Image-based parking slot determination

Abstract

The invention refers to a computer-implemented method for determining a position or pose (pps) of a parking slot (PS) in a reference coordinate system based on image data, the method comprising the steps of: - receiving an image (I(PS) showing at least one parking slot (PS); - determining four edges (p) of the parking slot (PS); and - determining the position or pose (pps) of the parking slot (PS) in the reference coordinate system based on the four edges (p). According to the present invention determining the position or pose comprises: - assuming a geometry (g(PS)) of the parking slot (PS); - building at reconstruction matrix (A) based on the assumed geometry (g(PS)) and the determined edges; and - determining the position or pose (pps) of the parking slot (PS) in the reference coordinate system by applying a singular value decomposition (SVD) of the reconstruction matrix (A).

Description

[0001]202401241 1 Description Image-based Parking Slot Determination The present invention is concerned with a computer-implemented method for determining a position or pose of a parking slot in a reference coordinate system based on image data, with a preferred use of the determined position or pose for a parking function of an advanced driver assistance system, or for a scene interpretation module, by a computer-program, computer-readable (storage) medium and by an advanced driver assistance system comprising at least one camera, which is embodied to carry out the method according to the present invention. Advanced driver assistance systems (ADAS) for vehicles are based on a processing of various data sensed by various ADAS sensors, such as radar-, LiDAR- and ultrasonic sensors as well as cameras. By means of the ADAS sensors information relating to an environment of the vehicle can be obtained which in turn is used to realize various ADAS functions. ADAS functions on the one hand may include an assistance for the driver while control of the vehicle remains with the driver. On the other hand, depending on the level of automation, a full autonomously driving vehicle may be realized. Known ADAS functions for instance are various methods for detecting and / or classifying objects and / or obstacles in the vicinity of the vehicle, methods for lane detection and / or lane departure, methods for rain detection or also various parking assistance functions. In case of parking assistance functions, various methods to support parking of a vehicle are known, e. g. support to search for a parking slot or to assist with a parking process, but also for autonomously performing the parking process. The higher the level or automation, the more reliable and detailed information about the precise position of the parking slot is needed. It is e. g. known to perform a backpropagation of 2D image points onto an assumed ground plane by computing an intersection of a plane and a line in 3D. The resulting 202401241 2 points are used as the edges of the parking slots. Even though such approach is rather straightforward to implement it may be error prone. For instance certain constraints such as the assumption of parallel border lines of the parking slot cannot be guaranteed. In addition, a subsequent fitting of a parking slot geometry onto the determined points by minimizing a cost function, e. g. a squared distance of fitted points and the computed points, is carried out as well. Such optimization may be performed by means of a least square method and by assuming Gaussian added noise. However, the fitting process minimizes an error on the plane, but not for the source of the noise, which is the image itself. The backpropagation furthermore magnifies errors in a non-linear way. Due to the non-linearity, the noise is not Gaussian which makes the optimization process rather complex. Accordingly, this approach is either accompanied by lower accuracy or higher computational demand. Alternatively, it is possible to transform the image into a birds eye view by a flat world assumption, i.e. assuming a vehicle environment as a known plane, and provide the bird eye view images to a neural network. However, this necessitates an accurate formulation of one single reference plane. In reality, however, parking slots could be located at different planes. That way, also this approach can cause significant error and / or require increased computational demand. Finally, it is as well known to use a 3D bird eye view representation of the image data directly as input for a neural network embodied to detect the parking slot based on the input data. Such procedure was e. g. described in “NVAutoNet: Fast and Accurate 360° 3D Visual Perception For Self Driving” by T. Pham et al., available on arXiv:2303.12976. A suitable network for this task, however, has a rather complex architecture accompanied by higher computational demand, and 3D ground truth data would be needed for its training. Another approach was described by X. Zhou et al. in “Objects as Points”, published on arXiv:1904.07850, suggested a detection of objects based on single points, i.e. 202401241 3 the center point of the object’s bounding box. Subsequently, further object properties are regressed, e. g. size, 3D location, orientation and pose. A review of the use of computer vision for parking was provided by C. Eising et al. in the article “Computer Vision in Automated Parking Systems: Design, Implementation and Challenges”, available on arXiv:2104.12537. Based thereon it is an object of the present invention to improve the determination of the pose or position of parking slots based on image data. This object is achieved by means of the method according to claim 1, the use of the determined pose or position for an advanced driver assistance system according to claim 8, as well as by means of the, the computer program according to claim 9, the computer-readable medium according to claim 10, and the advanced driver assistance system according to claim 11. With respect to the method, the problem underlying the present invention is solved by a computer-implemented method for determining a position or pose of a parking slot in a reference coordinate system based on image data, the method comprising the steps of: - receiving an image showing at least one parking slot; - determining four edges of the parking slot; and - determining the position or pose of the parking slot in the reference coordinate system based on the four edges. According to the present invention determining the position or pose comprises: - assuming a geometry of the parking slot; - building at reconstruction matrix based on the assumed geometry and the determined edges; and - determining the position or pose of the parking slot in the reference coordinate system by applying a singular value decomposition of the reconstruction matrix. 202401241 4 The method according to the present invention provides an alternative to determine a position or pose of a parking slot in a reference coordinate system, e. g. a world coordinate system. The invention specifically addresses how a position or pose in three dimensions may be estimated in an efficient manner based on noisy 2D camera images by leveraging certain constraints with respect to the parking slot, e. g. some geometrical considerations referring to the parking slot. In addition, the provided solution features a high accuracy with respect to the parking slot reconstruction and also has the advantage of high computational efficiency. The pose thereby refers to a combination of position and orientation of the parking slot. It is of advantage, if the parking slot and / or the four edges of the parking slot are determined by using a neural network embodied to detect the parking slot and / or edges of a parking slot based on image input data. The neural network may be configured in the form of an object detector and may be trained using a data set of images showing labeled parking slots or labeled features characterizing a parking slot, e. g. its edges. According to a preferred embodiment of the method, the geometry of the parking slot is assumed to be a rectangle or parallelogram. This means, that opposing border lines of the parking slot are assumed to be parallel. This simplifies the computational effort of estimating or determining the position or pose of the parking slot. In this regard, it is further of advantage, if the geometry of the parking slot in the form of a parallelogram or rectangle is described by equations, especially homogeneous equations, referring to two pairs of parallel lines. Another preferred embodiment comprises that a ground plane in which the parking slot lies is determined. The ground plane may especially be determined based on a position and / or orientation of the vehicle and / or the camera, i. e. based on known 202401241 5 coordinates of the vehicle and / or camera, preferably in the reference coordinate system. The ground plane may further advantageously be characterized by a normal vector. Then, a height of the ground plane may be estimated based on the normal vector and be used to determine a height of the position and / or pose of the parking slot in the reference coordinate system. This embodiment further has the advantage, that, if a neural network is utilized to identify the parking slot in an image, the network may be trained based on 2D labels, whereas no 3D labeling of used training data is necessary. This simplifies the set-up of the step of identification of the parking slot in the images significantly. In this regard, it is of advantage, if the plane is transformed into a horizontal plane. A majority of parking slots is positioned in an essentially horizontal plane. Finally, in another preferred embodiment a non-linear refinement is carried out. The objective problem underlying the present invention is further solved by using the position or pose of the parking slot determined by the method according to the present invention and according to any of the embodiments described for a parking function of an advanced driver assistance system, or for a scene interpretation module. The objective problem is as well solved by a computer program comprising instructions, which, when the program is executed by a computer, cause the computer to carry out the method according to the present invention and by a computer-readable (storage) medium comprising instructions executable by at least one processor to perform the method according to the present invention or on which the computer program according to the present invention is stored. Finally, the objective problem is also solved by an advanced driver assistance system comprising at least one camera, wherein the advanced driver assistance 202401241 6 system is embodied to carry out the method according to any embodiments of the method according to the present invention. The invention and its preferred embodiments will be further described based on the subsequent figures, wherein Fig.1 shows a block diagram to illustrate the method according to the present invention; and Fig.2 shows a comparison of a state of the art method to locate a parking slot in an image and the method according to the present invention. In the figures the same reference symbols are used for the same elements. In Fig.1 a block diagram illustrating the method according to the present invention is shown. In an image I(PS) received in a first method step 1 showing a parking slot PS, a second method step 2 includes determining four edges p of the parking slot PS. The determination may include an identification of the edges p in the image I(PS) which in turn e. g. may be performed by any suitable detection algorithm, preferably by using a neural network embodied to detect a parking slot PS in image data I(PS). The determination of the edges p of the parking slot may further include their transformation into a reference coordinate system. For instance, in an exemplary embodiment as indicated in Fig.1, 3D rays may be calculated from the edges p identified in the image I(PS), e. g. by using the camera calibration. In addition, a ground plane for the parking slot PS, here in the form of a horizontal plane, may be assumed. Then, both the rays and the ground plane may be transformed to a canonical position in which the ground plane has a normal vector n=[0;1;0] meaning that it is horizontal in the camera frame. Accordingly, in this case the camera coordinate system serves as the reference coordinate system. Subsequently, in a third method step 3, based on an assumed geometry g(PS) of the parking slot PS, and based on its determined edges p, the position or pose pps of the parking slot PS may be determined. The determination of the pose pps of the 202401241 7 parking slot PS includes to build a reconstruction matrix A based on the edges p and the assumed geometry g(PS). The reconstruction matrix A represents a linear set of equations which is solved by applying a singular value decomposition SVD and optionally a non-linear refinement. Finally, the parameters determined by solving the set of linear equations may serve to calculate the pose or position pps of the parking slot PS. Several preferred embodiments of method step 3 will be explained in more detailed in the following: In a 2D representation a parking slot PS may be represented by an assumed geometry g(PS) of a parallelogram. Such constrain accurately represents the most important and frequent parking cases, such as parallel parking, perpendicular parking and tilted parking. The assumption of a parallelogram for the parking slot geometry g(PS) allows a precise alignment of a vehicle relative to the parking slot PS determined. A parallelogram may be described by a set of two parallel pairs of lines and by using the following homogenous equations: ^ 1,2where i refers to an index of each pair of parallel lines. That way, there are six parameters needed to describe the parallelogram. The parameters may be estimated based on the determined edges p of the parking slot PS. Parallelogram Fitting in 2D Each of the four edges of the parking slot PS defines one equation of the parallelogram. Because the two pairs of parallel lines of the parallelogram do not share parameters, the problem may be solved by two independent homogeneous 202401241 8 least square problems Aq=0, where A is the corresponding reconstruction matrix with ^^ ^ ^^^; ^^; 1; 0^or ^^ ^ ^^^; ^^; 0; 1^depending on to which line of the parallelogram a point belongs; and with the variables Because the assumed geometry of the parallelogram is described by homogenous equations, the variables may be rescaled, so that: Parallelogram Fitting in 3D In order to obtain a 3D representation, it is assumed that the parallelogram lies on a plane in the 3D space. If the orientation of the plane is known, a normal vector of the plane may be constructed and a height of the plain may be estimated. In case of a horizontal orientation of the parking slot PS, corresponding to [0;0;1] in a vehicle frame or [0;1;0] in a camera frame, a distance from this plane needs to be determined. A non-horizontal plane of the parking slot PS with known normal vector may be transformed to a horizontal plane by rotating the edges of the parking slot PS determined from the image I(PS). As discussed already above, a line on a certain plane in 2D may be described by 202401241 9 ^^^ ^ ^^^ ^ ^ ^ 0Line in 2D: ^ ^ #"^ ^^^; ^^^Plane in 3D on plane with normal vector ^ # #$ ^ ^0; 1; 0^ : ^" ^ ^^^; 0; ^%^The line in 3D is an intersection of two planes with direction: 00 1* which can be ^ ^with p0 being any point on the line.A point , #- ^ ^^-; ^-; 0-^ needs to satisfy that is lies on both planes:ℎ by taking 0- ^ 0 and This line may as well be represented by a 3D plane passing through a center of the * 0Note, that since^ li→m ^- ^ ^ 1, this also holds if ^^ ^ 0.nc is a normal vector of a plane passing through the origin of the coordinate system,which is a homogenous equation. Accordingly, it can be divided by −ℎ, ℎ ≠ 0,resulting in: 202401241 10 ^^ Distance to the line (plane) For any point pj representing an image pixel laying on the 3D line, the following scalar product needs to be zero: ^#5 ∙ ,= ^ 0For a point pj with unit length >,=>^ ^ 1 an angle θ between the point and theplane / line is defined by Further defining C ^ ^ΔC and by using the cosine sum formula:cos^J − K^ ^ cos J cos K ^ sin J sin Kone obtains: where ΔC ^ E^ − C. Accordingly, the residuals rj measure the sin of the anglebetween the point and the plane without normalizing for the amplitude of the plane normal. Note, that this is linear, while the normalized residual would be non-linear.The line may be fitted by building a linear set of equations following:^^5 ^ 0with the reconstruction matrix A built by stacking the pj direction vectors corresponding to the line. 202401241 11 Parallelogram on the 3D plane The starting point are again two pairs of parallel lines. For two parallel lines in the 2D plane x-z one obtains: ^^^^^ ^ ^^^^0 ^ ^^^^ % ^^ ^ 0^ ^ 1,2 After dividing by ℎ, ℎ ≠ 0 and by renaming the variables to ^^^^, ^^^^, ^^^^ ^^^^ ^ P , ^Q oneobtains: The two pairs of parallel lines can be solved independently, but they both need to be set with the same distance h. From this, the linear set of equations following ^^ ^ 0 where one row of the reconstruction matrix A with ,= ^R^=; ^=; 0=S is given by 202401241 12 ^= ^ R^=; ^=; 0; 0=Sor ^= ^ R^=; 0; ^=; 0=Sdepending on from which line the point pj is chosen. This problem is similar to the 2D case. Solving the system of linear equations in 3D requires a comparable amount as a computation of the corresponding problem in 2D. Non-linear refinementSince a singular value decomposition (SVD) solves with ‖^‖ ^ 1 and not ‖^5‖ ^ 1,additionally a non-linear refinement may be carried out for Rectanlge fitting In another exemplary embodiment the geometry g(PS) of the parking slot PS is assumed to be rectangular. Then, ^^^^ ^ ^^^^ −^%and^^^^ ^ ^^^^% ^As a consequence, the same for the parallel lines: ^With Because nx, ny define a single direction, there are 5 free parameters and the rows ofthe reconstruction matrix A are^= ^ R^=; ^=; 0; 0; 0; 0=Sor^= ^ R^=; 0; ^=; 0; 0; 0=Sor^= ^ R0=; 0; 0; ^=; 0; −^=S 202401241 13or^= ^ R0=; 0; 0; 0; ^=; −^=SIn Fig.2 compares a back projection of parking slots PS identified in image data on a plane (Fig.2a) with a determination of the position or pose pps of the parking slot by using the method suggested by the present invention (Fig.2b). The state of the art determination of the pose or position pps of the parking slots PS in Fig.2a is muss less accurate compared to the results obtained by using the method suggested by the present invention. A similar comparison of different methods to estimate the pose or position pps of a parking slot PS based on image data I(PS) is shown in Fig. 2c. PS refers to the parking slot PS in the reference coordinate system, or to the reel parking slot PS. PS1 refers to image points referring to the edges p of the parking slot determined by a back projection onto a plane. PS2 refers to the pose of the parking slot PS obtained by a 2D fitting and PS3 a pose of the parking slot PS obtained by performing the method according to the present invention.

Claims

202401241 14 Patent claims 1. Computer-implemented method for determining a position or pose (pps) of a parking slot (PS) in a reference coordinate system based on image data, the method comprising the steps of: - receiving an image (I(PS) showing at least one parking slot (PS); - determining four edges (p) of the parking slot (PS); and - determining the position or pose (pps) of the parking slot (PS) in the reference coordinate system based on the four edges (p); wherein determining the position or pose comprises: - assuming a geometry (g(PS)) of the parking slot (PS); - building at reconstruction matrix (A) based on the assumed geometry ( g(PS)) and the determined edges; and - determining the position or pose (pps) of the parking slot (PS) in the reference coordinate system by applying a singular value decomposition (SVD) of the reconstruction matrix (A).

2. Method according to claim 1, wherein the parking slot (PS) and / or the four edges (p) of the parking slot (PS) are determined by using a neural network embodied to detect the parking slot (PS) and / or edges (p) of a parking slot (PS) based on image input data.

3. Method according to claim 1 or 2, wherein the geometry (g(PS)) of the parking slot (PS) is assumed to be a rectangle or parallelogram.

4. Method according to claim 3, wherein the geometry (g(PS)) of the parking slot (PS) in the form of a parallelogram or rectangle is described by equations, especially homogeneous equations, referring to two pairs of parallel lines.

5. Method according to any of the preceding claims, wherein a ground plane in which the parking slot (PS) lies is determined.202401241 15 6. Method according to claim 5, wherein the plane is transformed into a horizontal plane.

7. Method according to any of the preceding claims, wherein a non-linear refinement is carried out.

8. Use of the position or pose of the parking slot (PS) determined by the method according to any of the preceding claims for a parking function of an advanced driver assistance system, or for a scene interpretation module.

9. Computer program comprising instructions, which, when the program is executed by a computer, cause the computer to carry out the method of any of the claims 1-7.

10. Computer-readable (storage) medium comprising instructions executable by at least one processor to perform the method of any of the claims 1-7 or on which the computer program according to claim 9 is stored.

11. Advanced driver assistance system comprising at least one camera, wherein the advanced driver assistance system is embodied to carry out the method according to any of the claims 1-7.

Images