth inverse matrix Pl. To do this, the recurrent unit 13 calculates at each occurrence a depth inverse differential Ap-1 to be applied to the next occurrence to obtain the depth inverse matrix Pl that is as coherent as possible as a function of the displacement measured between the two images imgl and img2. To do this, the internal state Ei of the reflections of the recurrent unit 13 can be propagated at each occurrence. Thus, the depth inverse differential Ap-1 is introduced into a summer 14 causing, at each occurrence, the depth inverse matrix Pl initially set to a first value to evolve, for example 0 in the case of [Fig.4].It is noted that the recurrent unit 13 differs structurally from the recurrent unit 130 of the prior art because the calculation of a depth inverse matrix Pl is different from the calculation of a disparity matrix D. .

[0053] The recurrent unit 13 may correspond to that described in the publications “RAFT: Recurrent All-P airs Field Transforms for Optical Flow”, Zachary Teed and Jia Deng (2020) and “RAFT-Stereo: Multilevel Recurrent Field Transforms for Stereo Matching”, Lahav Lipson & ail (2021). Thus, the recurrent unit 13 may integrate a GRU, for “Gated Recurrent Unit” in the Anglo-Saxon literature, convolutional in concatenating the inputs. Such a recurrent unit 13 combines the advantages of GRU and CNN neural networks, for "Convolutional Neural Network" in the English literature. Alternatively, the structure of the GRU and the CNN can be adapted to the specific need of generating the inverse depth matrix Pl.

[0054] Furthermore, for each occurrence, a hypothesis generator 12 uses the previously calculated inverse depth matrix Pl to search, in the correspondence unit 11, for the possible evolutions of position and displacement corresponding to the matrices Ml, M2, Mn. More precisely, the hypothesis generator 12 corresponds to a depth inverse hypothesis generator and it preferentially integrates a first module 12a which determines a set of displacement hypotheses between the images imglet img2 and a second module 12b which extracts the possible evolutions corresponding to the hypotheses of the first module 12a. Thus, the first module 12a can be configured to generate, at each occurrence, a number of hypotheses between 60 and 100, for example 80.

[0055] These hypotheses can be generated in a non-uniform manner around the previously calculated depth inverse matrix Pl. When all of these hypotheses are generated, the second module 12b extracts the possible evolutions corresponding to the hypotheses in order to transmit them to the recurrent unit 13.

[0056] With this particular structure of recurrent neural networks, it is possible to carry out the learning of the recurrent neural network as a function of the optical properties integrated in the optical unit 16. At the end of this learning phase, it is possible to estimate the depth between two images provided that these two images imgl img2 have an overlap zone Zr.

[0057] As illustrated in [Fig.7], by using two pinhole type cameras 17a, it is possible to improve the overlap area compared to the state of the art since one of the two cameras can now be angularly oriented. Furthermore, as illustrated in [Fig.8], it is also possible to use other types of cameras, such as an omnidirectional camera 17b.

[0058] The embodiment of [Fig. 5] illustrates an alternative implementation of the invention in which the optical unit 16 is placed in the depth inverse hypothesis generator 12. More precisely, the optical unit 16 is placed between the first module 12a and the second module 12b so that the possible evolutions sought in the correspondence unit 11 do not take into account the optical properties but that these are taken into account at each occurrence in the depth inverse hypothesis generator 12. Thus, the optical unit 16 also receives the optical properties Optl, Opt2, Pos and it is implemented at each occurrence so as to transmit the hypotheses from the first module 12a to the second module 12b by integrating optical modifications.

[0059] The embodiment of [Fig. 6] differs from the embodiments of figures 4 and 5 in that the correspondence unit 11 is placed in the optical unit 16, itself placed between the first module 12a and the second module 12b. In this embodiment, the matrices M1, M2, Mn are not calculated before implementing the different occurrences but at each occurrence according to the displacement hypotheses generated by the first module 12a. The possible evolutions corresponding to these hypotheses are thus calculated by the correspondence unit 11 by integrating the optical properties Opt1, Opt2, Pos.

[0060] Regardless of the implementation selected, the invention can be implemented for pinhole, fisheye or even omnidirectional cameras.

[0061] Furthermore, the images imgl and img2 can be captured by a single camera at two consecutive times or by two separate cameras separated by a fixed distance and a potentially variable orientation.

[0062] The embodiment of [Fig.9] also illustrates that at least three cameras 17a can be used to capture three images img1, img2 and img3 of the same scene, for example to increase the detection area. To take into account more than two images, it is possible to integrate the optical properties and characteristics of the third image img3, as illustrated in [Fig. 10]. To do this, another module 10d for extracting characteristics of the new image img3 is used in order to transmit to the correspondence unit 11 or to the optical unit 16 the characteristics of the new image img3.

[0063] [Fig. 10] corresponds to the embodiment of [Fig. 6] by integrating a third image img3. Thus, in this embodiment of [Fig. 10], the third image img3 is integrated into the optical unit 16. Alternatively, the third image img3 could be integrated into the correspondence matrix 11 using the embodiments of FIGS. 4 or 5.

[0064] In addition to integrating the characteristics of the third image img3, the optical properties must also integrate the intrinsic matrix Opt3 of the third camera and the extrinsic matrix must illustrate the relative position between the first camera and the second camera Posl-2 and the relative position between the first camera and the third camera Posl-3. Alternatively, it is also possible to integrate an extrinsic matrix of the relative position between the second and the third camera. However, the more the number of images increases, the more complex the matrices calculated by correspondence unit 11 or the characteristics sought by the optical unit 16 are.

[0065] Furthermore, with more than two images to be analyzed, there are also possible overlap phenomena. As illustrated in [Fig.9], several areas of overlap come from the three cameras 17a: an overlap zone Zrl in which the images from the three cameras can be used and two overlap zones Zr2 and Zr3 in which only two cameras can be used to estimate the depth. When several images can be used to estimate the depth, it is possible to use a fusion module to determine which images should be used.

[0066] Thus, multiple implementations are possible for implementing the invention and estimating depth by means of a recurrent neural network in which optical properties are integrated.

[0067] The recurrent neural network can use as input images that are not rectified and variable vision systems provided that the training of the recurrent neural network is carried out on the basis of the optical properties of the optical unit 16.

[0068] The invention thus makes it possible to integrate very effectively into an autonomous mobile robot whose technical properties are not standard. This autonomous mobile robot conventionally integrates movement means and a vision system integrating at least one camera. According to the invention, this autonomous mobile robot can be controlled using a depth estimation device from two images captured by at least one camera of the vision system as previously described.

Claims

Claims

1. Device for estimating depths from at least two images (imgl, img2, img3) captured by at least one camera (17a-17b), the device comprising a recurrent neural network implemented on a computer and comprising: - a module (10a) for extracting characteristics from a first image (imgl); - a module (10b) for extracting characteristics from a second image (img2); - a module (10c) for extracting context elements from one of the images (imgl, img2); and - a correspondence unit (11) configured to determine possible evolutions of position and displacements (Ml, M2, Mn) between the characteristics of the images (imgl, img2, img3); characterized in that said recurrent neural network also integrates: - a recurrent unit (13) for determining a correction (Ap-1) of a depth inverse matrix (Pl) as a function of the possible changes in position and displacements (Ml, M2, Mn) between the characteristics of the images (imgl, img2, img3) from the correspondence unit (11) and the context elements from the context element extraction module (10c); at each occurrence of the recurrent unit (13), said depth inverse matrix (Pl) being modified in a summator (14); - a depth inverse hypothesis generator (12) configured to transmit, to said recurrent unit (13), the possible evolutions of position and displacements (Ml, M2, Mn) from the correspondence unit (11) as a function of hypotheses generated on the basis of said depth inverse matrix (Pl); and - an optical unit (16), integrating the optical properties (Optl, Opt2, Opt3, Pos, Posl-3, Posl-2) of the at least one camera used to capture the images (imgl, hng2, img3), configured to transmit, to said recurrent unit (13), the possible changes in position and displacements (Ml, M2, Mn) from the correspondence unit (11) by integrating said optical properties (Optl, Opt2, Opt3, Pos, Posl-3, Posl-2); said recurrent neural network being trained according to said optical properties (Optl, Opt2, Opt3, Pos, Posl-3, Posl-2) of the optical unit (16).

2. A depth estimation device according to claim 1, wherein said optical unit (16) is placed in said correspondence unit (11) so that the possible evolutions of position and displacements (Ml, M2, Mn) are calculated as a function of said optical properties (Optl, Opt2, Opt3, Pos, Posl-3, Posl-2).

3. A depth estimation device according to claim 1, wherein said optical unit (16) is placed in said depth inverse hypothesis generator (12) between a first module (12a), which determines a set of displacement hypotheses between the images (imgl, hng2, hng3), and a second module (12b), which extracts the possible evolutions corresponding to the hypotheses of the first module (12a), the possible evolutions being sought in the correspondence unit (11) as a function of said optical properties (Optl, Opt2, Opt3, Pos, Posl-3, Posl-2).

4. A depth estimation device according to claim 1, wherein said correspondence unit (11) is placed in said optical unit (16) between a first module (12a) of the depth inverse hypothesis generator (12), which determines a set of displacement hypotheses between the images (imgl, hng2, hng3), and a second module (12b) of the depth inverse hypothesis generator (12), which extracts the possible evolutions corresponding to the hypotheses of the first module (12a), the possible evolutions being searched in said optical unit (16) by said correspondence unit (11) as a function of said optical properties (Optl, Opt2, Opt3, Pos, Posl-3, Posl-2).

5. Depth estimation device according to one of claims 1 to 4, wherein said at least one camera (17a-17b) corresponds to a fisheye or omnidirectional type camera.

6. Depth estimation device according to one of claims 1 to 5, in which said at least two images (imgl, hng2, hng3) are captured by a single camera (17a-17b) at at least two consecutive times.

7. Depth estimation device according to one of claims 1 to 5, in which said at least two images (imgl, hng2, hng3) are captured by at least two separate cameras (17a-17b), the two cameras being separated by a fixed distance and a potential orientation. partially variable.

8. Depth estimation device according to one of claims 1 to 5, in which the depth estimation device uses at least three images (imgl, img2, img3) captured by at least three cameras (17a-17b), a module (10d) for extracting characteristics of a third image (img3) making it possible to transmit, to said correspondence unit (11) or to said optical unit (16), the characteristics of said third image (img3).

9. A depth estimation device according to claim 8, wherein the depth estimation device incorporates a fusion module for determining which images (imgl, img2, img3) should be used to estimate the depth in the event of overlap.

10. Autonomous mobile robot comprising: - movement means; - a vision system integrating at least one camera (17a-17b); and - a device for estimating depths from at least two images (imgl, img2, img3) captured by said at least one camera (17a-17b) of the vision system according to one of claims 1 to 9.