Method and device for determining depth using a self-supervised vision system.

Abstract

A method or device for determining depth using a vision system mounted in a vehicle (10), the vision system comprising an array of at least two cameras (11, 12), each configured to acquire an image of a scene from a different viewpoint, the optical axes representing an orientation of the field of view of each camera being oriented non-parallel. The method includes the steps of predicting depths using a stereoscopic vision system and a monoscopic vision system, generating images from the different depths by applying a similar function, training each vision system based on errors determined by comparing the generated images with acquired images, and adjusting the monoscopic vision system under the supervision of the stereoscopic vision system, the stereoscopic vision system providing training data. Figure for the abstract: Figure 1

Description

Title of the invention: Method and device for determining depth by a self-supervised vision system. technical field

[0001] The present invention relates to methods and devices for determining depth using a vision system installed in a vehicle, for example, in a motor vehicle. The present invention also relates to a method and device for measuring such depth. The present invention further relates to a method and device for controlling one or more ADAS systems installed in a vehicle based on the determined depth. Technological background

[0002] Many modern vehicles are equipped with Advanced Driver-Assistance Systems (ADAS). Such ADAS systems are passive and active safety systems designed to eliminate human error in driving all types of vehicles. ADAS systems use advanced technologies to assist the driver while driving and thus improve performance. ADAS systems use a combination of sensor technologies to perceive the environment around a vehicle and then provide information to the driver or act on certain vehicle systems.

[0003] There are several levels of ADAS, such as reversing cameras and blind spot sensors, lane departure warning systems, adaptive cruise control or automatic parking systems.

[0004] Vehicle-mounted ADAS systems are powered by data obtained from one or more on-board sensors such as, for example, cameras. These cameras make it possible, in particular, to detect and locate other road users or any obstacles present around a vehicle in order, for example: - to adapt the vehicle's lighting according to the presence of other road users; - to automatically regulate the vehicle's speed; - to act on the braking system in case of risk of impact with an object.

[0005] The quality of the data emitted by a vision system therefore determines the proper functioning of the driving assistance devices using this data. Summary of the present invention

[0006] One object of the present invention is to solve at least one of the problems of the technological background described above.

[0007] Another object of the present invention is to improve the quality of the data obtained of these cameras.

[0008] Another object of the present invention is to improve road safety, in particular by improving the reliability of AD AS systems powered by data obtained from at least one camera.

[0009] According to a first aspect, the present invention relates to a method of determining depth by a vision system embedded in a vehicle, the vision system comprising a set of cameras of at least two cameras arranged so as to each acquire an image of a three-dimensional scene from a different point of view, the method being characterized in that it comprises the following steps: - reception of first and second data respectively representative of a first and second image acquired by respectively a first and second camera of the set of cameras at the same first time instant of acquisition; - prediction of first depths associated with a first set of pixels of the first image and of second depths associated with a second set of pixels of the second image as a function of optical flux values ​​and first disparity values ​​induced by a relative position of optical axes of the first and second cameras, determined from the first and second sets of pixels, the first and second depths being predicted by a first convolutional neural network associated with a stereoscopic vision system including the first and second cameras; - generation of a fourth image from the first image and the first depths and of a fifth image from the second image and the second depths, the image generation being obtained by applying an image generation function; - learning the first convolutional neural network by minimizing a first error, the first error being determined by comparing the first and fifth images and by comparing the fourth and second images; - reception of third data representative of a third image acquired by the first camera at a second time instant of acquisition different from the first time instant of acquisition of the first image; - determination of data representative of a movement of the first camera between the first temporal instant of acquisition and the second temporal instant of acquisition as a function of the first and third images; - prediction of third depths associated with a third set of pixels from the first image and of fourth depths associated with a fourth set of pixels from the third image as a function of optical flux values ​​and second disparity values ​​induced by the relative position of optical axes of the first camera at the first and second time instants of acquisition determined from of the third and fourth sets of pixels, the third and fourth depths being predicted by a second convolutional neural network associated with a monoscopic vision system including the first camera; - generation of a sixth image from the first image and the third depths and of a seventh image from the third image and the fourth depths by application of the image generation function; - learning the second convolutional neural network by minimizing a second error, the second error being determined by comparing the sixth and third images and by comparing the seventh and first images; - adjustment of parameters of the second convolutional neural network according to the first depths corresponding to training data.

[0010] Such a method makes it possible to predict first depths for the stereoscopic vision system more reliable, thereby providing the monoscopic vision system with higher-quality training data, known as "annotated data." The prediction of second depths by the monoscopic vision system is then more precise and achieves a metric accuracy comparable to that of the stereoscopic vision system.

[0011] A driving assistance system using first or second depths as input data then has reliable data regarding the distance separating the vision system from objects present in the field of vision of one of the cameras of this vision system.

[0012] According to a variant of the method, the values ​​of the first and second induced disparities are determined by a third convolutional neural network.

[0013] According to another variant of the method, the image generation function is defined by the following function, respectively for the stereoscopic vision system and for the monoscopic vision system: = 7F ( [ W ( p^, D st ( p f ) j ) With : • Ps a pixel of a generated image, • 71 a function to convert from homogeneous coordinates to pixel coordinates by removing one dimension from a vector, • K is an intrinsic matrix of the second camera, respectively of the first camera, • M an extrinsic matrix of a stereoscopic vision system, respectively a matrix representing the movement of the first camera between the first and second time instants, • 0 a projection function in the three-dimensional scene of a pixel as a function of its depth, • K' an intrinsic matrix of the first camera, and • D^p ) is a depth of the pixel Pt determined by the first convolutional neural network, respectively by the second convolutional neural network.

[0014] According to yet another variant of the method, the first error is determined by a first loss function defined by: 4 - L p avg(L5X / A LM) Avec : • £0a first error, • avS an average of arguments, • Ls^p) an error for a pixel P of the first image compared to the fifth image, • Lx / p) an error for a pixel P of the second image compared to the fourth image.

[0015] According to yet another variant of the process, the first error is determined by a first loss function defined by: L, = LM) With: • £0a first error, • min a minimum value of an argument, • Ls^p) an error for a pixel P of the first image compared to the fifth image, • LSi(p) an error for a pixel P of the second image compared to the fourth image.

[0016] According to yet another variant, the second error is determined by a second loss function defined by: Lm = L / àVg(Lmi(p\ L„M)With: • y the second error, • av8 an average of arguments, • Lmfô)) an error for a pixel P of the first image compared to the seventh image, • Lm^p) an error for a pixel P of the third image compared to the sixth image.

[0017] According to a further method variant, the second error is determined by a second loss function defined by: L m = L fJ min(L mf (p), L ms (p) Avec : • T the second mistake, • min a minimum value of an argument, • Lm£p} an error for a pixel P of the first image compared to the seventh image, • Lms(p) an error for a pixel P of the third image compared to the sixth image.

[0018] According to another variant of the process, the third image is replaced by a black image, a black image being an image in which a set of colorimetric values ​​associated with the pixels of the image is zero.

[0019] According to a second aspect, the present invention relates to a device for determining depth by means of a vision system embedded in a vehicle, the device comprising a memory associated with at least one processor configured for the implementation of the steps of the process according to the first aspect of the present invention.

[0020] According to a third aspect, the present invention relates to a vehicle, for example of the automobile type, comprising a device as described above according to the second aspect of the present invention.

[0021] According to a fourth aspect, the present invention relates to a computer program which includes instructions adapted for carrying out the steps of the process according to the first aspect of the present invention, in particular when the computer program is executed by at least one processor.

[0022] Such a computer program may use any programming language and be in the form of source code, object code, or an intermediate form between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0023] According to a fifth aspect, the present invention relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for carrying out the steps of the process according to the first aspect of the present invention.

[0024] On the one hand, the recording medium can be any entity or device capable of storing the program. For example, the medium can include a storage means, such as a ROM, a CD-ROM or a microelectronic circuit-type ROM, or a magnetic recording means or a hard disk drive.

[0025] On the other hand, this recording medium can also be a transmissible medium such as an electrical or optical signal, such a signal being able to be transmitted via an electrical or optical cable, by conventional or radio frequency, by self-directing laser beam, or by other means. The computer program according to the present invention can, in particular, be downloaded from an Internet-type network.

[0026] Alternatively, the recording medium may be an integrated circuit in which the computer program is incorporated, the integrated circuit being adapted to execute or to be used in the execution of the process in question. Brief description of the figures

[0027] Other features and advantages of the present invention will become apparent from the description of the particular and non-limiting embodiments of the present invention below, with reference to the attached Figures 1 to 3, in which:

[0028] [Fig-1] schematically illustrates a vision system equipping a vehicle, according to a a particular and non-limiting example of a realization of the present invention;

[0029] [Fig.2] illustrates a flowchart of the different stages of a process for determining a depth by a vision system on board the vehicle of the [Fig.1], according to a particular and non-limiting example of the present invention;

[0030] [Fig. 3] schematically illustrates a device configured for determining depth using a vision system mounted in the vehicle of [Fig. 1], according to a particular and non-limiting embodiment of the present invention. Description of embodiment examples

[0031] A method and device for determining depth by means of a vision system mounted in a vehicle will now be described in what follows with joint reference to Figures 1 to 3. The same elements are identified with the same reference signs throughout the description that follows.

[0032] The terms "first," "second" (or "firsts," "seconds"), etc., are used in this document by arbitrary convention to allow for the identification and distinction of different elements (such as operations, means, etc.) implemented in the embodiments described below. Such elements may be distinct or correspond to a single element, depending on the embodiment.

[0033] According to a particular and non-limiting example of an embodiment of the present invention, a method for determining depth by means of a vision system embedded in a vehicle is for example implemented by a computer of the vehicle's embedded system controlling this vision system.

[0034] The vision system comprises a set of cameras of at least two cameras arranged so as to each acquire an image of a scene from a different point of view.

[0035] For this purpose, the method of determining a depth by a vision system on board a vehicle includes the steps of receiving first, second and third data respectively representative of first, second and third images.

[0036] The first and second images are acquired respectively by the first and second cameras at the same initial acquisition time instant. The first and second cameras then form a stereoscopic vision system.

[0037] The third image is acquired by the first camera at a second acquisition time that differs from the first acquisition time. The first camera then forms a monoscopic vision system, the first camera having different viewpoints of the three-dimensional scene depending on the movement of the vehicle 10.

[0038] According to a first particular embodiment, the second temporal instant of acquisition is prior to the first temporal instant of acquisition.

[0039] According to a second particular embodiment, the second temporal instant of acquisition is subsequent to the first temporal instant of acquisition.

[0040] First and second depths are predicted by a first convolutional neural network associated with the stereoscopic vision system and third and fourth depths are predicted by a second convolutional neural network associated with the monoscopic vision system.

[0041] Fourth and fifth images are generated from, respectively, the first and second images and the first and second depths by applying an image generation function.

[0042] Similarly, sixth and seventh images are generated from, respectively, the first and third images and the third and fourth depths by application of the image generation function.

[0043] Each vision system is then learned by minimizing errors determined by comparing generated images to acquired images and the second depth predictions by the second convolutional neural network associated with the monoscopic vision system are then improved by adjusting the parameters of the second convolutional neural network according to the first depths predicted by the first convolutional neural network, the stereoscopic vision system providing training data, called "annotated data".

[0044] Fig. 1 schematically illustrates a vision system equipping a vehicle, according to a particular and non-limiting embodiment of the present invention.

[0045] Such an environment 1 corresponds, for example, to a road environment consisting of a network of roads accessible to the vehicle 10.

[0046] In this example, vehicle 10 corresponds to a vehicle with an internal combustion engine, with electric motor(s), or a hybrid vehicle with an internal combustion engine and one or more electric motors. Vehicle 10 thus corresponds, for example, to a land vehicle such as a car, a truck, a bus, or a motorcycle. Finally, vehicle 10 corresponds to an autonomous or non-autonomous vehicle, that is to say, a vehicle operating according to a determined level of autonomy or under the total supervision of the driver.

[0047] The vehicle 10 advantageously comprises several on-board cameras 11, 12, each configured to acquire images of a scene in the environment of the vehicle 10. This set of cameras 11, 12 forms the stereoscopic vision system. Two cameras 11 and 12 are illustrated in [Fig. 1]. The present invention is not limited, however, to a stereoscopic vision system comprising two cameras but extends to any vision system comprising two or more cameras, for example, two, three, four, or five cameras.

[0048] The two cameras 11, 12 have known intrinsic parameters. These parameters include, in particular: - the focal length fl of the first camera 11; - the focal length f2 of the second camera 12; - distortions which are due to imperfections in the optical system of each camera; - the direction Cl of the optical axis of the first camera 11; - the C2 direction of the optical axis of the second camera 12; and - the respective resolutions of cameras 11, 12.

[0049] The intrinsic parameters characterize the transformation that associates, for an image point, the camera coordinates to the pixel coordinates, in each camera. These parameters do not change if the camera is moved.

[0050] Distortions, which are due to imperfections in the optical system such as defects in the shape and positioning of camera lenses, will deflect the light beams and thus induce a positioning error for the projected point relative to an ideal model. It is then possible to complete the camera model by introducing the three distortions that generate the most significant effects, namely radial, decentering, and prismatic distortions, induced by defects in lens curvature, parallelism, and coaxiality of the optical axes. In this example, the cameras are assumed to be perfect, that is, distortions are either not taken into account or their correction is addressed during image acquisition.

[0051] These two cameras 11, 12 are arranged so that each acquires an image of a scene from a different viewpoint. The first viewpoint is, for example, located on or in the left-hand rearview mirror of the vehicle 10 or at the top of the windshield of the vehicle 10. The second viewpoint is, for example, located on or in the right-hand rearview mirror of the vehicle 10 or at the top of the windshield of the vehicle 10. If both cameras are located at the top of the windshield of the vehicle, they are then positioned at a certain distance. In this example, the first camera 11 is located at the top of the windshield of the vehicle 10, and the second camera 12 is located in the right-hand rearview mirror of the vehicle 10.

[0052] A first marker is associated with the first camera 11: - the direction of the y-axis is defined by the position of the second camera 12, so as to place the second camera 12 on the y-axis of the first camera 11. The distance B separating the two cameras 11,12 is called the reference base (in English "baseline") and the direction separating the two cameras 11,12 is that of the y-axis; - the direction of the x axis is defined orthogonal to that of the y axis and orthogonal to that of the optical axis Cl of the first camera 11; - The direction of the z-axis is defined as orthogonal to the directions of the x and y axes. The three axes x, y and z thus form an orthonormal coordinate system.

[0053] The extrinsic parameters related to the position of cameras 11, 12 are the following parameters: - 3 translations in the x, y, and z directions: Tx, Ty, and Tz forming the translation vector T; and - 3 rotations around the x, y and z axes: Rx, Ry and Rz, constituting the rotation matrix R.

[0054] Determining the extrinsic parameters constitutes the problem of calibrating a stereoscopic vision system.

[0055] A key constraint of stereoscopic vision systems used in automobiles is, for example, the large distance between the two cameras. Indeed, to cover a measurement range of 200 meters, the reference base must be 60 cm for cameras commonly used in this field.

[0056] The two cameras 11, 12 acquire images of a scene located in front of the vehicle 10, the first camera 11 alone covering a first acquisition field 13, the second camera 12 alone covering a second acquisition field 14 and the two cameras 11, 12 both covering a third acquisition field 15. The first and third acquisition fields 13, 15 thus allow a monoscopic view of the scene by the first camera 11, the second and third acquisition fields 14, 15 allow a monoscopic view of the scene by the second camera 12 and the third acquisition field 15 allows a stereoscopic view of the scene by the stereoscopic vision system composed of the two cameras 11, 12.

[0057] An obstacle 18 is placed in the acquisition field of the cameras, for example in the third acquisition field 15. The presence of the obstacle 18 defines an occlusion field for the stereoscopic vision system composed here of the three fields 16, 17 and 19.

[0058] Among these three fields, field 16 is visible from the second camera 12. The part of the scene present in this field 16 is therefore observable using the monoscopic vision system composed of the second camera 12.

[0059] Field 17 is visible from the first camera 11. The part of the scene present in this field 17 is therefore observable with the help of the monoscopic vision system composed of the first camera 11.

[0060] Finally, field 19 is not visible from any of the cameras. The part of the scene present in this field 19 is therefore not observable.

[0061] According to one embodiment, the directions Cl, C2 of the optical axes representing an orientation of the field of vision of each camera are oriented non-parallel so as to obtain the third acquisition field 15 of the environment 1 as wide as possible.

[0062] According to another embodiment, the directions Cl, C2 of the optical axes representing an orientation of the field of vision of each camera are oriented parallel.

[0063] It is evident that it is possible to use such a stereoscopic vision system to take images of scenes located on the sides or behind the vehicle 10 by equipping it with cameras placed and oriented differently.

[0064] The images acquired by cameras 11, 12 at a given acquisition time t1 are in the form of data representing pixels characterized by: - coordinates in each image; and - data relating to the colors and brightness of objects in the observed scene, for example in the form of RGB (Red Green Blue) or HSL (Hint, Saturation, Luminosity) colorimetric coordinates.

[0065] The images acquired by cameras 11 and 12 represent views of the same scene taken from different viewpoints, the camera positions being distinct. For example, this scene includes: - buildings; - road infrastructure; - other stationary users, for example a parked vehicle; and / or - other mobile users, for example another vehicle, a cyclist or a moving pedestrian.

[0066] These images are sent to a computer of a device equipping the vehicle 10 or stored in a memory of a device accessible to a computer of a device equipping the vehicle 10.

[0067] A process for determining depth by a vision system on board the vehicle 10 is advantageously implemented by the vehicle 10, i.e. by a computer or a combination of computers of the on-board system of the vehicle 10, for example by the computer or computers in charge of the stereoscopic vision system of the vehicle 10.

[0068] In a first operation, the computer receives initial data representing a first image acquired by a first camera 11 of the set of cameras at a first temporal instant of acquisition.

[0069] In a second operation, the computer receives second data representing a second image acquired by a second camera 12 of the set of cameras at the same first time instant of acquisition.

[0070] The two images received correspond to two views of the same three-dimensional scene taking place around the vehicle 10 at the same first given time instant of acquisition.

[0071] In a third operation, first depths associated with a first set of pixels of the first image are predicted as a function of optical flux values ​​determined from the first set of pixels and a second set of pixels of the second image corresponding to the first set of pixels.

[0072] Since the fields of vision of the first 11 and the second 12 camera are concurrent, a part of the pixels of the first image and a part of the pixels of the second image are associated with the same objects present in the three-dimensional scene.

[0073] The third operation is carried out by implementing a method called optical flow calculation. Such a method is notably described in "PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume" by Deqing Sun, Xiaodong Yang, Ming-Yu Liu and Jan Kautz of June 25, 2018 and in "RAFT: Recurrent Ail-Pairs Field Transforms for Optical Flow" by Zachary Teed and Jia Deng of August 25, 2020.

[0074] The optical flow calculation method is implemented by a first convolutional neural network (CNN), a tool commonly used in image processing. An intermediate input of the third operation is an optical flow representing a displacement between each pixel of the first set of pixels and a second pixel corresponding to each first pixel in the second image, the set of second pixels corresponding to the second set of pixels.

[0075] The rotation and translation from the optical axis of the first camera 11 to the second camera 12 create, in particular, a first disparity known as the "induced" disparity. This first induced disparity is, moreover, included in the optical flow determined from the first and second images.

[0076] The rotation is due, for example, to the fact that the optical axes of the first 11 and second 12 cameras are not parallel. The translation is due, for example, to the fact that the first 11 and second 12 cameras are not positioned at the same height on the vehicle 10.

[0077] According to a particular embodiment, in order not to take into account this first induced disparity which is likely to distort the prediction of first and For second depths, a third fully convolutional neural network (FCN), hereafter called an FCN, is implemented to determine the first induced disparity and filter the optical flux value before predicting the depth of a pixel.

[0078] The input data for such a FCN is a one-dimensional HxWxN matrix, where H is the height of an image, W is the width of an image, and N is the number of channels including the non-nuisance elements of the rotation matrix (maximum 9), the non-nuisance elements of the translation vector (maximum 3), and the two optical fluxes (horizontal and vertical). The maximum number of channels is therefore 14. The number of intermediate layers and the number of channels in the matrix at each layer depend on the deviation of the parallel system.

[0079] The structure of this FCN is constructed experimentally in such a way as to minimize the size of the FCN while offering an acceptable result.

[0080] The structure of this FCN is constructed experimentally in such a way as to minimize the size of the FCN while offering an acceptable result.

[0081] Thus, according to this particular embodiment, in a fourth operation, a first induced disparity is determined by the FCN from the first and second images allowing the first induced disparity to be subtracted from the optical flux to obtain a corrected optical flux, the first depths are then predicted as a function of this corrected optical flux, a depth being proportional to a norm of an optical flux from which the first induced disparity is subtracted.

[0082] Similar to the third and fourth operations, in the fifth and sixth operations, second depths associated with the second set of pixels are predicted based on optical flux values ​​determined from the second set of pixels of the second image and the first set of pixels of the first image. The second depths are also predicted by the first convolutional neural network.

[0083] Such operations are permitted because the optical flow can be defined in both directions, i.e. from the first image to the second image and vice versa.

[0084] In a seventh operation, a fourth image is generated from the first image and the first predicted depths by applying an image generation function and in an eighth operation a fifth image is generated from the second image and the second depths by applying the same image generation function.

[0085] According to a particular embodiment, the image generation function is defined by the following function: [Math.l] Ps = tt (K [M¢ (pDst (pf))])

[0086] With: • Ps one pixel of the fourth image, respectively fifth image •77 a function to convert from homogeneous coordinates to pixel coordinates by removing one dimension from a vector, • K an intrinsic matrix of the second camera 12, respectively of the first camera 11, • M is an extrinsic matrix of the stereoscopic vision system, • 0 a projection function in the three-dimensional scene of a pixel as a function of its depth, • K' an intrinsic matrix of the first camera 11, respectively of the second camera 12, • D^pj is a first depth of the pixel Pt of the first image, respectively a second depth of a pixel Pt of the second image.

[0087] The principle of image generation consists of projecting a pixel from an image acquired by the first camera 11 (the first image) onto a point in the three-dimensional scene according to its first depth, and then reprojecting this point onto an image as it would be acquired from the viewpoint of the second camera 12 (the fourth image). Thus, if the first depth value associated with a pixel in the first image is correct, then the point in the scene is well-defined and the second image should have, at the location of this projected point in the scene, a pixel of the same value.

[0088] The same applies in the other direction, that is to say by projecting a pixel from an image acquired by the second camera 12 (the second image) towards a point in the three-dimensional scene according to its second depth, and then reprojecting this point into an image as it would be acquired from the point of view of the first camera 11 (the fifth image).

[0089] In a ninth operation, the first convolutional neural network associated with the stereoscopic vision system comprising the first 11 and second 12 cameras is learned by minimizing a first error, the first error being determined by comparing the first and fifth images and by comparing the fourth and second images.

[0090] According to a first particular embodiment, a first reconstruction error for a pixel of a generated image is further determined by comparing the generated image to an image acquired by a function defined as follows: [Math.2] = EJ(la) ■ \I(p)-I(p)\+a-(1-jSSIM(i(p),I(p)))]

[0091] With: • a first reconstruction error for pixel P, • Kp} a value of pixel P in an acquired image (the first image, respectively the second image), • 2^ a value of pixel P in a generated image (the fifth image, respectively the fourth image), • SSIM is a function that takes into account a local structure, and • has a weighting factor depending in particular on the type of environment.

[0092] This function is based on photometric error. Once an image is generated, it is compared to the true image acquired by the stereoscopic vision system.

[0093] According to a second particular embodiment, a second error in the reconstruction of a generated image is further determined by the following function: [Math.3] (p,). W,o) = H^P,)IV^Pj )

[0094] With: • LmMOth^O. W, o) a second reconstruction error, • Dt^ p ) is the first depth, respectively second depth Dt of a pixel Pt predicted by the stereoscopic vision system; • VF is a parameter matrix; • 0 is the order of a smoothing gradient; • an L1 norm of the second order depth gradients is calculated with VF =1, and ° =2; •1 and 3' are the dimensions of the second image, respectively first image; • is an environment-dependent hyperparameter; and • / ( p ) is the value of the pixel Pt in the fourth image, respectively fifth image.

[0095] This second loss function is generally used to deal with the discontinuity at the edge of objects (in English, "edge aware smoothness").

[0096] The reconstruction error is thus defined from the first and second reconstruction errors previously defined.

[0097] According to a particular embodiment, the first error is determined by one of the following loss functions.

[0098] A first loss function is defined by: [Math.4] L s = E p avg(£ Ar ( / ?), Mp))

[0099] With: • The first mistake, • avS an average of arguments, • Lj^p) a reconstruction error for a pixel P of the first image compared to the fifth image, • Ls£p) a reconstruction error for a pixel P of the second image compared to the fourth image.

[0100] By calculating the average of the two reconstruction errors, the two predictions are treated equally, which improves the prediction reliability when one of the predictions is better in one direction than the other (from the first image to the second image or vice versa), that is, when the reliability of the first and second depths is not equivalent.

[0101] A second loss function is defined by: [Math.5] = lLpmm(Lsl{p\

[0102] With: • the first mistake, • min a selection of the minimum value from among arguments, • Lst(p) is a reconstruction error for a pixel P of the first image compared to the fifth image, • L^(p) a reconstruction error for a pixel P of the second image compared to the fourth image.

[0103] This second loss function is faster and allows the number of iterations in the learning operation to be reduced.

[0104] A third loss function is defined by: [Math.6] Ls=Ep(Lsl(p)+Ljp))

[0105] With: • the first mistake, • Lx£p) a reconstruction error for a pixel P of the first image compared to the fifth image, • Ls£p) a reconstruction error for a pixel P of the second image compared to the fourth image.

[0106] The operations for predicting the first and second depths are thus implemented by the first self-supervised convolutional neural network. Such a convolutional neural network is known to those skilled in the art as PWCNet (from the English "Pyramidal processing, Warping, and the use of a Cost volume Network", in French "Réseau de traitement pyramidal, perturbation et utilisation d'un volume de coûts").

[0107] This learning also makes it possible to find or refine the correct values ​​of the parameters (called "weights" in English) of a chosen convolutional neural network in order to make the first predicted depths more reliable.

[0108] In a tenth operation, third data representing a third image acquired by the first camera 11 at a second time instant of acquisition different from the first time instant of acquisition of the first image are received.

[0109] The third data has, for example, been saved in a memory associated with the computer or in a memory of a device embedded in the vehicle 10 and accessible to the computer implementing the process.

[0110] If the vehicle 10 is in motion, then the third image corresponds to a third view of the scene taken from a third point of view, that of the first camera 11 at its position at the second time instant of acquisition.

[0111] The position of the first camera 11 at the second acquisition time instant is defined by the movement of the vehicle 10 between the first and second acquisition times. This movement of the first camera 11 is therefore linked to the speed and direction of movement of the vehicle 10 during the time interval between the first and second acquisition times.

[0112] The first camera 11, in the two positions defined at the first and second acquisition times, forms a monoscopic vision system. The intrinsic parameters of this system remain the same as those previously defined for the first camera 11 for the stereoscopic vision system. The extrinsic parameters of this monoscopic vision system are as follows: - 3 translations in the x, y and z directions: Tx', Ty' and Tz' constituting the translation vector T'; and - 3 rotations around the x, y and z axes: Rx', Ry' and Rz', constituting the rotation matrix R'.

[0113] In an eleventh operation, data representing the movement of the first camera 11 between the first and second acquisition times are determined based on the first and third images. The extrinsic parameters of the monoscopic vision system are, for example, determined by a computer associated with this same monoscopic vision system. The determination of the extrinsic parameters of the monoscopic vision system is known to those skilled in the art and is presented, for example, in the document Unsupervised Leaming of Depth and Ego-Motion from Video by Tinghui Zhou, Matthew Brown, Noah Snavely and David G. Lowe published August 1, 2017.

[0114] According to a particular embodiment, in a twelfth operation, the third image is replaced by a black image or by the first image when the third image is identical to the first image or when there is no second image, a black image being an image in which a set of colorimetric values ​​associated with the pixels of the image is zero or corresponds to a black object.

[0115] The objective of the black image here is to be able to predict the depth for the first image of a video according to the first embodiment, that is to say when there is no previous image.

[0116] The objective of the identical image is to be able to predict the depth when the vehicle is stationary.

[0117] Such an operation makes it possible to limit the time of the following operations and thus promotes the learning of the monoscopic vision system.

[0118] In a manner similar to that of the third, in a thirteenth operation, third depths associated with a third set of pixels of the first image are predicted as a function of optical flux values ​​and second disparity induced by a relative position of optical axes of the first camera 11 at the first and second time instants of acquisition determined from the third set of pixels and a fourth set of pixels of the third image corresponding to the third set of pixels.

[0119] The thirteenth operation is carried out by implementing the so-called optical flow calculation method, implemented by a second convolutional neural network associated with the monoscopic vision system comprising the first camera 11. An intermediate data of the thirteenth operation is an optical flow representing a displacement between each pixel of the third set of pixels and a fourth pixel corresponding to each third pixel in the second image, the set of fourth pixels corresponding to the fourth set of pixels.

[0120] The rotation and translation to go from the optical axis of the first camera 11 at the first time instant and at the second time instant create in particular a second induced disparity which is, moreover, included in the optical flow determined from the first and third images.

[0121] The third FCN-type convolutional neural network is implemented in a fourteenth operation to determine the second induced disparity and filter the optical flux value before predicting the depth of a pixel.

[0122] It should be noted that the FCN training for the monoscopic vision system is longer than for a stereoscopic vision system. Indeed, the parameters ex The triangular values ​​of the monoscopic vision system change constantly according to the movements of the vehicle 10.

[0123] Thus, according to this particular embodiment, in a fourteenth operation, a second induced disparity is determined by the FCN from the first and third images allowing the second induced disparity to be subtracted from the optical flux to obtain a corrected optical flux, the third depths are then predicted as a function of this corrected optical flux, a depth being proportional to a norm of an optical flux from which the induced disparity is subtracted.

[0124] In a manner similar to the thirteenth and fourteenth operations, the fourth depths associated with the fourth pixel set of the third image are predicted in a fifteenth operation based on optical flux values ​​and disparity induced by the relative position of optical axes of the first camera 11 at the first and second acquisition times determined from the fourth pixel set and the third pixel set of the first image corresponding to the fourth pixel set. The induced disparities are predicted by the FCN in a sixteenth operation.

[0125] The fourth depths are thus also predicted by the second convolutional neural network.

[0126] In a manner similar to that of the seventh and eighth operations, in a seventeenth operation a sixth image is generated from the first image and the third depths by application of the image generation function and, in an eighteenth operation, a seventh image is generated from the third image and the fourth depths.

[0127] According to a particular embodiment, the image generation function is defined by the following function for the monoscopic vision system: [Math.7] Ps = ^(K[M¢iP^ DÀPt] ) ] )

[0128] With: • Ps one pixel of the sixth image, respectively seventh image, • 77 a function to convert from homogeneous coordinates to pixel coordinates by removing one dimension from a vector, • K an intrinsic matrix of the first camera 11 • M is a matrix representing the movement of the first camera 11 between the first and second time instants, • 0 a projection function in the three-dimensional scene of a pixel as a function of its depth, * D^p ) is a third depth of the pixel Pt of the first image, respec- tivement a fourth depth of the pixel Pt of the third image.

[0129] In a nineteenth operation, the second convolutional neural network associated with the monoscopic vision system comprising the first camera 11 is learned by minimizing a second error, the second error being determined by comparing the sixth and third images and by comparing the seventh and first images.

[0130] According to the first particular embodiment, a third reconstruction error is determined in a similar manner to the first reconstruction error by the following function: [Math. 8] L.(p) = L,,[ (ia) • | / (p)-r(p) I +«■ Up) )) ]

[0131] With: • LAp) a third reconstruction error for a pixel P, • I(p) a value of pixel P in an acquired image (the first image, respectively the third image), * / (a value of pixel P in a generated image (the seventh image, respectively the sixth image), • SSIM is a function that takes into account a local structure, and • has a weighting factor depending in particular on the type of environment.

[0132] According to the second particular embodiment, a fourth reconstruction error is determined by the following function: [Math.9] W, o) = W( P,} )

[0133] With: • LP(h(O, W, o) the fourth reconstruction error, • Dt(p) is the third depth Dt of a pixel Pt; respectively the fourth depth Dt of a pixel Pt, - W is a parameter matrix; - 0 is the order of a smoothing gradient; - an L1 norm of the second order depth gradients is calculated with W = 1, et° = 2; - x and P are the dimensions of the first image; - is an environment-dependent hyperparameter; and ■ is 'a va^eur du pixel Pt dans la sixième image, respectively septendre image.

[0134] According to a particular embodiment, the second error is determined by one of the following loss functions.

[0135] A fourth loss function defined by: [Math. 10] Lm = E / ?avg(Lm^p\ Lm^p))

[0136] With: • the second mistake, • avg an average of arguments, • Lmt(p) is a reconstruction error for a pixel P of the first image compared to the seventh image, • Lmskp) a reconstruction error for a pixel P of the third image compared to the sixth image.

[0137] A fifth loss function is defined by: [Math. 11] Lm = Lms(p))

[0138] With: • £' the second error, • avS an average of arguments, • a reconstruction error for a pixel P of the first image compared to the seventh image, • LmXp) a reconstruction error for a pixel P of the third image compared to the sixth image.

[0139] A sixth loss function is defined by: [Math. 12] Lm 52^ ( Lm, ( p ) + )

[0140] With: • T the second mistake, • Lm^p) a reconstruction error for a pixel P of the first image compared to the seventh image, • Lm^p) a reconstruction error for a pixel P of the third image compared to the sixth image.

[0141] In a twentieth operation, the parameters of the second convolutional neural network are adjusted according to the initial depths determined by the first convolutional neural network, the initial depths then corresponding to the training data. Thus, the parameters of the second convolutional neural network are adjusted according to any learning method from training data known to a person skilled in the art.

[0142] The first convolutional neural network associated with the stereoscopic vision system thus predicts reliable first depths, allowing the second convolutional neural network associated with the monoscopic vision system to be provided with good quality training data.

[0143] This training data is used to adjust the parameters of the second convolutional neural network, according to any learning method known to a person skilled in the art.

[0144] If the ADAS uses the first or second depths as input data to determine the distance between a part of the vehicle 10, for example the front bumper, and another road user, the ADAS is then able to determine this distance precisely. For example, if the ADAS is designed to activate the braking system of the vehicle 10 in the event of a risk of collision with another road user, and the distance between the vehicle 10 and that road user decreases sharply, then the ADAS is able to detect this sudden closeness and activate the braking system of the vehicle 10 to avoid a possible accident, even if the other road user is only seen by a single camera.

[0145] Figure 2 illustrates a flowchart of the different steps of a method 2 for determining depth using a vision system mounted in a vehicle, for example the vehicle of Figure 1, according to a particular and non-limiting embodiment of the present invention. Method 2 is, for example, implemented by a device mounted in the vehicle 10 or by the device 4 of Figure 3.

[0146] In a step 21, first representative data of a first image acquired by a first camera 11 of said camera set at a first time instant of acquisition are received.

[0147] In a step 22, second data representing a second image acquired by a second camera 12 of the camera set at the same first time instant of acquisition are received.

[0148] In a step 23, first depths associated with a first set of pixels of the first image and second depths associated with a second set of pixels of the second image are predicted as a function of optical flux values ​​and first disparity induced by a relative position of optical axes of the first 11 and second 12 cameras.

[0149] The optical flux and induced disparity values ​​are determined from the first and second sets of pixels, the first induced disparity arising from a relative position of optical axes of the first 11 and second 12 cameras.

[0150] The first and second depths are then predicted by a first convolutional neural network associated with a stereoscopic vision system comprising the first 11 and second 12 cameras.

[0151] In a step 24, a fourth image is generated from the first image and the first depths by applying an image generation function and a fifth image is generated from the second image and the second depths, the image generation being obtained by applying the same image generation function.

[0152] In a step 25, the first convolutional neural network is learned by minimizing a first error determined by comparing the first and fifth images and by comparing the fourth and second images.

[0153] In a step 31, third data representing a third image acquired by the first camera 11 at a second time instant of acquisition are received.

[0154] The second time instant of acquisition is different from the first time instant of acquisition of the first image.

[0155] In a step 32, data representing a movement of the first camera 11 between the first time instant of acquisition and the second time instant of acquisition are determined as a function of the first and third images.

[0156] In a step 33, third depths associated with a third set of pixels of the first image and fourth depths associated with a fourth set of pixels of the third image are predicted as a function of optical flux values ​​and second induced disparity.

[0157] The optical flux and second induced disparity values ​​are determined from the third and fourth sets of pixels, the second disparity being induced by a relative position of optical axes of the first camera 11 at the first and second time instants of acquisition.

[0158] The third and fourth depths are predicted by a second convolutional neural network associated with a monoscopic vision system including the first camera 11.

[0159] In a step 34, a sixth image is generated from the first image and the third depths by applying the image generation function and a seventh image is generated from the third image and the fourth depths by applying the image generation function.

[0160] In a step 35, the second convolutional neural network is learned by minimizing a second error, the second error being determined by comparing the sixth and third images and by comparing the seventh and first images.

[0161] In a step 36, the parameters of the second convolutional neural network are adjusted according to the first depths, the first depths corresponding to training data.

[0162] According to one variant, the variants and examples of the operations described in relation to [Fig.1] apply to steps of process 2 of [Fig.2].

[0163] Figure 3 schematically illustrates a device 4 configured for determining depth by a vision system mounted in a vehicle 10, according to a particular and non-limiting embodiment of the present invention. The device 4 corresponds, for example, to a device mounted in the first vehicle 10, for example a computer.

[0164] Device 4 is, for example, configured to carry out the operations described opposite [Fig. 1] and / or steps described opposite [Fig. 2]. Examples of such a device 4 include, but are not limited to, embedded electronic equipment such as a vehicle's on-board computer, an electronic control unit such as an ECU (Electronic Control Unit), a smartphone, a tablet, or a laptop computer. The elements of device 4, individually or in combination, may be integrated into a single integrated circuit, into several integrated circuits, and / or into discrete components. Device 4 may be implemented in the form of electronic circuits or software (or computer) modules, or a combination of electronic circuits and software modules.

[0165] The device 4 comprises one (or more) processor(s) 40 configured to execute instructions for carrying out the steps of the process and / or for executing instructions from the software embedded in the device 4. The processor 40 may include integrated memory, an input / output interface, and various circuits known to those skilled in the art. The device 4 further comprises at least one memory 41, for example, volatile and / or non-volatile memory, and / or includes a memory storage device that may include volatile and / or non-volatile memory, such as EEPROM, ROM, PROM, RAM, DRAM, SRAM, flash, magnetic disk, or optical disk.

[0166] The computer code of the embedded software(s), including the instructions to be loaded and executed by the processor, is for example stored on memory 4L

[0167] According to various particular and non-limiting embodiments, the device 4 is coupled in communication with other similar devices or systems (for example other computers) and / or with communication devices, for example a TCU (Telematic Control Unit), for example via a communication bus or through dedicated input / output ports.

[0168] According to a particular and non-limiting embodiment, the device 4 includes a block 42 of interface elements for communicating with external devices. The interface elements of the block 42 include one or more of the following interfaces: - radio frequency RF interface, for example of the Wi-Fi® type (according to IEEE 802.11), for example in the 2.4 or 5 GHz frequency bands, or of the Bluetooth® type (according to IEEE 802.15.1), in the 2.4 GHz frequency band, or of the Sigfox type using UBN (Ultra Narrow Band) radio technology, or LoRa in the 868 MHz frequency band, LTE (Long-Term Evolution), LTE-Advanced; - USB interface (from the English "Universal Serial Bus" or "Universal Serial Bus" in French); HD MI interface (from the English "High Definition Multimedia Interface", or "High Definition Multimedia Interface" in French); - LIN interface (from the English "Local Interconnect Network", or in French "Réseau interconnecté local").

[0169] According to another particular and non-limiting embodiment, the device 4 includes a communication interface 43 which enables communication with other devices (such as other computers in the embedded system) via a communication channel 430. The communication interface 43 corresponds, for example, to a transmitter configured to transmit and receive information and / or data via the communication channel 430. The communication interface 43 corresponds, for example, to a wired network of the CAN (Controller Area Network), CAN FD (Controller Area Network Flexible Data-Rate), FlexRay (standardized by ISO 17458) or Ethernet (standardized by ISO / IEC 802-3) type.

[0170] According to a particular and non-limiting embodiment, the device 4 can provide output signals to one or more external devices, such as a display screen 440, touch or not, one or more loudspeakers 450 and / or other peripherals 460 (projection system) via the output interfaces 44, 45, 46 respectively. According to a variant, one or more of the external devices is integrated into the device 4.

[0171] Of course, the present invention is not limited to the embodiments described above but extends to a method for determining depth and measuring such a quantity by means of a vision system mounted in a vehicle, which would include secondary steps without departing from the scope of the present invention. The same would apply to a device configured for implementing such a method.

[0172] The present invention also relates to a vehicle, for example a motor vehicle or more generally an autonomous land-powered vehicle, comprising the device 4 of the [Fig.3].

Claims

Claims

1. Method for determining a depth by a vision system on board a vehicle (10), the vision system comprising a set of cameras of at least two cameras (11, 12) arranged so as to each acquire an image of a three-dimensional scene from a different point of view, said method being characterized in that it comprises the following steps: - reception (21, 22) of first and second data respectively representative of a first and second image acquired by respectively a first and second camera (11, 12) of said set of cameras at the same first acquisition time instant; - prediction (23) of first depths associated with a first set of pixels of the first image and second depths associated with a second set of pixels of the second image as a function of optical flow values ​​and first disparity values ​​induced by a relative position of optical axes of the first (11) and second (12) cameras, said optical flow values ​​and first induced disparity being determined from said first and second sets of pixels, said first and second depths being predicted by a first convolutional neural network associated with a stereoscopic vision system comprising the first (11) and second (12) cameras; - generation (24) of a fourth image from said first image and the first depths and of a fifth image from said second image and the second depths, the image generation being obtained by applying an image generation function; - training (25) said first convolutional neural network by minimizing a first error, the first error being determined by comparing the first and fifth images and by comparing the fourth and second images; - reception (31) of third data representative of a third image acquired by said first camera (11) at a second acquisition time instant different from said first acquisition time instant; - determination (32) of data representative of a movement of the

2.

3. first camera (11) between the first acquisition time instant and the second acquisition time instant as a function of said first and third images; - prediction (33) of third depths associated with a third set of pixels of the first image and of fourth depths associated with a fourth set of pixels of the third image as a function of optical flow values ​​and of second disparity values ​​induced by a relative position of optical axes of the first camera (11) at the first and second acquisition time instants, said optical flow values ​​and second induced disparity values ​​being determined from said third and fourth sets of pixels, said third and fourth depths being predicted by a second convolutional neural network associated with a monoscopic vision system comprising the first camera (11); - generation (34) of a sixth image from said first image and the third depths and of a seventh image from said third image and the fourth depths by applying said image generation function; - training (35) said second convolutional neural network by minimizing a second error, the second error being determined by comparing the sixth and third images and by comparing the seventh and first images; - adjustment (36) of parameters of said second convolutional neural network as a function of said first depths corresponding to training data. The method of claim 1, wherein said first and second induced disparity values ​​are determined by a third convolutional neural network. Method according to one of claims 1 to 2, for which said image generation function is defined by the following function, respectively for the stereoscopic vision system and for the monoscopic vision system: Ps = Æ (K [(P^ Dst (Pt))) With : • Ps a pixel of a generated image, • 77 a function to go from homogeneous coordinates to pixel coordinates by removing a dimension from a vector, • K an intrinsic matrix of the second camera (12), respectively of the first camera (11), • M an extrinsic matrix of the stereoscopic vision system, respectively a matrix representing the movement of the first camera (11) between the first and second time instants, • 0 a projection function in the three-dimensional scene of a pixel as a function of its depth, • K' an intrinsic matrix of the first camera (11), and • D^p ) is a depth of the pixel Pt determined by the first convolutional neural network, respectively by the second convolutional neural network.

4. Method according to one of claims 1 to 3, for which the first error is determined by a first loss function defined by: Lx = E;?avg( L^p), Lss(p) ) With: • L the first error, • avS an average of arguments, • Lslp) an error for a pixel P of the first image compared to the fifth image, • Lxx(p) an error for a pixel P of the second image compared to the fourth image.

5. Method according to one of claims 1 to 3, for which the first error is determined by a first loss function defined by: Lx = Epmin(£yf(p), Lss(p)) With: • L the first error, • min a minimum value of an argument, • Ls^p) an error for a pixel P of the first image compared to the fifth image, • Lxx(p) an error for a pixel P of the second image compared to the fourth image.

6. Method according to one of claims 1 to 5, for which the second error is determined by a second loss function defined by: Lm = Epavg(Lwj77), Lm^p}) With: • T the second error, • with an average of arguments, • L^p} an error for a pixel P of the first image compared to the seventh image, • Lms(p) an error for a pixel P of the third image compared to the sixth image.

7. Method according to one of claims 1 to 5, for which the second error is determined by a second loss function defined by: Lm = L,m(p)) With: • l' the second error, • min a minimum value of an argument, • Lm^p) an error for a pixel P of the first image compared to the seventh image, • Lms(p) an error for a pixel P of the third image compared to the sixth image.

8. Method according to one of claims 1 to 7, for which said third image is replaced by a black image, a black image being an image of which a set of colorimetric values ​​associated with the pixels of the image is zero.

9. Device (4) for determining a depth by a vision system on board a vehicle (10), said device (4) comprising a memory (41) associated with at least one processor (40) configured for implementing the steps of the method according to any one of claims 1 to 8.

10. Vehicle (10) comprising the device (4) according to claim 9.

Images