Method for detecting objects within parking spaces or in the immediate vicinity of parking spaces

Abstract

The invention relates to a method for detecting objects (24, 26, 44) within parking spaces (10) or in the immediate vicinity of parking spaces (10) and for estimating dimensions (32, 38) of the objects (24, 26, 44). At least the following method steps are run through: The objects (24, 26, 44) are represented in a simple geometric manner by lines (28, 30) which run parallel or at an angle (62) to be estimated in relation to a front line (14) of the parking space (10). The distances / the distance of the lines (28, 30) relative to the parking space (10) is / are then estimated as the output of an artificial neural network which receives data from a sensor, in particular a camera, as input.

Description

[0001] Description

[0002] title

[0003] Methods for detecting objects within parking lots or in the immediate vicinity of parking lots

[0004] Technical field

[0005] The invention relates to a method for detecting objects within parking lots or in the immediate vicinity of parking lots and for estimating the dimensions of the objects.

[0006] State of the art

[0007] A method for detecting parking spaces using machine learning methods or neural networks is known. When detecting parking spaces, systems currently estimate the center of an entry line, the angle of the entry line, the length of the entry line, as well as the angle to the side lines and the length of the side line. Objects within parking spaces are usually represented as bounding boxes or polygons and are learned independently of the parking spaces. However, this requires a subsequent assignment of the objects to the respective parking spaces.

[0008] To ensure safe parking, it is necessary to detect objects within or in the immediate vicinity of parking spaces and estimate their respective dimensions. Of particular importance are objects that can prevent entry into a parking space, such as parking barriers, or that restrict entry lengthwise, such as tire stops.

[0009] Disclosure of the invention: According to the invention, a method for detecting objects within parking lots or in the immediate vicinity of parking lots is proposed, which further estimates the dimensions of the objects, wherein at least the following method steps are carried out: a) Simplified geometric representation of the objects by lines which run parallel to or at an angle to be estimated to a front line; b) Estimation of the distance(s) of the lines relative to the parking lot as output of an artificial neural network which receives data from a sensor, in particular a camera, as input.

[0010] In an advantageous further development of the method proposed according to the invention, it is proposed that for each parking space, it is estimated whether an object belonging to that parking space is located in or on the parking space.

[0011] Advantageously, in the method proposed according to the invention, the distances of a front edge and / or a rear edge of the object relative to the parking space are estimated for each parking space.

[0012] In the method proposed according to the invention, a midpoint of the front line or a midpoint of the parking space itself is referenced. Advantages include the reduction of the distance to be estimated, which is beneficial for the convergence speed and estimation accuracy. Furthermore, it should be emphasized that the mutual referencing focuses on a semantic relationship between the parking space and objects near the parking space, which also leads to an acceleration of convergence and an increase in robustness.

[0013] In the method proposed according to the invention, an angular deviation from the front line is advantageously estimated using an artificial neural network. Furthermore, the angle between the front line of the parking space and the front line of the object can be estimated using the artificial neural network, in addition to the distance. According to the invention, it is further proposed that a distance / dimension / depth dimension between the front edge and the rear edge of the object be estimated using the artificial neural network.

[0014] In the method proposed according to the invention, a transverse dimension or a width of the object is further estimated by means of the artificial neural network, which has multiple outputs.

[0015] In the method proposed according to the invention, the estimation processes are carried out using an artificial neural network. The artificial neural network used here is preferably one based on posture operations. Alternatively, this could also be replaced by an attention-based network, for example, a vision transformer network. In the method proposed according to the invention, the objects are represented in the most advantageous way possible in order to accelerate the training and to find the most robust and meaningful description possible.

[0016] Advantages of the invention

[0017] The implementation of the method proposed according to the invention ensures safe parking, as objects within or in the immediate vicinity of a parking space can be detected and their dimensions estimated. Particular emphasis is placed on objects whose presence makes driving into a parking space impossible or restricts entry into the parking space in the longitudinal direction. Such objects include, in particular, parking barriers or rolling resistance devices in the form of tire stops, which may be located at the sides of the parking space. The method proposed according to the invention allows for a significant reduction in the necessary parameters, thus reducing the size of an artificial neural network (ANN).Furthermore, it is advantageous to emphasize that the training of the KNN (Kinky-Nearest Neighbors) can be simplified because the aforementioned objects, particularly parking barriers or tire stops, are often very similar in geometry and are also arranged in similar positions relative to the parking area. Consequently, the artificial neural network's estimation deviates only slightly from the statistical mean of the parameters, which significantly reduces the effort required. Moreover, a semantic connection is established between the parking space and the object. The artificial neural network used in the above context, which preferably has multiple outputs, is trained not to estimate objects and parking spaces independently, but to predict a direct connection between the parking space and the objects associated with it.This can be achieved, in particular, by defining related parking spaces and objects as belonging together using an index that also needs to be estimated. Alternatively, it is possible to group the outputs of the artificial neural network so that one or more related objects can potentially be directly assigned to a parking space instance, with their probability of occurrence and geometric dimensions directly corresponding to the same output of the multi-output artificial neural network. This semantic connection represents a significant simplification in post-processing compared to independently estimating the parking space and object followed by mapping. Furthermore, the semantic connection between parking space and object eliminates the possibility of mapping errors occurring during merging and independent estimation of the parking space and object.

[0018] The solution proposed according to the invention advantageously optimizes and improves the training speed of the artificial neural network through significantly simpler convergence. Detection is considerably more robust due to the semantic association of the parking space with its associated objects, resulting in easier learning and shifting the focus to the co-occurrence of the parking space and objects. The method proposed according to the invention reduces post-processing effort because the parking space and its associated objects can be directly linked within the artificial neural network, preferably with multiple outputs.The individual objects are estimated using parameters within meaningful value ranges. For example, a distance from the front line of the parking lot to the front line of the object is defined as positive, since otherwise the object would be located in front of the parking lot. This allows for plausibility checks of the estimated parameters. Brief description of the drawings.

[0019] Embodiments of the invention are explained in more detail with reference to the drawings and the following description.

[0020] They show:

[0021] Figure 1 shows a possible geometric representation of tire stoppers on a parking area of ​​a parking lot.

[0022] Figure 2 shows an alternative geometric representation of tire stoppers on a parking area of ​​a parking lot.

[0023] Figure 3 shows a possible geometric representation of an object, for example a parking barrier, on a parking space of the parking lot and

[0024] Figure 4 shows a parking lot representation within an artificial neural network.

[0025] Embodiments of the invention

[0026] In the following description of embodiments of the invention, identical or similar elements are designated by the same reference numerals, and repeated descriptions of these elements are omitted in individual cases. The figures represent the subject matter of the invention only schematically.

[0027] Figure 1 shows a possible geometric representation of objects, for example tire stoppers, on a parking area 12 of a parking lot 10.

[0028] The schematic representation in Figure 1 shows that a first object 24 and a second object 26, for example, tire stops, are located near a first side line 20 and a second side line 22 of a parking area 12 of the parking lot 10. The parking area 12 of the parking lot 10 is bounded by a front line 14, the midpoint of which is designated by reference numeral 16, and a rear line 18. The possible geometric representation of the first object 24 and the second object 26 shown in Figure 1 indicates that a line representing a front edge 28 of the objects 24 and 26, i.e., a first distance 34, is estimated. Furthermore, an estimated second distance 36 indicates the rear edge 30 of the first and second objects 24 and 26 in the form of tire stops. These are located at a first object position 40 and a second object position 42, respectively.A width dimension 38 identifies the widths of the first and second objects 24, 26. The estimated first and second distances 34, 36 allow us to estimate the depth at which the two tire stopper objects 24, 26 are arranged in relation to the extent of the parking area 12, and the depth dimension, i.e., the distance between the front edge 28 and the rear edge 30, represented by the first and second distances 34, 36, can be estimated.

[0029] Figure 2 shows an alternative geometric representation of the first and second objects 24, 26 on parking area 12 of parking lot 10.

[0030] As shown in Figure 2, in this embodiment, an artificial neural network (ANN) estimates the first distance 34, i.e., the position of the leading edge 28 of the first and second objects 24, 26. In contrast to the procedure in the first embodiment according to Figure 1, in the alternative embodiment according to Figure 2, the artificial neural network estimates a depth dimension 32 between the leading edge 28 and the trailing edge 30 of the first and second objects 24, 26. In this embodiment as well, the first distance 34 extends, for example, from the midpoint 16 of the front line 14 to the leading edge 28 of the first and second objects 24, 26, which, for example, represent tire stoppers. The first object 24 is also shown in the first object position 40, while the second object 26 is located in the aforementioned second object position 42.

[0031] Figure 3 schematically illustrates a possible geometric representation of a third object 44 on parking area 12 of parking lot 10. Figure 3 shows that the third object 44 is located in a third object position 46, significantly closer to the center point 16 of the front line 14. The third object 44 in Figure 3 is, for example, a parking barrier, which occupies a position on parking area 12 that differs from the first and second positions 40, 42 of the first and second objects 24, 26. The third object 44 is characterized by its front edge 28 and its rear edge 30, both represented by lines. A depth dimension 32 exists between these, indicating the depth of the third object 44. For a third object 44 that represents a parking barrier, the width, i.e.,The width dimension 38 is of particular relevance and an important size for the representation in the artificial neural network.

[0032] Figure 4 schematically shows a representation of a parking lot within a neural network.

[0033] Figure 4 shows parking spaces 12 of adjacent parking lots 10. A vehicle 50 passes through a lane 54 in the direction of travel 52. The scene according to Figure 4 is overlaid by a grid 58, within which a grid cell 60 is highlighted by hatching. The parking space scene targeted by the vehicle 50, which is still moving in the direction of travel 52, is located in the grid cell 60 highlighted by hatching. Its parking space 12, or rather the midpoint 16 of the front line 14, is oriented at a vector angle 62 away from the vehicle 50.

[0034] The invention is not limited to the embodiments described here and the aspects highlighted therein. Rather, within the scope specified by the claims, a multitude of modifications are possible that fall within the bounds of what is considered skilled in the art.

Claims

Claims 1. Method for detecting objects (24, 26, 44) within parking lots (10) or in the immediate vicinity of parking lots (10) and for estimating the dimensions (32, 38) of the objects (24, 26, 44), comprising at least the following procedural steps: a) Simplified geometric representation of the objects (24, 26, 44) by lines (28, 30) (endpoint - endpoint) corresponding to the dimensions (32, 38), which run parallel or at an angle (62) to be estimated to a front line (14) of the parking lot (10), b) Estimation of the distance(s) of the lines (28, 30) relative to the parking lot (10) as output of an artificial neural network which receives as input data from a sensor, in particular a camera.

2. Method according to claim 1, characterized in that for each parking space (10) it is estimated whether an object (24, 26, 44) belonging to the parking space (10) is located in or near it.

3. Method according to claims 1 and 2, characterized in that for each parking space (10) distances (34, 36) of a front edge (28) and / or a rear edge (30) of the object (24, 26, 44) relative to the parking space (10) are estimated.

4. Method according to claims 1 and 2, characterized in that a midpoint (16) of the front line (14) or a midpoint - endpoint - endpoint of the parking space (10) is referenced.

5. Method according to claims 3 and 4, characterized in that an angular deviation to the front line (14) is estimated using an artificial neural network.

6. Method according to claims 1 to 3, characterized in that a depth dimension (32) between the line representing the front edge (28) and the line representing the rear edge (30) of the object (24, 26, 44) is estimated by means of the artificial neural network.

7. Method according to claims 1 to 6, characterized in that a width dimension (38) of the object (24, 26, 44) is estimated by means of the artificial neural network which has multiple outputs.

8. Method according to claims 1 to 7, characterized in that the estimation processes according to the preceding claims are carried out using an artificial neural network.

Images