Method for performance evaluation, method for adapting a selection calculation model and method for identifying at least one reference target object

Abstract

The invention relates to a method for performance evaluation (10) of a target object selection of a selection calculation model (12) for automated distance control of a vehicle, comprising the steps of providing (14) input data (16) based on measurement data (18) of environmental objects (20) of a vehicle environment (22), providing (30) at least one reference target object (34) having reference target object data (32) from the environmental objects (20) in the input data (16) for the automated distance control as a control target, operating (36) the selection calculation model (12), which selects at least one target object (38) from the environmental objects (20) of the input data (16) as a control target for the automated distance control, object comparison (42) between the selected target object (38) and the reference target object (34), and calculating at least one evaluation parameter (44) of the target object selection depending on the object comparison (42).Furthermore, the invention relates to a method for adapting (46) a selection calculation model (12) and a method for identifying (62) reference target objects (34).

Description

[0001] The invention relates to a method for performance evaluation according to claim 1. Furthermore, the invention relates to a method for adapting a selection calculation model and a method for identifying reference target objects. State of the art

[0002] Adaptive Cruise Control (ACC) is a driver assistance system in a vehicle that increases driving comfort and safety on highways and in other traffic situations. ACC enables automated distance-following by monitoring and adjusting the relative distance and speed between the vehicle and a vehicle in front selected as the control target.

[0003] Environmental sensors are used to detect the vehicle's surroundings, specifically measuring the distance between the vehicle and the target object. If the vehicle in front slows down or the distance decreases, the vehicle's speed is reduced to maintain the specified safety distance. Conversely, if the distance to the vehicle in front increases, the vehicle accelerates, as long as the vehicle's speed remains within a speed range preset by the driver.

[0004] DE 10 2012 210 608 A1 describes a method for generating a control parameter for Adaptive Cruise Control (ACC) by evaluating environmental data obtained by means of environmental object recognition and image processing. An image- or video-based distance assistance system takes into account lane information, the position and dimensions of environmental objects, and the direction of travel to generate the control parameter.

[0005] If there are several surrounding objects that could be considered as target objects, the target object is selected by choosing the surrounding object that is closest to the vehicle as the target object. DE 10 256 529 A1 describes a distance control system that, in addition to the distance to the vehicle immediately ahead, also takes into account the distances between the vehicle and vehicles traveling further ahead.

[0006] The target object selection in automated distance control is crucial for the functionality and reliability of the automated distance control and thus for the road safety of the vehicle. Disclosure of the invention

[0007] According to the present invention, a method for evaluating the performance of target object selection is proposed, comprising the features of claim 1. This allows the performance of target object selection in automated distance control to be objectively evaluated. Target object selection can be standardized. Automated distance control can be performed more accurately, safely, and reliably.

[0008] The vehicle can be a motorized vehicle, in particular a motor vehicle, a two-wheeled vehicle, or a commercial vehicle. The vehicle can be a mobile robot. The vehicle can be the one used for applying the automated distance control, while the measurement data comes from at least one reference vehicle other than the vehicle.

[0009] The automated distance control can be an adaptive cruise control system. The automated distance control can automatically maintain a distance, in particular a minimum distance, between the vehicle and the target object by adjusting the vehicle's longitudinal dynamics, in particular by adjusting the vehicle's speed and / or acceleration, in particular while maintaining a predefined speed range. The automated distance control can implement a distance-following mode of the vehicle.

[0010] When operating the automated distance control, sensor data from a vehicle sensor device that detects a vehicle environment of the vehicle can be used as input data.

[0011] The vehicle sensor device may comprise at least one vehicle sensor, in particular a radar sensor, an ultrasonic sensor, a LIDAR sensor or a camera.

[0012] The environmental objects can be other vehicles, plants, buildings, living beings, facilities, equipment, mobile devices or the like.

[0013] The measurement data can be obtained partially or completely from data sources other than the current vehicle environment of the vehicle. The measurement data can include measurement data previously recorded by at least one reference vehicle. The vehicle environment with the environmental objects of the measurement data can be that of at least one reference vehicle, in particular during a reference run of the reference vehicle. This vehicle environment can differ from the current vehicle environment of the vehicle.

[0014] The target object is a reference object for the automated distance control, to which a distance and / or relative speed is specified. The target object can be another vehicle or mobile device. The target object can be traveling ahead of the vehicle or reference vehicle.

[0015] The selection calculation model calculates a target object selection for the vehicle's automated distance control. Target object selection is the selection of at least one target object. The selection calculation model can be a deterministic and / or stochastic model. The selection calculation model can comprise a neural network. The neural network can have multiple layers, each with multiple interconnected neurons. The connections between the neurons can be described by weights. The weights can be learnable. The neural network can be a deep neural network with a large number of neurons.

[0016] The input data can be formed from the measurement data or by processing the measurement data. The input data can include an environmental object list with the detected environmental objects of the vehicle's surroundings.

[0017] The input data may comprise at least one image or a temporal sequence of images of a vehicle's surroundings. The image may be an image. The temporal sequence of images may be a video recording. The measurement data may be radar measurement data and / or LIDAR measurement data.

[0018] The input data may be generated by a vehicle sensor device that is identical or comparable to that used in an application of the selection calculation model for automated distance control. The input data may have a data form, data integrity, and / or data structure that is identical or comparable to that used in an application of the selection calculation model for automated distance control.

[0019] At least two reference target objects can be determined from the input data, in particular during a target object change from a first reference target object to a second reference target object, for example during a lane change, a cutting-in operation of the second reference target object or a cutting-out operation of the first reference target object.

[0020] Providing the reference target object may be preceded by identifying the reference target object in the input data. The reference target object may be linked to the input data.

[0021] The identification of the reference target object in the input data can be performed manually, i.e., by a person, or supervised. The identification of the reference target object can be verified. The identification and / or verification can be supported by images or video recordings corresponding to the input data.

[0022] Providing the reference target object can be done by applying a deterministic model. Providing the reference target object can be done rule-based. Providing the reference target object can be a retrieval of the reference target object associated with the input data.

[0023] By operating the selection calculation model, at least two target objects can be selected, in particular during a target object change from a first target object to a second target object, for example, during a lane change, a merge of the second target object, or a merge of the first target object. The object comparison can be performed between the two target objects and two reference target objects.

[0024] In a preferred embodiment of the invention, it is advantageous if the evaluation parameter is calculated depending on a match and / or discrepancy between the selected target object and the reference target object. The evaluation parameter can comprise at least one numerical value. The evaluation parameter can be binary or have multiple values. The evaluation parameter can be one-dimensional or multi-dimensional.

[0025] In a preferred embodiment of the invention, the measurement data are recorded measurement data, and the provision of the input data includes access to the recorded measurement data. The measurement data can comprise measurement data recorded on a reference route. The provision of the input data can be preceded by a recording of the measurement data during a journey of the reference vehicle, in particular on the reference route.

[0026] In a preferred embodiment of the invention, it is advantageous if the input data comprise environmental object data relating to the environmental objects. The target object can be selected by the selection calculation model depending on the environmental object data. The environmental object data can comprise at least a position of the environmental object, in particular a relative distance between the reference vehicle and the environmental object, a relative speed of the environmental object with respect to the reference vehicle, and / or a dimension of the environmental object. The environmental object data can be partially or completely contained in the measurement data. The environmental object data can be calculated from the measurement data. Some or all of the environmental object data can be contained exclusively in the input data.

[0027] In a preferred embodiment of the invention, the reference target object data includes at least a relative distance between the reference vehicle and the reference target object, a relative speed of the reference target object with respect to the reference vehicle, and / or a dimension of the reference target object. The reference target object data can include a time point and / or a time range of a target object change from a first reference target object to a second reference target object.

[0028] In a preferred embodiment of the invention, it is advantageous if target object data derived from the environmental object data relating to the target object are assigned to the selected target object, and the object comparison comprises a comparison of the target object data with the reference target object data. The target object data can comprise a time and / or a time range of a target object change from a selected first target object to a selected second target object. The object comparison can comprise a comparison of the times and / or time range of the target object change of the reference target objects and the target object change of the selected target objects.

[0029] According to the present invention, a method for adapting a selection calculation model with the features of claim 7 is further proposed. In a preferred embodiment of the invention, the selection calculation model is modified if the evaluation parameter deviates from a target parameter. The target parameter can be a threshold value. If the evaluation parameter or the inverted evaluation parameter reaches or exceeds the threshold value, the selection calculation model can be adapted.

[0030] In a preferred embodiment of the invention, it is advantageous if the selection calculation model remains unchanged if the evaluation parameter at least corresponds to the target parameter. The target parameter can be a threshold value. If the evaluation parameter or the inverted evaluation parameter falls below the threshold value, an adjustment of the selection calculation model can be omitted.

[0031] According to the present invention, a method for identifying at least one reference target object is further proposed, having the features of claim 10. The labels may include a classification. The labels may indicate a surrounding object type, in particular a vehicle, building, traffic sign, road marking, living being, or plant, and / or an object behavior, for example, braking, turning, or crossing, of the surrounding objects.

[0032] The measurement data can be obtained during a journey with a reference vehicle. The measurement data can be stored. Providing the measurement data can include retrieving the stored measurement data and / or capturing the measurement data using a vehicle sensor device of the reference vehicle. The assignments can be created during and / or after the acquisition of the measurement data. The assignments can be created exclusively or additionally with the stored measurement data.

[0033] The assignments can include distance-dependent assignments and distance-independent assignments. The distance-dependent assignments refer to environmental objects and the behavior of the environmental objects for which the distance to the vehicle is decisive. The distance-dependent assignments include other environmental objects driving ahead, lane changes, static and / or dynamic obstacles on the roadway, and / or traffic density. The distance-independent assignments refer to environmental objects and the behavior of the environmental objects that are not, or only indirectly, decisive for the distance to the vehicle. The distance-independent assignments can be road conditions, weather conditions, traffic signs, traffic sign signals, lighting conditions, and / or roadway properties, in particular a road surface condition or roadway type.

[0034] The first computational model can be an object recognition model. The measurement data can be present as image recordings and / or video recordings. The measurement data can be recordings from a video camera in the reference vehicle. The first computational model can be an image-based computational model and / or a sensor fusion model. The first computational model can be provided for environmental perception of the vehicle's surroundings. The first computational model can include a convolutional neural network (CNN).

[0035] The first computational model can include single-stage detectors or multi-stage detectors. Single-stage detectors perform object detection and localization in a single pass, i.e., a single stage. Examples of single-stage detectors include YOLO (You Only Look Once), which divides the image into a grid and makes predictions about bounding boxes and class probabilities for each grid box; SSD (Single Shot Multibox Detector), which uses multiple convolutional neural networks applied to different image resolutions to generate bounding boxes and object classifications in a single step; and RetinaNet, which uses a Feature Pyramid Network (FPN) and a special focal loss to address the problem of imbalance between foreground objects and background objects.

[0036] Multi-stage detectors perform object detection in multiple steps, usually starting with the generation of region proposals, followed by object classification and bounding box regression. Examples of multi-stage detectors include R-CNN (Region-based Convolutional Neural Network), which uses Selective Search to generate region proposals and applies a CNN to each proposed region to classify surrounding objects and adjust bounding boxes; Fast R-CNN, which improves R-CNN by introducing a Region of Interest (ROI) pooling layer, thereby increasing the model's efficiency; and Faster R-CNN, which adds a Region Proposal Network (RPN) to generate region proposals, eliminating the need for an external method like Selective Search.

[0037] The assignments can include bounding boxes for the environmental objects. The bounding boxes can specify the localization of the environmental objects. The bounding boxes can specify the object positions and spatial extents of the environmental objects. The bounding boxes can specify object recognition, in particular, classification.

[0038] The target object can be identified based on the distance-dependent and distance-independent assignments. The second computational model can be a deterministic and / or stochastic model. The second computational model can include a neural network. The second computational model can be a rule-based computational model that identifies the target object primarily or exclusively based on rules.

[0039] Further advantages and advantageous embodiments of the invention emerge from the description of the figures and the illustrations. Character description

[0040] The invention is described in detail below with reference to the figures. They show in detail: Fig. 1: A method for adapting a selection calculation model and a method for performance evaluation in a specific embodiment of the invention. Fig. 2: A method for identifying at least one reference target object in a specific embodiment of the invention.

[0041] Fig. 1 shows a method for adapting a selection calculation model and a method for performance evaluation in a specific embodiment of the invention. The method for performance evaluation 10 of a target object selection of a selection calculation model 12 for an automated cruise control (ACC) system of a vehicle first comprises providing 14 input data 16 based on measurement data 18 of environmental objects 20 of a vehicle environment 22 of a reference vehicle 24. The measurement data 18 are measurement data 26 recorded by the reference vehicle 24 or multiple reference vehicles 24. Providing 14 the input data 16 comprises accessing the recorded input data 16. In particular, the input data 16 comprises environmental object data 28 relating to the environmental objects 20.The environmental object data 28 preferably comprise at least a relative distance of the reference vehicle 24 to the environmental object 20, a relative speed of the environmental object 20 relative to the reference vehicle 24 and / or a dimension of the environmental object 20. For example, the environmental object data 28 can be partially contained in the measurement data 18 and partially calculated from the measurement data 18.

[0042] Furthermore, at least one reference target object 34 comprising reference target object data 32 from the surrounding objects 20 is provided in the input data 16 for the automated distance control as a control target. In particular, the reference target object data 32 includes at least a relative distance of the reference vehicle 24 from the reference target object 34, a relative speed of the reference target object 34 with respect to the reference vehicle 24, and / or a dimension of the reference target object 34.

[0043] Furthermore, the selection calculation model 12 is operated 36, which selects at least one target object 38 as the control target for the automated distance control from the environmental objects 20 of the input data 16, depending on the environmental object data 28. In particular, target object data 40, which are derived from the environmental object data 28 relating to the target object 38, are assigned to the selected target object 38.

[0044] Subsequently, an object comparison 42 is performed between the selected target object 38 and the reference target object 34, and at least one evaluation parameter 44 of the target object selection is calculated based on the object comparison 42. The object comparison 42 may include a comparison of the target object data 40 with the reference target object data 32. Preferably, the evaluation parameter 44 is calculated based on a match and / or discrepancy between the selected target object 38 and the reference target object 34.

[0045] Depending on the calculated evaluation parameter 44, a method for adapting 46 the selection calculation model 12 for the automated distance control is executed by changing the selection calculation model 12 depending on the evaluation parameter 44 calculated by the previously described performance evaluation method 10. The selection calculation model 12 is changed in particular if the evaluation parameter 44 deviates from a target parameter 48 and preferably remains unchanged if the evaluation parameter 44 at least corresponds to the target parameter 48.

[0046] Fig. 2 shows a method for identifying at least one reference target object in a specific embodiment of the invention. The method for identifying 50 reference target objects 34 initially comprises providing 52 measurement data 18 indicating environmental objects 20 of a vehicle's surroundings 22 and creating 54 associations 56 with the environmental objects 20 in the measurement data 18 by applying a first calculation model 58. Providing 52 the measurement data 18 may comprise acquiring the measurement data 18 by a vehicle sensor device 59 of a reference vehicle 24. The first calculation model 58 may be an object recognition model 60.

[0047] Furthermore, an identification 62 of at least one target object 38 takes place depending on the assignments 56 by applying a second calculation model 64. The second calculation model 64 can be a rule-based calculation model 66 that identifies the target object 38 mainly or exclusively in a rule-based manner.

[0048] Subsequently, an output 68 of the identified target object 38 takes place as a reference target object 34, for example for the provision of the reference target object 34 in the Fig. 1 described performance evaluation procedures. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2012 210 608 A1

[0004] DE 10 256 529 A1

[0005]

Claims

[1] Method for performance evaluation (10) of a target object selection of a selection calculation model (12) for an automated distance control of a vehicle, comprising the steps Providing (14) input data (16) based on measurement data (18) of environmental objects (20) of a vehicle environment (22), Providing (30) at least one reference target object (34) having reference target object data (32) from the surrounding objects (20) in the input data (16) for the automated distance control as a control target, Operating (36) the selection calculation model (12), which selects at least one target object (38) from the surrounding objects (20) of the input data (16) as the control target for the automated distance control, Object comparison (42) between the selected target object (38) and the reference target object (34), Calculating at least one evaluation parameter (44) of the target object selection depending on the object comparison (42). [2] Performance evaluation method (10) according to claim 1, characterized by that the evaluation parameter (44) is calculated depending on a match and / or deviation between the selected target object (38) and the reference target object (34). [3] Performance evaluation method (10) according to claim 1 or 2, characterized by that the measurement data (18) are recorded measurement data (26) and the provision (14) of the input data (16) includes access to the recorded measurement data (26). [4] Performance evaluation method (10) according to one of the preceding claims, characterized bythat the input data (16) comprise environmental object data (28) relating to the environmental objects (20), and the environmental object data (28) comprise at least a relative distance to the environmental object (20), a relative speed of the environmental object (20), and / or a dimension of the environmental object (20). [5] Performance evaluation method (10) according to claim 4, characterized by that the selected target object (38) is assigned target object data (40) which are derived from the surrounding object data (28) relating to the target object (38), and the object comparison (42) comprises a comparison of the target object data (40) with the reference target object data (32). [6] Performance evaluation method (10) according to one of the preceding claims, characterized by that the reference target object data (32) comprise at least a relative distance to the reference target object (34), a relative speed of the reference target object (34) and / or a dimension of the reference target object (34). [7] Method for adapting (46) a selection calculation model (12) for an automated distance control by changing the selection calculation model (12) depending on the evaluation parameter (44) calculated by a performance evaluation method (10) according to one of the preceding claims. [8] Method of adaptation (46) according to claim 7, characterized by that the selection calculation model (12) is changed if the evaluation parameter (44) deviates from a target parameter (48). [9] Method of adaptation (46) according to claim 7 or 8, characterized by that the selection calculation model (12) is left unchanged if the evaluation parameter (44) at least corresponds to the target parameter (48). [10] Method for identifying (50) at least one reference target object (34), comprising the steps Providing (52) measurement data (18) indicating environmental objects (20) of a vehicle environment (22), Creating (54) assignments (56) to the environmental objects (20) in the input data (16) by applying a first calculation model (58), identifying (62) at least one target object (38) depending on the assignments (56) by applying a second calculation model (66), Output (68) of the identified target object (38) as a reference target object (34).

Images