Patent Application
20230377245
The invention pertains to the technical field of methods for simulating the effects of optical distortions of a windshield on the image recording quality of a digital image recording device, in particular of a digital image recording device from an advanced driver-assistance system.
Advanced/automated driver-assistance systems (ADAS) help divers in controlling their vehicles to improve safety of driver and people in the outside surrounding environment. They are on-board systems which are able to provide real-time information on the state of road traffic and/or on the state of the vehicle's mechanical and/or electrical equipment and components, assess the driver's state of fatigue or distraction, detect and anticipate possible threats from the vehicle's external environment, or help the driver perform certain difficult manoeuvres. Outstanding functions are adaptive cruise control, lane keeping, emergency braking, that are now mandatory to fulfil the Euro NCAP 5* requirements, parking assistance, autonomous driving, traffic signs detection, or anti-collision.
ADAS integrates numerous devices or sensors to collect data on the driver, the vehicle and/or their environment to perform their functions. In particular, they embed one or more digital image recording devices for real-time recording of images and/or video of the outside environment. The recording devices are often installed inside the vehicle, behind the windshield which in turn also acts as a protective shield.
The recorded images are processed by the ADAS depending on the desired functionality. For example, for night driving assistance, an infrared camera records a video of the outside environment which is then displayed in real time on the dashboard of the vehicle. For autonomous driving or emergency braking, a camera records images and/or video which are then processed by object detection and classification algorithms to extract the relevant data for further processing by the vehicle's automatic control unit.
For optimal operation of ADAS, high quality images and/or video are mandatory. For example, ADAS often comprises object detection and classification algorithms which real-time process recorded images/videos for real-time detection and classifications of objects from the surrounding environment of the vehicle. These algorithms are sensitive to colours, light intensity and light deviation in the recorded images/videos. In particular, these algorithms often contain an estimation of the position, distance and/or velocity of the objects in relation to the car position, velocity and/or trajectory. This quantity is important in many ADAS functions, for instance to calculate the risk of a collision. Any artefacts in the images/videos may cause wrong detection and/or classification and/or estimation of the distance, position or velocity of the target which may lead to defective operations of ADAS and safety failures, for instance a wrong estimation of the risk of collision.
It is then a strong requirement that the optical path of ADAS is free from optical distortions which may degrade the quality of the recorded images or/ videos. In this scope, lot of efforts has been made and is still being made to improve the optical quality of windshields, since they are key components in the optical system of ADAS and have a key role in their overall performances.
Beside, windshields, as efficient as they could be, still remain optical systems in themselves with their own optical properties and limitations. Calibrating tools for ADAS has then been developed which try to evaluate and manage the effects of windshield on recorded images and/or videos by ADAS.
EP 3 293 701 B1 describes a calibration method for ADAS to compensate the optical distortions from windshields. Optical distortions are calculated from the comparison of recorded images of a planar checkerboard pattern through and out a reference windshield. Only static parameters are corrected, i.e. parameters that are considered to be the same for all windshields of a given category of product and, once corrected, are corrected also for all unmeasured windshields of same category. The static parameters are limited to tilt angle, refractive index or thickness of the pane.
In this method, intrinsic non static optical properties of each individual windshield are completely overlooked, which may cause inefficient calibration in long term use, in particular if a windshield departs too much from the reference windshield regarding optical distortions. Windshields of same kind often show variations of their respective distribution of optical properties owing to normal tolerances of the manufacturing process, either because they come from different manufacturing lines or, within the same line, because of normal tolerances on the bending process, inhomogeneity of the furnace temperatures, defects during lamination, coating if coating is present . . . . This distribution of optical properties is different from a windshield to another within the same category and are by definition non static, which means that they cannot be considered as identical for all windshields in the same class. In this regard, each windshields of same kind have to be considered as an optical system in itself, different from the others, and its intrinsic optical properties have not to be neglected.
FR 3 032 820 A1 describes a calibration method for an ADAS to compensate the optical distortions from the windshield in front of which it is installed. Correction parameters are calculated while the car is driving from the deviations between measured positions of a static objet from recorded images of said object through the windshield and predicted positions for said object.
The method shows several disadvantages. It has to be performed once the ADAS is already installed or in use on the vehicle and while the vehicle is driving. This can be hardly implemented in a production line. A system for calibrating the ADAS, comprising both the calibration device to perform the method and the facilities to record images of the road, has to be available on the operation place, which may become tedious and expensive in an industrial context.
As an alternative, WO 2020164841 A1 provides a computer implemented method to form training datasets containing simulated images viewed through a windshield which can be later fed to an ADAS for calibration or be used for windshield qualification. In the method, the effect of different windshields is simulated by an optical filter applied on the images of a dataset.
Using an optical filter to simulate the effect of windshield is not always an appropriate strategy to accurately simulate the effect of certain optical particularities. In particular, as a homogeneous filter, it cannot account for variation of said optical power along the windshield. Since the optical filter is used as a static filter, the effect of the intrinsic optical properties of each individual windshield is overlooked, with the same drawbacks mentioned above.
Finally, in the ADAS field, in particular because the effect of windshield is often overlooked during calibration and/or the windshield is falsely considered as a negligible component, it is also a common practice to require optical specifications for windshields which may be hard to fulfil in the future, in particular regarding the optical power in the view zone of the digital image recording devices. In this scope, anticipating this technical limitation by a better considering the particularities of windshield during ADAS calibration may be very valuable.
There is a need for a method which solves all the afore mentioned problems. In particular, there is a need for a method for simulating the effects of distortions of a windshield on the image recording quality of a digital image recording device, which accounts accurately for the intrinsic optical particularities of each windshield considered individually, is able to provide valuable information of said effects for calibrating an ADAS without requiring cumbersome calibrating systems, and is compatible with production throughputs.
The present invention relates to a computer implemented method and a process as described in claims.
In particular, the invention is based on adapted stochastic ray tracing method to simulate the effects of the optical distortions of a windshield on the image recording quality of a digital image recording device from a measured optical quality function of the windshields related to those optical distortions and an image of a landscape likely to be recorded by a digital image recording device. It provides a simulated image of the image as viewed through said windshield taking into account its intrinsic optical particularities.
The method according to the invention can be used in a process for evaluating the optical quality of windshields for a use with a digital image recording device, in particular with a digital image recording device of an automated driving and advanced safety system. It can also be used in a process for calibrating digital image recording device of an automated driving and advanced safety system.
An aspect of the invention also relates to a process for evaluating the performances of an object detection and classification algorithm of an automated driving and advanced safety system, in which set of simulated images of one or several images of landscapes as viewed through one or several windshields are fed to an object detection and classification algorithm of an automated driving and advanced safety system to monitor its performances in object detection and classification.
A first outstanding advantageous effect of the invention is that the intrinsic optical properties of the windshield, i.e. the optical properties of the windshield as an optical system in itself are not overlooked or neglected anymore, but are taken into account in their worth influence on the quality of recorded image, even when the optical quality of the windshields is deemed to fulfil requirements. Comparing to methods of prior art, potential negative side-effects on image quality when a windshield departs too much from a reference windshield regarding optical distortions are completely eliminated.
A second advantage of the invention is that it does not rely on the computation of surrogate parameters, such as optical filters, to model the optical properties of the windshield. This kind of parameters are often limited in their application and not accurate enough for certain properties. Instead, the method of invention takes a measured quality function as a basis for the simulation and relies on the laws of optics, so that the accuracy and reliability of the simulated images are greatly improved.
Preferably, the measured quality function of the windshield which is related to the optical distortions of said windshield come from real-time measurements performed on production line of windshields. A measuring device may be at some location on the production line and configured to measure or collect data related to the quality function on the fly, on each windshield being produced. This approach is highly valuable as it is compatible with most common production throughputs and does not require later measurement on batches of already produced windshields, which is a time-consuming and tedious operation. The measured quality functions of the windshields can be stored in a database for later processing with a method according to the invention.
In the invention, the image of landscape may be of any kind. It may a recorded or a simulated 2D or 3D image of a 2D or a 3D real or simulated landscape. Preferably, it's a 2D image of landscape as they may require less computing resources for similar results compared to 3D images.
In the context of the invention, an image of landscape has to be understood as a representative image of a representative landscape which is likely to be recorded by a digital image recording device, in particular a digital image recording device of an automated driving and advanced safety system, through a windshield.
Because it is based on the use of images of landscapes, another advantage of the invention compared to prior art is that the need of cumbersome material calibration systems to evaluate the effects of windshields on recorded image of a material landscape by an ADAS, is alleviated, and even eliminated. Furthermore, the versatility in the choice of the landscape is greatly improved as it does not require to adapt and/or transform the existing material landscape to get a new one.
The method according to the invention allows to accurately simulate batches or sets images of landscape as viewed by a digital image recording device through windshields of a set or batches of windshields without requiring inaccurate approximations of the optical properties of windshields regarding their potential optical distortions and side-effects on quality of recorded images. As consequence, it can be advantageously used to evaluate the optical quality of windshields for applications in automated driving and advanced safety system, to calibrate digital image recording device of such system or to evaluate the performances of an object detection and classification algorithm of such system, as described in certain embodiments of the invention.
With reference to
On the windshield 1002, the view zone 1003 through which the digital image recording device 1001 is recording images/videos is usually demarcated by glaze or enamel stripes that are used as ornament and/or for concealing some elements such as fixing joints and electronics.
The inclination angle between the median axis (C) of the demarcated zone 1003 and the optical axis (B) of the digital image recording device 1001 may be equal to or may differ from the inclination angle between the median axis (C) of the demarcated zone 1003 and the median axis (A) of the vehicle chassis, which generally is horizontal. In most transport vehicles, the windshield 1002 is not vertical but inclined at a certain angle which is the installation angle of the windshield 1002 on the vehicle frame. In other words, with respect to the digital image recording device 1001, the windshield 1002 is inclined at an angle which corresponds to its condition of use.
In this configuration, the quality of the recorded images or/videos may partly depend on the optical quality of the demarcated view zone 1003 of the windshield 1001. In particular, the windshield 1002 must be devoid of optical distortions. However, as efficient as they could be, windshields still remain optical systems in themselves with their own optical properties and limitations, and their effects on the recorded images/videos remain to be evaluated for each of them.
Furthermore, as a windshield 1002 is often inclined with respect to the digital image recording device 1001, the distance between the surface of its demarcated view zone 1003 and the lens of the digital image recording device 1001, in one hand, and the length travelled by light ray inside the windshield 1002 itself, in other hand, vary vertically, i.e. along the (C) axis. Because the optical path length is not constant along this axis, further optical distortions, and/or deflections, may occur which have also to be considered while evaluating the effects of windshields on the image/video quality.
With reference to
In the field of ray tracing, ‘global illumination’ is a well-defined expression. It encompasses all kinds of illumination, whether direct or indirect, coming from a light source, i.e. the image 2001 of a landscape in the present invention. Direct light is the light coming directly form the light source and indirect light is the reflected, refracted and/or diffused light from surfaces and/or volumes in the scene.
The frustum 2004f of the virtual camera 2004 corresponds to the field of view of said camera 2004 in the ray tracing scene. It is usually represented as a kind of pyramid of vision which is considered representative of the field of view of a real camera. Concretely, it constitutes the region of the ray tracing scene which may appear on the screen. ‘Frustum’ is well-defined term in the art.
In the context of the invention, optical distortions are to be understood as encompassing all the optical aberrations of the windshield which may affect the optical path of light so that the magnification varies across the field of view of said windshield at a given working distance. In other words, optical distortions are all aberrations which induces deviation from rectilinear projection; projection in which straight lines remain straight across the field of view. At some extent, in the context of the invention, optical distortions may also comprise blur as long as it comes from the optical particularities of the windshield itself. Optical distortions can come from variations in refractive index, in surface roughness/profile, or in thickness of the windshield. They may also occur from defects.
As explained above, the measured optical quality function OQF of the windshield 1002 which is related to the optical distortions of said windshield 1002 may come from real-time measurements performed on production line of windshields. In this context, it may be the measured transmitted wavefront error of the windshield, the measured surfaces profiles and/or the measured distribution of complex refractive index. In a preferred embodiment, the optical quality function OQF related to the optical distortions is the measured transmitted wavefront error, whose measurement can be relatively easily implemented on production lines of windshields since it can be performed quickly with acquisition rates which are compatible with throughputs of most production lines.
Object detection and classification algorithms implemented in data processing systems of ADAS are often sensitive to colours, i.e. light wavelengths, and may perform better in colour filtered, for instance red or green filtered image. On the other hand, colour filtering may help to reduce image artefacts or anomalies, in particular for night-time images. Furthermore, digital image recording device may sometimes comprise colour filter array on some pixels. On this scope, the image 2001 of a landscape 1005 may advantageously undergo some prior image processing in order to improve the image quality and/or extract wavelength specific information from said image and/or to render the effect of colour filter arrays.
In certain advantageous embodiments of method according to the invention, before step (a), the input image 2001 of a landscape 1005 may be digitally preprocessed, first with a colour filter array, for instance a clear-red-clear-clear or a clear-clear-red-clear filter array, and/or, second, with a demosaicing algorithm, in particular with a nearest-neighbour interpolation kernel.
At step (a) of the method according to the invention, at least one of the faces of the modelled sheet 2002 of transparent mineral glass is textured with the measured quality function I3001. The texture of the textured surface may be modelled in different ways. For instance, it may be modelled with a bump map or a displacement which is representative of the measured optical quality function.
At step (b) of the method according to the invention, the global illuminance GI is calculated with a stochastic ray tracing method. Compared to conventional, i.e. non-stochastic, ray tracing methods which relies on a given number of light rays to be drawn for each pixel, stochastic ray tracing methods needs less computer resources and computing time. They allow to focus calculation on light rays reaching the virtual camera and to those light rays going nowhere. In the invention, beside these advantages, using a stochastic ray tracing method also allows to get the same level of accuracy as with conventional ray tracing methods. Monte Carlo ray tracing methods, in particular Metropolis Light Transport ray tracing methods, provide valuable results.
Different stochastic ray methods are available in the art. Any of them may be used at step (b) providing it is adapted to calculate the global illuminance arriving through the modelled inclined sheet of mineral glass. In particular, a stochastic path tracing method may be used, as it allows to get a high level of precision and sharpness.
Several physically based rendering engines are available in the art to implement stochastic ray tracing in step (b). Examples of engines are Indigo renderer or Lux-CoreRender.
According to the invention, the 3D projection O3001 of the global illuminance GI is performed with a optical camera model (OCM). Preferably, it may be a physically realistic camera projection matrix. Alternatively, any other relevant quantity representative of the imaging operation, i.e. operation which converts simulated lights into colour values in a given colour space, e.g. RGB or XYZ, performed by the camera may be used, e.g. a point-spread function (PSF), an optical transfer function (OTF) or a real modelling of each interface of the different optics of the camera lens (optical design). Ideally, the physically realistic camera used to model the virtual camera 2004 may be the camera projection matrix, the optical transfer function, the point-spread function, the optical design or the like, of the digital recording device. However, in most cases, such models of the digital recording device may be not available or difficult to obtain. Nevertheless, accurate projection may still be obtained by considering the virtual camera as an ideal thin lens and using with the camera projection matrix of an ideal thin lens as a camera model.
In certain embodiments of the method according to the invention, the optical camera model may be the measured or the simulated optical transfer function of the digital image recording device 1001. A direct advantage may be that the 3D projection may be closer to the 3D projection performed by the digital image recording device and more accurate for digital image recording device with specific optics.
On the
Alternatively, according to a certain embodiment of method according to the invention, with reference to
In the context of the invention, an environment map is a map projected on the scene coordinate system in order to make it the background of the scene in a ray tracing method. The environment map converts pixel locations in an image into angles of incidence. This makes any object to appear at an infinite location. In the embodiment, the image 2001 of the landscape 1005 may be projected on the all or part of the inside surface of the sphere with spherical mapping technique, so that the image 2001 used as light source 2003 corresponds to an infinitely distant illumination which is afterwards refracted, reflected and/or diffused on the modelled sheet of transparent mineral glass. Main advantage may be that the light rays coming from each pixel of the image 2001 used as a light source 2003 form a rectilinear and parallel beam from an angular direction, as if the image was located at infinite distance. The computing workload may be alleviated and the computing time may be reduced without prejudicing the accuracy of the output image O3001.
Placing the image 2001 of the landscape 1005 at finite or infinite distance in the scene of ray tracing is a matter of choice and will depend on the specifications of both the digital image recording device and the ADAS.
At step (a) of the method according to the invention, the modelled sheet 2002 of transparent mineral glass may represent the whole surface of the windshield 1002, and the surface of the at least one of two main faces 2002a, 2002b of said modelled sheet 2002 may be textured with the measured quality function OQF of the whole surface of said windshield 1002.
In a certain embodiment of method according to the invention, in step (a), the sheet 2002 of transparent mineral glass is modelled so that said sheet 2002 of transparent mineral glass represents only the demarcated zone 1003 of the windshield 1002 in front of which a digital image recording device 1001 may be placed. The demarcated zone 1003 is the region of interest for the recording of image through the windshield 1002 by the digital image recording device 1001. Therefore, it may be advantageous regarding computing time and speed of simulation to model only this demarcated zone 1003.
Alternatively, in a certain embodiment of method according to the invention, in step (a), the sheet 2002 of mineral glass is modelled so that the global illuminance GI calculated at step (b) is only the global luminance GI arriving through the part of modelled sheet 2002 of mineral glass which corresponds to a demarcated zone 1003 of the windshield 1002 through which a digital image recording device 1001 may record an image.
For instance, in step (a), the modelled sheet 2002 of transparent mineral glass may be modelled to represent the whole surface of the windshield 1002, and the surface of the at least one of two main faces 2002a, 2002b of said modelled sheet 2002 may be textured with the measured optical I3001 quality function of the whole surface of said windshield 1002. The portion of the surface of the modelled sheet 2002 which does not correspond to the demarcated zone 1001 may then be blackened in order to prevent part of the light coming from the light source from going through that portion. The global illuminance (GI) calculated at step (c) is then global luminance (GI) arriving through the non-blackened portion of the surface which corresponds to the demarcated zone 1003.
In certain embodiment of the method according to the invention, in step (a), a second sheet of transparent mineral glass and a polymeric interlayer may be modelled so that the first and second sheets of transparent mineral glass are assembled to each other by the polymeric interlayer, for instance a PVB interlayer, to form a laminated glazing, such as a windshield. The optical quality function I3001 of the windshield 1002 may then the measured optical quality function of the windshield or a calculated combination of measured optical quality functions of each sheet of transparent mineral glass and the polymeric interlayer.
The method of the invention is computer implemented. With reference to
As image processing and ray tracing often require a lot of computational power to process large amounts of data, the data processing system 5000 may advantageously comprise one or more Graphical Processing Units (GPU) whose parallel structure makes them more efficient than CPU, in particular for image processing in ray tracing.
Another object of the invention is to provide a computer program I5001 comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the invention according to any embodiments described herein.
Any kind of programming language, either compiled or interpreted, can be used to implement the steps of the method of the invention. The computer program can be part of a software solution, i.e. part of a collection of executable instructions, code, scripts or the like and/or databases.
Another object of the invention is to provide a computer-readable medium 5002 comprising instructions which, when executed by a computer, cause the computer to carry out the method according to any of the embodiments described herein.
The computer-readable storage 5002 is preferably a non-volatile storage or memory, for example hard disk drive or solid-state drive. The computer-readable storage can be removable storage media or a non-removable storage media as part of a computer.
Alternatively, the computer-readable storage may be a volatile memory inside a removable media. This can ease the deployment of the invention into many production sites.
The computer-readable storage 5002 can be part of a computer used as a server from which executable instructions can be downloaded and, when they are executed by a computer, cause the computer to carry out a method according to any of the embodiments described herein.
Alternatively, the program may be implemented in a distributed computing environment, e.g. cloud computing. The instructions can be executed on the server to which client computers can connect and provide encoded data as inputs to the method of the invention. Once data are processed, the output can be downloaded and decoded onto the client computer or directly send, for example, as instructions. This kind of implementation can be advantageous as it can be realised in a distributed computing environment such as a cloud computing solution.
The method according the invention is well adapted to be used in a process for evaluating the optical quality of windshields for a use with a digital image recording device, in particular with a digital image recording device of an automated driving and advanced safety system. In this scope, the effects of the optical distortions may be simulated for each windshield of a batch of windshields from their respective measured quality function and a reference image of landscape or a set of reference images. Once the effects of the optical distortions are simulated, the windshields may be compared, and those less altering the image quality regarding required specifications may be selected for use with a digital image recording device. Alternatively, some criteria may be defined on the simulated images or on the output of ADAS algorithms fed with the simulated images in order to reject non-compliant windshields.
The method according the invention is also well adapted to be in a process for calibrating digital image recording device of an automated driving and advanced safety system. For instance, the effects of optical distortions of a reference windshield or a windshield from a set of reference windshields, i.e. windshields whose optical quality is considered to fulfil requirements for a use with a digital image recording device, may be simulated from a set of images of landscapes. The output images may then be used to calibrate and/or select the relevant features of a digital image recording device.
According to an aspect of the invention, with reference to
| Number | Date | Country | Kind |
|---|---|---|---|
| 20315428.1 | Oct 2020 | EP | regional |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2021/076676 | 9/28/2021 | WO |
