EP2923331A1

EP2923331A1 - Method of panoramic 3d mosaicing of a scene

Info

Publication number: EP2923331A1
Application number: EP13795213.1A
Authority: EP
Inventors: Raphaël HORAK
Original assignee: Thales SA
Current assignee: Thales SA
Priority date: 2012-11-21
Filing date: 2013-11-20
Publication date: 2015-09-30
Also published as: WO2014079880A1; FR2998400B1; IL259954B; JP6483075B2; WO2014079879A2; EP2923330A2; IL238939B; US9998660B2; FR2998400A1; IL238938A0; FR2998401A1; JP2017097871A; IL259954A; EP2923330B1; EP3144881B1; DK3144881T3; US20150304558A1; US9729789B2; US20170310892A1; EP3144881A1

Abstract

The invention relates to a method of mosaicing a scene into a 3D mosaic, characterized in that at least one 3D reconstruction having been obtained and in that it furthermore comprises the following steps: A) Acquiring successive images by means of a panoramic sensor moving along a predefined mosaicing trajectory, the last image acquired being designated current panoramic image, B) Choosing one or more projection surfaces tied to the scene and on which the mosaic will be constructed, C) Textures of the panoramic images covering each 3D reconstruction, retrieving the textures originating from the current panoramic image, D) From among these textures, selecting the parts visible in each projection surface with the aid of the 3D reconstruction, E) Projecting these textures onto each projection surface and merging the textures in each projection surface so as thus to obtain a conforming mosaic on each projection surface.

Description

METHOD FOR PANORAMIC 3D MOSQUERY OF A SCENE

The field of the invention is that of the 3D mosaicing of a scene from successive panoramic images of this scene, one or more 3D reconstructions of this scene having previously been performed.

The 3D reconstruction of a scene consists in obtaining, from successive 2D images of this scene taken from different points of view, a so-called 3D reconstructed image such that at each pixel of the reconstructed image, that is to say at any point where the reconstruction declares that there is a scene element, are associated the coordinates of the point of the corresponding scene, defined in a reference X, Y, Z linked to this scene.

The classic mosaicization called 2D mosaicing consists of successive images of a scene to project successively them on a main plane of the scene and to assemble them to make a mosaic.

There are different methods of 3D reconstruction of a scene:

- Static methods using two or more static cameras targeting the same scene, these are stereoscopic or multi-scopic methods.

- Methods adapted to a moving camera. They comprise the following main steps:

- Acquisition of successive images by a device moving along a trajectory known as 3D reconstruction, and for each selected pair of images acquired,

- Rectification of these images on a rectification plan,

- Matching rectified images,

- 3D reconstruction in an intermediate plane from these matched images.

- And finally active methods not using cameras but active sensors, mobile or not.

Techniques for passive 3D scene reconstruction from cameras are described in various reference works: - R. Horaud & O. Monga. Computer Vision: Basic Tools, Hermès Editions, 1995.

http://www.inrialpes.fr/movi/people/Horaud/livre-hermes.html

- Olivier Faugeras. Three-Dimensional Computer Vision, MIT Press, 1993

- Frédéric Devernay, INRIA Grenoble, "Computer Vision 3-D" course. http://devernay.free.fr/cours/vision/

- Tébourbi Riadh, SUP'COM 2005 IMAGING 3D 08/10/2007

- "OpenCV Learning: Computer Vision with the OpenCV Library", Gary Bradsky, 2008

These books all cite 3D scene reconstruction techniques from pairs of stereoscopic images from cameras positioned at different points of view, which can be either fixed cameras positioned at different places in space, or a camera whose the position varies temporally, always with the same basic principle of matching the images of the cameras taken 2 to 2 to form a stereoscopic 3D reconstruction of the portion of space visualized by the cameras.

They also explain the principle of epipolar rectification where the focal plane image of each camera is rectified in accordance with the attitude of the camera on a so-called rectification plane to facilitate the matching between the images of the stereoscopic pair and make reconstruction possible. 3D. The process is relatively optimized by different authors but is still based on the principle that we must first correct the optical distortions of the camera and then use the relative attitudes of the two cameras to determine the correction plan from which is made in correspondence and 3D reconstruction.

Other passive 3D reconstruction techniques exist in the literature, for example so-called silhouetting techniques, which are not considered here because they apply to particular cases and require prior knowledge on the scene. In the techniques of active reconstruction of a scene we can cite those based on lidar which allow to reconstruct directly by a distance calculation the 3D mesh of the scene.

Reference works include:

- MATIS works for the IGN: "Using Full Waveform Data Lidar for Mapping of Urban Areas", Thesis of Doctorate, Clément Mallet, 2010

- "Airborne Laser and Photogrammetric Data Coupling for Three-dimensional Scenes Analysis", PhD Thesis, Frédéric Bretar, 2006.

An interesting article shows that these techniques have limits to reconstruct 3D objects of complex shape (eg concave): Planar structuring of an unstructured 3D point cloud and detection of obstacle zones, Vision interface conference, 1999, Nicolas Loémie, Laurent Gallo, Nicole Cambou, George Stamon.

Regarding mosaicing, the following reference works can be cited:

- L. G. Brown, "A Survey of Image Registration Techniques," in ACM Computing Surveys, Vol. 24, No. 4, 1992,

- "Mosaic of multiresolution images and applications", Thesis of doctorate, University of Lyon. Lionel Robinault, 2009.

If we summarize the prior art concerning the 3D reconstruction, we can say that the 3D reconstruction can be partially obtained by using:

- Couples of cameras performing a spatially stereoscopic image of the scene and merging these images to achieve a 3D reconstruction and possibly a mosaic of the scene. This solution has several disadvantages:

o cameras are difficult to calibrate (vibration problems), o imprecision in 3D reconstruction restitution because of a stereo base limited by the spacing between cameras, o a weak field and small extent restitution due to the field limited optics of the cameras. On the other hand, the finalized 3D reconstruction is not obvious, since it consists of the assembly of local 3D reconstructions (resulting from the process of stereoscopic restitution of 2 images often small field) which can be very noisy because of the limited number images that made it possible to build it, the limited field of cameras and the fact that the reconstruction plans dependent on the respective attitudes of the cameras have a geometry that is difficult to measure with precision (the relative position and geometry of the cameras used for the reconstruction 3D is often imprecise in practice because they are cameras that are 1 or 2 meters apart from each other and can vibrate relative to each other: it is even more obvious when these cameras are motorized). The precise way of assembling intermediate 3D reconstructions is never described in detail and in practice there are many errors in the finalized 3D reconstruction which remains weak in spatial and angular extent anyway (typically less than 200m x 200m in spatial extent with an angular extent typically less than 30 °).

Finally, the process of rectification and matching itself, depending on the attitudes of the cameras and passing through a preliminary step of derotation of the focal plane in the rectification process, makes it possible that there are cases where the 3D reconstruction presents holes, especially if the system exhibits temporal rotational movements.

Finally, the stereoscopic system renders poorly the plans that are almost perpendicular to one of the two cameras (this is the problem of the return of sloping roofs in stereoscopic aerial or satellite imagery).

A low or medium field camera moving, but the 3D reconstruction is limited by the course and orientation of the camera and is not omnidirectional; moreover, the reconstruction may have holes due to uncontrolled movements of the camera or non-overlapping of the camera during its movement. The algorithms used for the 3D reconstruction impose a reconstruction in a reference linked or close to the plane focal point of the camera which limits the possibilities of reconstruction (only one main plane of reconstruction and very limited reconstruction when the camera changes orientation). The result of the reconstruction is also very noisy and can present many errors due to the low overlap between images, a constant reconstruction plan of the reconstructed scene (and a camera that can deviate from this plane) and the use of algorithms that exploit for 3D reconstruction only two images separated by a relatively small distance. The mosaicization obtained by the plating on the ground of the successive images does not work and is not compliant when the scene is not flat and / or comprises elements 3D.

Active sensors that is to say with telemetry, but again the 3D reconstruction is not omnidirectional and is not necessarily segmented, the measurements being obtained in the form of point clouds difficult to use automatically. In addition, the mesh obtained by these active sensors has the disadvantage of being non-angularly dense (typically less than 4 points per m ² for airborne applications at 1 km height). The technique is not suitable for the moment to be able to produce a textured image of the scene and must almost always be corrected manually.

All of the above solutions are unsuitable for 3D mosaicing for a large 3D scene, for example covering the scene over more than 120 ° of angular extent on all sides and can be built continuously along the entire path. . The 3D instantaneous mosaics obtained have deformations and are limited in angular (typically <30 °) or spatial extent. The assembly of mosaics is complex when the terrain is 3D and the final result does not conform to the geometry of the scene.

The disadvantages of the methods of the state of the art are not limiting, other disadvantages are described in the patent. The object of the invention is to overcome these disadvantages. 3D mosaicing according to the invention is defined as a generalization of 2D mosaicing by operating this mosaic on any non-planar scene comprising a set of curved surfaces and 3D objects in relief.

The object of the invention is to be able to perform a 3D mosaic on any 3D scene from a panoramic sensor performing successive acquisitions of the scene along any trajectory and following different points of view, which ensure both a significant spatial and angular extent of mosaicing and conformity thereof. This 3D mosaic design assumes that one or more 3D reconstructions of the scene have been established on the same scene as the panoramic sensor. The proposed solution is based on the use:

a panoramic system covering a very large angular sector of the scene and can go if you want to cover the entire sphere, and the exploitation of the displacement of it in the scene with a 2D image processing obtained , and

a mosaicing system exploiting this information to produce a mosaic with a very large spatial and angular extent representing this scene according to all the points of view of the panoramic system during its movement along a path called mosaicing.

The panoramic system comprises one or more sensors whose images do not necessarily have overlapping between them, and allow to cover the entire scene to reconstruct instantly (with holes if the sensors are not overlapping) or during the movement .

The subject of the invention is a method of mosaicing a scene into a 3D mosaic. It is mainly characterized in that at least one 3D reconstruction of the scene has been obtained during the following steps of: acquisition of successive images by a sensor moving along a 3D reconstruction trajectory,

- rectification of these images on a rectification plane, matching the rectified images,

- 3D reconstruction in an intermediate plane from these matched images,

and in that it comprises the following steps:

A) Acquiring successive images by means of a moving panoramic sensor following a predefined mosaic path, the last acquired image being designated as a current panoramic image,

B) Select one or more projection surfaces linked to the scene on which the mosaic will be built,

C) Textures of the panoramic images covering each 3D reconstruction, recover the textures coming from the current panoramic image,

D) Among these textures, select the visible parts of them in each projection surface using the 3D reconstruction,

E) Project the visible textures on each projection surface and merge the projected textures on each projection surface to obtain a conformal mosaic on each projection surface.

When corrected images resulting from the 3D reconstruction and belonging to the field of the current image can be determined, and these belong to rectification planes closest to the geometric sense of the projection surfaces, then the textures recovered are those of the rectified image (s).

According to another characteristic of the invention, step C) includes a step of determining sectors of the current panoramic image that correspond to the selected projection surfaces. This characteristic is preferably applied when the previous case (with rectified images) can not be used. These steps are preferably repeated with each new 2D panoramic image acquisition.

A 3D mosaicing is thus carried out which is a generalization of the 2D mosaicing in the sense that the projection can be done on any 3D surface, which itself can be made of several flat or non-planar surfaces with discontinuities. This 3D mosaicking consists of successive 2D images of a scene (taken from different points of view) and the 3D reconstruction of the scene in the previous sense, to project and assemble the different 2D images on the geometric modeling of the 3D reconstruction, allowing to restore the whole scene in the form of a textured mosaic plated on the various 3D elements of this scene. It makes it possible to reproduce in an appropriate manner an assembly of images on any scene presenting relief or 3D elements. The reconstituted 3D mosaic is a 3D textured reconstruction of the scene.

These methods make it possible to perform 3D mosaicing over the largest possible spatial and angular extent.

According to one characteristic of the invention, each texture having a resolution, it comprises a step for determining the resolution of the textures and the fusion of the textures of the step E) is carried out according to these resolutions.

The reconstruction and mosaicization trajectories can be the same trajectory.

Preferably, the one or more 3D reconstructions have been obtained by using several reconstruction planes to benefit from the possibility of projecting the textures of the mosaic on all projection planes, in particular in very different planes of directions, and thus extend to maximum the possibility of angular and spatial extent of 3D mosaicking as indicated in the first example of 3D reconstruction method described later.

The invention also relates to a 3D mosaicization equipment of a scene that comprises: a panoramic system capable of forming 2D images of the scene, called 2D panoramic images, provided with localization means and,

connected to this panoramic system, a calculator comprising:

means for implementing the mosaicing method as defined above,

automatic processing means for complementary images possibly associated with or replaced by a man-machine interface.

In this way, a simple, accurate method is obtained, for example to produce textured maps on which measurements can be made, to reconstruct the scene over a large spatial extent (up to 180 ° x 360 °) and angular and time real, without constraints on the trajectory, on any scene (without assumptions of flat scene and without the help of previous scene model for example).

The proposed solution:

makes it possible to provide a compact solution to the problem of 3D mosaicing of the scene by requiring only one panoramic system, whereas those of the state of the art require several independent sensors or several independent trajectories with different viewing angles and are more complex to implement,

produces a mosaic of the scene

o conforms that is to say without geometric deformations and thus superimposable on a map,

o over a very wide spatial and angular extent, and without holes, o complete, that is to say that can be done in different directions of the planes which is very different from conventional processes producing only one mosaic plan in exit, and not allowing to restore or poorly restoring objects of the scene having different faces of the chosen restitution plan,

o robust thanks to the temporal redundancies implemented, o accurate thanks to the temporal stereovision which alone produces a large virtual stereoscopic base, which explains the precision,

instantaneous, in the sense that at each instant is recalculated and updated the mosaic,

compatible for example low-end MEMS attitude units when they are used to find out the trajectory, or simple means of relative displacement measurements such as an odometer or a basic GPS, compatible with large or uncoordinated movements of the sensor, which does not allow a weak field sensor, o and applying to any type of trajectory including curvilinear and in any direction.

Other benefits can be cited, such as:

allow the operator to choose any reconstruction and projection planes (for example to reproduce both what is on the ground and on the facades, or following a cylindrical projection). The solution is also suitable for the reproduction of concave objects, which is very difficult to achieve by other methods, producing textured reconstructions on which precise measurements are possible (the reproduced images are compliant),

allow any movements of the panoramic system in the scene including the approximation,

do not require any external measurements other than those measuring the relative displacement in position and attitude between 2 shots, with compatible measurement accuracy of low-end COTS instruments (MEMS control unit, basic GPS or odometer).

do not require any other information a priori on the scene to rebuild,

enable real-time operation on a PC.

Other features and advantages of the invention will appear on reading the detailed description which follows, given by way of non-limiting example and with reference to the appended drawings in which: FIG. 1 diagrammatically represents an example of equipment for implementing the 3D reconstruction and mosaicing method according to the invention,

FIG. 2 diagrammatically represents different steps of an example of a multi-plane 3D reconstruction method,

FIG. 3 schematically represents various steps of the mosaicing method according to the invention,

FIG. 4 illustrates measurement ambiguities produced by a concave object when there is only one reconstruction plane,

FIG. 5 represents an example of sectoral decomposition of a panoramic image resulting from a panoramic system,

FIG. 6 represents examples of rectified images of sectors of the panoramic image of FIG. 5 projected on different grinding planes,

FIG. 7 schematically represents, for an exemplary trajectory, examples of rectification planes, lines of sight of the panoramic system Ldv1 and Ldv2 being independent of these planes,

FIG. 8 schematically represents an example of time evolution of grinding planes and 3D reconstruction planes according to the invention, for a given trajectory.

FIG. 9 represents the result of a mosaic of successive panoramic images in the case where each projected image takes into account the 3D reconstruction according to the invention (FIG. 9b) and in the case where it does not take it into account (FIG. 9a).

The general idea of the invention consists in exploiting to the maximum the frontal angular field (= whose direction of line of sight travels a plane in the direction of the movement of the panoramic system) and transverse field (= whose direction of line of sight travels a plane in the direction perpendicular to the movement of the panoramic system) of a panoramic system moving in a scene according to a known trajectory, to restore according to different points of view, the relief and the texture of this scene.

The exploitation of the transverse field is done by reconstructing the relief and the texture according to all the lateral points of view seen by the system. panoramic view that can be presented to the operator according to different reconstruction plans.

The exploitation of the frontal field is done exploiting the temporal fusion of the previous reconstructions and mosaiquages observing the objects of the scene according to different points of view. These different reconstructions and mosaics of an object seen according to different points of view make it possible to produce an extended, precise and consistent overview of the scene that can be presented to an operator according to different points of view.

Exploiting the temporal stereoscopy in different angular directions that the displacement of the panoramic optical (or optronic) system moving in a scene can produce, simultaneously makes it possible to produce a 3D reconstruction of the projectable scene in different directions and a compliant and multi-surface mosaic of that -this.

The proposed solution uses the following new concepts: temporal stereoscopy with panoramic system, which differs from conventional stereoscopy using two low field cameras,

simultaneous rectification according to different planes whose directions are freely chosen, which differs from the conventional rectification which is only on a single plane whose direction is imposed by the line of sight direction of the two sensors used. Another innovation is the direct rectification that is done directly between any part of the 2D image of the panoramic system and the rectification plane chosen, unlike the conventional rectification used in stereovision which imposes an intermediate recovery plan, which produces information loss,

merging of intermediate reconstructions using very different lines of sight, to gain precision and robustness,

a trusted map related to the hierarchy of the quality of the information extracted from the 2D images targeting an object of the scene on very different points of view and which is directly related to the temporal exploitation of the 2D images of a moving panoramic system. Exploitation of a 3D reconstruction to treat only the visible parts of the textures in a 3D mosaic applied on any 3D surface, which can include complex objects, including concave

- 3D mosaicing process simultaneously projecting the textures of a panoramic image in different directions

3D mosaicing process that can exploit images rectified in place of the sectors of the current panoramic image More specifically, the method is implemented by means of equipment, an example of which is shown in FIG. 1, which comprises:

a panoramic system 1 capable of forming 2D panoramic images of the scene, comprising a sensor 14 associated with an optical 11 and provided with locating means such as a GPS 12 and an inertial unit 13, and

connected to this panoramic system, a computer 2 comprising: means 21 for implementing the mosaicking method

3D as described, and

automatic processing means for complementary images possibly associated with or replaced by a man-machine interface 22.

As indicated in the preamble, there are different methods of 3D reconstruction of a scene.

In order to show the feasibility of obtaining 3D data obtained on the extended space of the scene that we want to mosaic, a first example of 3D reconstruction process of a scene using 2D panoramic images of the scene is proposed.

According to this first example, which is preferred, the 2D images come from the panoramic system 1 traveling along a known reconstruction trajectory, which can be measured in relative image-to-image as the displacement progresses, thanks to the means location and calculator 2.

The system is panoramic in that it provides a 2D panoramic image. For this purpose, it may include a lens 1 1 large fish-eye type field, or any conventional optical medium or catadioptric large field capable of providing a 2D panoramic image, or from a weaker field optics but animated more or less ample movements to capture the different portions of scenes that we want to rebuild in their entirety. A 2D image covering a large field greater than 60 ° is for example obtained from a 45 ° field system 1 with a movement allowing it to cover this total field of 60 °. The choice of the panoramic system 1 technology is not limited: it can be passive but it can be generalized to an active system as long as it allows to implement the multi-plane fusion step presented here. -above ; this also includes large-field optics exceeding 360 ° x180 ° or full-sphere optics (for example, 2 fisheye-optic sensors that explore the full sphere of observability). This panoramic system may also comprise a set of optical sensors that are not independent of each other, covering together a determined panoramic angular coverage, for example identical from one image to the other, or maximum. All these optical sensors may not be overlapping, that is to say that the overall image obtained at a time by this set is not continuous (may include holes), the "holes" being filled when moving this set. An example of a 2D panoramic image obtained with fish-eye type optics, and sectors (5 in this example) is shown in FIG.

This trajectory can be arbitrary and / or determined as the 3D reconstruction process unfolds. The trajectory can be calculated as the panning system is moved by locating means measuring the relative displacements of position and attitude of the panoramic system in the scene such as GPS 12, inertial unit 13 or other. This movement can be controlled by an operator via a human-machine interface 22 or be autonomous. The images thus obtained are such that the image of at least one point of the scene is in at least 3 panoramic images respectively obtained according to different panoramic-point system directions of the scene.

The processing step of these 2D panoramic images respectively obtained at successive instants, by the processing unit 21 comprises the following substeps described in connection with FIG. a 3D reconstruction process preceding the actual mosaicization process.

Step a) Determine reconstruction plans in the scene. Different reconstruction plans Cj can be chosen to establish the 3D reconstructions by highlighting different aspects of the scene, for example to cover the scene over a large spatial and angular extent, or which will allow to have a better representation of it . Each reconstruction plane in the scene is determined experimentally by the operator or automatically according to the scene already restored; it can also be determined according to the trajectory of the panoramic system, typically around the average of this trajectory calculated between two shots, and according to the complexity of the scene and is independent of the line of sight of the panoramic system.

In the total absence of an initial 3D reconstruction and in the initialization phase (= 1 ^st iteration), by default, the selected reconstruction planes may be, for example, 3 or four planes that are tangent to a cylinder which surround the average trajectory of the system, so as to ensure a reconstruction in the different directions visible by the panoramic system. For example, for a horizontal trajectory located at 100 m from the ground, one could choose the following reconstruction planes: the plane of the ground, a plane perpendicular to the ground and tangent on one side to the cylinder surrounding the trajectory, a plane perpendicular to the ground and tangent on the other side of the cylinder, a plane parallel to the ground located at a height greater than 100m. Once an initial 3D reconstruction begins to be constructed, these previously defined reconstruction planes can be updated to approach or merge with the planar surfaces of the current reconstruction, extractable automatically or experimentally by an operator. . When a single reconstruction plane is not enough to give a sufficient 3D representation of the scene, several parallel or perpendicular planes are used to restore the uniqueness and completeness of the 3D representation. This is the case for example when the scene comprises a concave object, or in the case where a single reconstruction plane provides different measurements of 3D dimensions depending on the angle at which the measurement is made, and is therefore unable to provide a single measure, as shown in Figure 4. This figure illustrates the Z reconstruction ambiguity of the point (X, Y): the acquisitions at positions 1 and 2 of the trajectory reconstruct z1 on the reconstruction plane P1, but the acquisitions at the 2 and 3 of the trajectory reconstruct z2 on the same plane of projection P1. We then choose a new reconstruction plan P2 to remove the ambiguity because we will have z1 for P1 and z2 for P2. A plane P3 is also chosen to find the lateral limits of the concave object.

As the panning system moves, when new shots are unveiled or disappear in the scene, it may be necessary to renew the chosen reconstruction plans as well.

Step b): A concept of generalized rectification is introduced to be able to rectify two successive 2D panoramic images (forming a pair of images) in any direction. This rectification consists of calculating at least one projection plane best adapted to the rectification and applying the transformation transforming any sector of each of the two 2D panoramic images on each plane.

It is therefore necessary to determine for each pair of images chosen, the rectification plane corresponding to each reconstruction plane and to directly project a sector of the first image of the couple to obtain a 2D rectified image, and project in the same plane of rectification and directly a sector of the other image to obtain another image rectified 2D.

Each projection plane used for the rectification, called the rectification plane, can be chosen freely by the operator according to the trajectory of the panoramic system, from an infinite choice of positions and orientations all parallel to the trajectory of the panoramic system; the plane or each of them is independent of the evolution of the line of sight of the panoramic system (which can rotate on itself during its movement along its trajectory), contrary to the classic stereoscopy where the plane of chosen correction depends on the evolution of the line of sight and the choice of grinding plans is very limited.

An example of rectification planes referenced R1, R2, R3 is shown in FIG. 7; they are parallel to the trajectory. Is also indicated direction of the LdV (LdV1, LdV2) of the panoramic two-point trajectory sensor, which illustrates that the choice of these plans is independent of the LdV.

Examples of rectification and reconstruction plans are shown in FIG. 8 which is a top view of a scene comprising 3D objects. In the path are indicated position pairs (1, 2, 3) of the panoramic sensor corresponding to 3 pairs of panoramic images acquired during this step b); each pair of positions is associated with two grinding planes (R1 1, R21 for the pair 1, R12, R22 for the pair 2 and R13, R23 for the pair 3). Three reconstruction planes P1, P2, P3 were chosen in the example of this figure.

In order to optimize the 3D reconstruction, the rectification plane which is linked to the reconstruction planes chosen in a), is chosen so as to be the closest in the geometric sense of the reconstruction plane determined in step a). It is also chosen so as to guarantee a minimum of pixellic resolution and must not be too far in the angular direction of the associated reconstruction plane (typical angular difference of less than 30 °).

The transformation (= projection) that passes from the image to the rectification plane is direct, that is to say does not require to go through an intermediate step of recovery in a focal plane as in conventional stereovision; it is carried out without passing through one or more intermediate planes depending on the line of sight of the panoramic system. This makes it possible to obtain a corrected image which is:

- independent of the rotation movement of the sensor and

- without holes contrary to what can be found in conventional grinding,

- more precise because obtained by a direct calculation floating without intermediate quantized image.

The mathematical steps of this correction for a panoramic image obtained at time t, are in the case of a fish-eye type sensor, the following:

\ X Y. 7 I

Choice of a grinding plane P ,, a reference ^! ' ^! '''associated with this plane P ,, and a sector of the panoramic image (up to the complete panoramic image if the field thereof is included in the selected area on the rectification plane) to be projected on this rectification plane, this sector advantageously making it possible to cover the rectification plane as much as possible. If the sector of the projected image does not cover the entire panoramic image, the sectors remaining in the image are projected in other grinding planes, as in the example of Figure 6 where the sector 1 is projected on a horizontal rectification plan and does not cover the entire image to maintain a certain resolution; other vertical rectification planes are needed to project the other sectors of the panoramic image.

Calculation of the transformation which transforms a point (x, y) of the panoramic image into a point (X ,, Y,) of the plane P,; for this purpose, the correspondence between the angular direction (θ, φ) of a point in the scene and the coordinate (x, y) of the corresponding point in the image which depends on the chosen panoramic system is used. In the case of a rectilinear panoramic system, we can write this relation simply:

If R is the radius of the position of the point (x, y) with respect to the optical center, we have: tg9 = (y - yc) / (x - xc) where (xc, yc) are the coordinates of the optical center

φ = k.R with k = rectilinear factor of the sensor

Then simply write the equation of the plane P, as a function of the (θ, φ) found.

For a plane P, whose normal is oriented along (θ ,, φ,), with for particular case the focal plane for θι = φι = 0, we prove that we have the following relation for the particular case of the projection centered, f being the focal point of the panoramic system:

<pcos φ _ί cos (<9 - #, ·) - cos sin φ _ί

sin ç sin φ _ί cos (<9 - 6 _t ) + cos <pcos φ _ί

sin ç? sin (6> - 6>_; )

f

sin ç sin φ _ί cos (<9 - #, ·) + cos çcos, ç _t

This transformation is an example of transformation in the case of a rectilinear panoramic optics (fisheye type); it does not understand the distortion parameters of the panoramic system that can be calculated and offset elsewhere. The transformation can easily be generalized and adapted to any panoramic system having its own optical formula.

It follows that for any point (x, y) of the sector of the panoramic image, one can find its corresponding rectified point in the rectification plane chosen and thus build the rectified image in this plane.

The various rectification plans chosen during the iterations, and the preceding relation make it possible to define a sector correction on the different rectification plans. A sector of the panoramic image corresponds to an equivalent portion projected on a rectification plane. The sectoral decomposition of the panoramic image depends on the rectification plans chosen and the influence of the projection on these plans.

Examples of rectified images are shown in FIG. 6. The first results from the projection of the sector 1 of the image of FIG. 5 on a vertical rectification plane, the second results from the projection of the sector 2 of the image of the Figure 5 on another vertical rectification plane, the third results from the projection of the sector 3 on a vertical rectification plane different from the first two, the fourth results from the projection of the sector 5 on a horizontal rectification plane.

This projection is reiterated in the same rectification plane P, for a sector of another 2D panoramic image obtained at time t + Δΐ to obtain another corrected image, Δΐ being predetermined experimentally or determined so that the displacement From the system between t and t + Δΐ realizes a stereo base large enough to be compatible with the desired accuracy for 3D reconstruction. In the case, for example, of an overflight at an average distance H from the scene, and assuming, for example, that minimum disparities of 1/8 pixel (current value) can be measured by the sensor 14, the displacement De to obtain the desired reconstruction accuracy d _H is: Dc = (resol / 8) ^* H ² / d _H , where resol is the resolution of the sensor 14 (for example 3mrd for a Mpixel I sensor equipped with a fisheye).

In the example cited, and assuming that the desired reconstruction accuracy of _H is 20 cm for H = 50m, De must at least be equal to 5m, which corresponds to a minimum angle difference of 6 ° between 2 acquisitions by the panoramic system.

For the same precision d _H and H = 100m, De must at least be equal to 19m, which corresponds to an angular difference of 1 1 ° minimum between 2 acquisitions by the panoramic system.

The use of a panoramic system makes it possible to increase the reconstruction accuracy by increasing the distance De and the angular separation between two acquisitions, beyond what a weak or medium-field sensor can do for the same spatial coverage of reconstruction. 3D. The 3D stereoscopic base for 3D reconstruction may be larger than that of a conventional stereoscopic process due to the use of a panoramic field (and the longer presence of objects in this field), and this allows the This process has a greater ultimate reconstruction accuracy, which is also increased by the fusion of the measurements offered by the process.

Taking the previous example of an average distance flight of 100 m from the scene (ground reconstruction over a field of at least 120 ° corresponding to a restored band of at least 350 m wide without counting the reconstruction on the sides), the theoretical reconstruction accuracy d _H becomes 10 cm for Dc = 38m and an angular difference of 21 °, and 2 cm for Dc = 200m and an angular separation of 60 °; it is preferable to take into account the uncertainties of measurements on the relative location between the points of view to obtain d real _H.

If we take the context of a visual inspection made by a panoramic system with a fisheye camera I Mpixel, at a distance H = 20cm from the scene, and assuming a displacement of 10 cm between two acquisitions, details of 15μιτι height or depth can be restored (d _H = 15 μιτι).

In order to average the different 3D reconstructions obtained during the iterations, and thus benefit from a significant reduction of the errors and of the restitution noise, the actual acquisition of the panoramic system can be faster while keeping the displacement between the pairs of imagery rectified 2D used to reconstruct the 3D of the scene. The method then consists in taking a first pair of 2D panoramic images from a displacement De, doing an intermediate 3D reconstruction with this pair, then taking another pair of 2D images always from a displacement of to the next acquisition to remake an intermediate 3D reconstruction as long as the points of the scene concerned by these different pairs of images remain in the field of the panoramic system.

Step c): The stereoscopic pair of rectified images in the rectification plane P is used to define an intermediate 3D reconstruction in a reference relative to this P ,.

The intermediate 3D reconstruction in a 3D reference linked to the P ,, said intermediate 3D reference is obtained by matching point-to-point the two rectified images in P, and with the knowledge of the movement of the panoramic system. This mapping is a dense process that matches, as far as possible, each of the points in a 2D image of the stereoscopic pair with a point in the other image. It can be carried out by a more or less hierarchical local correlation process and can be used with mappings made at t-Δΐ or t-ΝΔΐ, where N is an integer> 1; the wide-field nature of the panoramic system and the very possibility of seeing the same scene from a different angle, which is not possible with a weak field system traditionally used in stereoscopy, makes it possible here to remove certain occultations or ambiguities.

Step d): transform this intermediate 3D reconstruction into a fixed 3D reference (= absolute) including the reconstruction plane determined in step a), called 3D reconstruction mark. A transformed intermediate 3D reconstruction is thus obtained.

Step e): repeating at least once steps b) to d) from a new pair of panoramic images (it may be a new pair of images formed from previous images, or this new pair results from a new acquisition coupled with one of the previously acquired images) and from at least one other rectification plane P ', to obtain at minus another transformed 3D intermediate reconstruction; we keep the same 3D reconstruction mark as in step d). These iterations can be successive in that steps b) to d) are successively linked in this order; these iterations can also be performed in parallel (several steps b) are performed in parallel with several rectification planes P, determined in parallel, etc.).

Preferably, these steps b) to d) are repeated as long as at least one point of the reconstructed scene remains in the field of view of the panoramic system.

Step f): The transformed intermediate 3D reconstructions are temporally fused by a specific fusion process that exploits the spatial and temporal redundancies of the intermediate reconstructions. This is obtained by temporally merging at least two transformed intermediate 3D reconstructions obtained in the 3D reconstruction mark, to obtain a corresponding 3D reconstruction.

This 3D reconstruction method makes it possible to find the most appropriate dense 3D mesh for representing the scene, such that at each point of this mesh are associated the coordinates of the corresponding point in a frame Χ, Υ, Ζ linked to the scene.

Step g) possible: repeat steps b) to f) for each reconstruction plane chosen in a), with the same panoramic images but with different sectors, so as to obtain as many 3D reconstructions as selected reconstruction plans. These 3D reconstructions or the intermediate 3D reconstructions obtained during these iterations are advantageously spatially merged to update the final 3D reconstruction (s), and thus increase the accuracy and robustness of these reconstructions. The spatial fusion of the 3D reconstructions constructed according to different planes takes into account the reconstruction precision of the different elements of each reconstruction which is not the same according to the different planes and which can be predicted mathematically. This spatial fusion is obtained by exploiting several rectification planes corresponding to the different sectors of each image used. The set of steps a) to g) are also preferably repeated at least once with new pairs of panoramic images, for example with temporally offset intermediate images of the previous ones, or with other sectors of the panoramic images already considered. This allows for a continuous process of updating the final 3D reconstructions. These new pairs of panoramic images can come from each panoramic image acquisition but not necessarily.

Again, these iterations can be conducted successively or in parallel.

The exploitation of the redundancies and the quality of the 2D rectified images (quality defined for example by the angular difference existing between the rectification plan and the reconstruction plan, or by a coefficient of confidence of the mapping which led to each intermediate 3D reconstruction) allows the process to produce a confidence map reflecting the quality of the final reconstruction. This confidence map is constructed pixel by pixel for each 3D reconstruction, considering the number of times each pixel has been built and if the conditions of this construction were good, these being for example defined according to a threshold of match quality determined experimentally or mathematically. Also considered are the cases where several 3D dimensions are obtained for the same pixel as a function of the angle of observation, in which case additional rectification and reconstruction plans are created to remove the ambiguity, for example for concave objects which require more than one reconstruction plan to be correctly reconstructed, as in the example of Figure 4. A second example of 3D reconstruction to serve as a preliminary step in 3D mosaicing is the particular case of the reconstruction method described above where only a reconstruction plan is maintained along the trajectory.

In this case, the rectification plans of the various pairs of images chosen to carry out the reconstruction are those which are the most Geometrically close to the construction plan and only one sector is chosen to perform the correction on these different rectification plans.

The final 3D reconstruction is performed only on a single reconstruction plane (the one initially chosen). It follows that only the faces and surfaces that are not too far angularly (<30 °) from the construction plan will be restored, the others may be either poorly restored or may have gaps.

For a given panoramic image, the sectors that will be extracted for mosaicing will be those corresponding to the rectification plans used with this panoramic image in the reconstruction. They will be well adapted to the mosaic of 3D reconstruction, but will be less suitable and sometimes poorly adapted to mosaicing badly reconstructed parts, for example perpendicular or near parts to be perpendicular to the reconstruction plane.

A third example of 3D reconstruction that can be used is performed by repeated stereoscopy: the rectification plane depends entirely on the movement of the line of sight of the panoramic sensor and is made in a plane that is as close as possible to the focal planes of the two images. used for successive reconstructions. The reconstruction plan is often confused or close to the first rectification plan used for reconstruction (most often we take the focal plane of the first image or a parallel plan as a reference).

It follows that successive reconstructions can only be properly constructed if the focal planes do not move too far from the initial reconstruction plane.

Similarly, the projection used for the rectification is not direct but passes through a first projection along the focal plane, followed by a second transformation linked to that which connects the focal plane of each image to the plane of rectification. This second transformation may present a significant rotation because of the movements of the line of sight of the sensor, thus inducing a ground image filling very little the plane of rectification and may even induce the impossibility of matching with the second ground image, or worst matching errors. The second problem is the significant distance possible in the angular direction of the rectification plane with respect to the reconstruction plane inducing in the final reconstruction, if it is made in a frame linked to the scene, significant noise reconstruction. It follows path portions without 3D reconstruction or with 3D reconstruction erroneous. At best, the method decreases compared to the two preceding methods the number of images that can be used to correctly reconstruct a point and therefore increases the reconstruction noise.

Most often therefore, the 3D reconstruction presents in this case the following characteristics: correctly restored but strongly noisy for the geometrically closest surfaces of the reconstruction plane, which may present significant errors in places or gaps, and restored in a very partial way or poorly rendered for faces or surfaces almost perpendicular to the reconstruction plane.

During mosaicing, the association of the texture from the panoramic sensor with the 3D model is very difficult in this case and can only be done when the line of sight of the sensor is close to perpendicular to the reconstruction plane. In this case, it only concerns the surfaces of the model that are not too geometrically far from the reconstruction plane. It is no longer possible to systematically use rectified images because they are mostly too small in extent, and one has to be content with using the sensor's point of view to mosaic, which makes the mosaicization process less specific.

A fourth example of a 3D reconstruction is one in which the 3D reconstruction has been carried out according to another trajectory and from other points of view than those used for the 3D mosaicization envisaged. This is referred to as external 3D reconstruction as opposed to previous examples of internal 3D reconstruction. In this fourth example, the sensor used for the 3D reconstruction is not necessarily panoramic. In addition, the 3D reconstruction and 3D mosaicking paths may be spatially different and / or temporally different.

The rectification plans used by this 3D reconstruction can no longer be used to match the texture of the panoramic image with the 3D reconstruction, and it becomes necessary to use in this case only the point of view to delimit the image. texture to project. The projection on the surfaces, assuming that the choice of surfaces is adapted to the texture to be projected during the mosaicing, will be tainted by the relative point of view estimation error between the sensor and the 3D model.

We now consider mosaicing 2D images of the scene acquired by a panoramic sensor moving along a mosaic path, the composition of these 2D images forming a global image called mosaic. This mosaic generally has several 2D textured planes present in the 3D scene or that approximate it, but can also be on a 3D surface.

The exploitation of the 3D reconstruction of the progressively created scene makes it possible to project each 2D image coming from the sectoral decomposition of the 2D panoramic image on different projection planes (or surfaces) also called mosaicization plans. These projection surfaces are the surfaces on which the mosaic is built; they can be freely chosen by the operator or can be determined automatically from the 3D reconstruction. As indicated above, some of these surfaces may be left (curves) or even be a 3D surface of which modeling is known.

In the case of a panoramic system viewing a highly 3D scene having different faces, several mosaicization planes (or surfaces) may (and have interest to) be chosen. By highly 3D scene, we mean a scene containing many 3D elements producing significant disparities between two successive acquisitions, as is the case for example of a drone flying over a low-flying urban environment. The mosaicing method exploits the fact that the surfaces or projection planes have different orientations for projecting the textures of the images on each of the projection surfaces or planes. Remember that the texture is a set of intensities of pixels on an image region.

The exploitation of the 3D reconstruction makes it possible to keep only the visible parts of the projected images. This makes it possible to avoid projecting portions of images that belong to other portions of the scene and that would be hidden in the projection onto a mosaicization plane. The multi-plane (or multi-surface) mosaicing process is reiterated preferably with each new 2D image acquisition performed by the panoramic system, and the new mosaic is merged with the old one (obtained at t-1). to update this one.

The result of these different projections and the continuous fusion of the mosaics is a conformal image (that is to say without geometric deformations) and very extensive on each projection plane. This is a direct result of the fact that the 3D mosaicing method explained below in detail and described in connection with FIG. 3 simultaneously calculates the 3D reconstruction of the scene and the projections of the textures thereon, that the method eliminates the hidden parts or badly resolved parts in the projection and that this process is reiterated in all directions and along the entire trajectory.

The 3D reconstruction of the scene and the projections of the textures on it are calculated at each acquisition of the so-called initial 2D image, an acquisition being separated from the previous one by a time interval Δί defined above.

According to one alternative, the 3D reconstruction of the scene and the projections of the textures on it are calculated at each image acquisition of the panoramic system and at a high frequency from previous images previously stored. The reconstruction and mosaicization trajectories are then the same. More precisely: the intermediate images comprised between two successive images of Δί successive for the 3D reconstruction, are stored so that they can also be used for 3D reconstruction in the manner of a FIFO acronym for the English expression "First In First Out "(each new image acquired is compared with the first image stored to establish a new 3D reconstruction instance, this first image is then deleted from the list and the last added to the updated list). On the other hand the intermediate images can also be used to facilitate the correspondence between the first and last image, or to fill in "holes" in the 3D model.

A mosaic is then obtained at the end of the following steps A) to E) described in connection with Figure 3, for each new 2D image acquired by the panoramic system. According to a first embodiment, 3D reconstruction and mosaicing are carried out successively after each acquisition; this supposes that a new 3D reconstruction of reference has just been made following (one or) 3D reconstructions already made.

According to a second embodiment, 3D reconstruction and mosaicing are performed in parallel after each acquisition; this assumes that the mosaicization is performed while a new 3D reconstruction is still in progress, in which case the reference 3D reconstruction is the one performed to one of the previous acquisitions of 2D images, or a 3D reconstruction performed previously.

These different steps will be described in more detail.

A) Acquire successive images by means of a moving panoramic sensor following a predefined mosaic path.

B) Select one or more projection-related (or 3D) projection planes on which the mosaic will be built.

This step consists of choosing the 3D projection planes or surfaces on which the mosaic is built. These 3D projection planes or surfaces can be freely chosen by the operator at a given moment of the mosaicing or computed automatically from the current 3D reconstruction (or reference) of the scene according to predetermined criteria (for example, planes parallel to the reconstituted soil surface or main planes extracted from the 3D reconstruction). Non-planar 3D projection surfaces may also be used if the scene is suitable and if the operator sees an interest therein; this allows for example to represent objects of the scene or a backdrop that have particular geometric shapes, but this does not detract from the conformity that can be obtained by multiple projections that would be exclusively flat.

The choice of projection surfaces can be made for each panoramic image in all possible directions of the panoramic image, these surfaces being chosen preferentially from those of the reconstructed 3D model and as close as possible to the geometric sense of the available correction plans containing rectified images from the current panoramic image. As the choice of rectification plans is directly related to the reconstruction process, it is advantageous to choose the multi-plane reconstruction method (= first reconstruction example) which favors the multiplicity of rectification planes in different directions. The possibilities in the direction of the projection surfaces that can be used by mosaicing (and therefore the choice of possible mosaicization directions in the panoramic image) are directly related to the number of reconstruction planes used by the reconstruction method.

C) For each current acquired panoramic image, recover textures used for mosaicing. There are two cases depending on whether these textures come directly from the current image or not.

We first consider the case where these textures do not come directly from the current image. When rectified images have been calculated during one or more previous 3D reconstructions and are included in the field of the current image, and at least one of them is sufficiently close to the geometric sense of the surface of the current image. projection according to a predefined threshold criterion, one chooses the closest (in the geometric sense) of the projection plane and one calculates the parameters of the direct projection (without intermediate stage) on the plane of projection.

We now consider the case where these textures come directly from the current image: it is in the case where no corrected image could be found sufficiently close to the projection plane with respect to a predetermined threshold by the operator for example . The current 2D panoramic image is used and the parameters of the direct projection of this current 2D panoramic image on the projection plane and the sectors of this current image which will be used during the direct projection of step E are calculated.

In both cases, the projection is not performed immediately but the projection parameters are stored in memory for use by step E).

D) Select with the help of the (or) reconstruction (s) 3D, among the previous textures the parts of these which are visible on each projection surface and preferably from the latter those which are exploitable (that is to say having a sufficient resolution) for the projection of step E, from:

the corrected 2D image from the current image if the first case of step C applies,

or sectors of the current panoramic image that will be used in the direct projection, if the second case of step C applies.

The newly calculated 3D reconstruction automatically calculates the hidden or weakly resolved parts (and conversely the exploitable visible parts) in the projection plane that would result from masks present in the scene. This is like selecting the textures to keep in the mosaic dataset. This calculation is accurate because the 3D reconstruction was first performed in the reference linked to the panoramic system.

One of the peculiarities of 3D mosaicing according to the invention is to take advantage of the calculation of the hidden parts to eliminate the masks generated by the scene on the different projection planes (surfaces) and to consider only the visible parts on these planes (surfaces). This makes it possible to temporally mosaic only portions of scenes that are always visible and thus avoid deformations due to the projection of parts of the scene not belonging to the projection plane (default present in a conventional mosaic which projects from all the image on the projection plane without being able to take into account the masking by the elements of the scene evolving as the sensor moves, and Figure 9 shows the case of a mosaic with taking into account that the parts visible and that resulting from the projection of the textures without delimitation of the always visible parts: it results in the second case of a progressive deformation of the textures due to the projection of textures of which a part should have been masked).

It should be noted that the method of limiting the textures projected to the visible parts in the model depends on the accuracy of the 3D model and the precision of rendering point of view between the current image and this 3D model. This favors a method for obtaining the 3D model by a multi-plane method applied in parallel to the mosaicing process, because the multi-plane method makes it possible to correctly reconstruct the 3D model. in all directions that does not allow a single plane reconstruction process. The imperfections of restitution of surfaces perpendicular or close to being perpendicular to the reconstruction plane in the case of a single-plane process, prevent correct calculation of the projection of textures on these surfaces and to delimit correctly the masking of these imperfections on d other surfaces.

E) Project the textures selected in the previous step on the planes or more generally on the projection 3D surfaces and merge the textures in each projection plane or 3D surface so as to obtain a conformal mosaic in several planes.

As well as a direct transformation between rectification plane and reconstruction plane that maximizes the surface of the rectified image is favored, a direct projection of the rectified image used or the selected sector in the panoramic panoramic image is also preferred. current, in order to maximize the surface of the projected textures on the projection plane (s).

The textures selected in the previous step are projected onto the projection planes (or surfaces), and temporally merged with the current mosaic to form a new mosaic.

It is important to note that projecting onto surfaces or planes resulting from a 3D reconstruction made from the same basic images as those used to project the textures, allows a very high precision in the projection (and in the transformations between image mark and mark of the reconstructed scene). This is also what ensures the conformity of the product mosaic. The main element of compliance results from the fact that mosaicing exploits a 3D reconstruction performed in the same frame as the mosaic and uses only the portions of unmasked images in its mosaic process. In the case where one would use an external 3D reconstruction that would not come from the basic images used for the projection, there would necessarily be uncertainties in the relative position of the sensor to the scene and in the recovery of the projection. F) Present the mosaic to the operator according to different planes or more generally according to different 3D surfaces of presentation, by projection of the 3D reconstruction (s) textured (s) on these plans of presentation. These presentation plans are freely chosen by the operator and serve only to present the results of the mosaic according to different perspectives chosen by it. The mosaic can be presented to the presentation plan operator by presentation plan, or according to the plans representing the unfolding of the curved surfaces on which the textures have been projected (in the case of a projection on a cylinder for example). One can of course also present the textured 3D result directly in the form of virtual 3D using a suitable software. The projection result provides a still consistent image, which is not necessarily the case, as has been explained, with a conventional mosaic method.

FIG. 9a shows the mosaicization results obtained without applying the method according to the invention, in particular without applying a selection criterion on the textures: deformations can be seen. FIG. 9b shows the mosaicization results obtained by using, according to the invention, a 3D reconstruction for projecting the unmasked textures onto the projection surface bound to the ground: there is no deformation.

This method of 3D reconstruction and simultaneous omnidirectional mosaicking is not limited to an optical panoramic system. The textures measured over a large directional field can be very well exploited by means other than optics, for example by an active means of lidar or sonar type; the process could then also exploit the distances given by the instruments.

Among industrial applications, we can consider: the 3D and textural real-time rendering of a scene overflown by a drone or an aircraft (application to the production of 3D maps, orthophotographs, application to surveillance, ...) ,

assistance to land or airborne navigation,

industrial, medical or other visual inspection, 33

Claims

Method for mosaicing a scene in a 3D mosaic, characterized in that at least one 3D reconstruction having been obtained during the following steps of:

acquisition of successive images by a sensor moving along a 3D reconstruction trajectory,

- 3D reconstruction in an intermediate plane from these matched images,

and in that it further comprises the following steps:

D) Among these textures, select the visible parts in each projection surface using the 3D reconstruction,

E) Project these textures on each projection surface and merge the textures in each projection surface to obtain a matching mosaic on each projection surface.

Method for mosaicing a scene according to the preceding claim, characterized in that when corrected images resulting from the 3D reconstruction and included in the field of the current image can be determined, and that these belong to rectification planes closest to the geometric sense of the projection surfaces, then the textures recovered are those of the rectified image (s).

A method of mosaicing a scene according to claim 1, characterized in that step C) includes a step of determining sectors of the current panoramic image that correspond to the selected projection surfaces.

Method of mosaicing a scene according to the preceding claims, characterized in that steps B) to E) are repeated with each new panoramic image acquisition.

Method for mosaicing a scene according to one of the preceding claims, characterized in that the choice of projection surfaces is made in all directions of the panoramic image, compatible reconstruction planes that generated the 3D reconstruction.

Method of mosaicing a scene according to one of the preceding claims, characterized in that each texture having a resolution, it comprises a step for determining the resolution of the textures and in that the fusion of the textures of step E) is carried out according to these resolutions.

Method for mosaicing a scene according to one of the preceding claims, characterized in that the projections of the textures are direct.

Method for mosaicing a scene according to one of the preceding claims, characterized in that the 3D reconstruction and mosaicization paths are the same trajectory.

Method for mosaicing a scene according to one of the preceding claims, characterized in that the 3D reconstruction or reconstructions have been obtained by using several reconstruction planes.

10. The method of mosaicing a scene according to the preceding claim, characterized in that rectification planes are chosen in different directions. 1 1. Method for mosaicing a scene according to one of the preceding claims, characterized in that the 3D reconstruction (s) have been obtained by direct transformation.

12. Equipment for 3D mosaicing a scene that includes a panoramic system (1) capable of forming images of the scene and moving along a mosaic path, provided with image-to-image relative location means (12). , 13) and, connected to this panoramic system, a computer (2) comprising means (21) for implementing the mosaicing method according to one of the preceding claims.