METHOD FOR SEGMENTING OBJECTS WITH SELF-MOTION

Abstract

Method for segmenting objects (1) with their own motion using an ego-vehicle (2) for execution in a neural network, wherein the ego-vehicle (2) comprises at least the following components: - at least one computer (3) with a processor for processing digital data and with a data storage for holding digital data for the processor; and - at least one measuring sensor (4) for sensing the current vehicle environment (5) of the ego-vehicle (2), wherein the method comprises at least the following continuously repeating steps: a. using the measuring sensor (4), generating a plurality of measurement data (6, 7) by sensing objects (1) in a current vehicle environment (5) of the ego-vehicle (2) and providing the measurement data (6, 7) from the sensed vehicle environment (5) to the computer (3); and b.By means of the processor, to generate segmented motion data (8), the current measurement data (6) are enriched with a number of measurement data (7) generated in previous sequences and position-corrected to the current measurement data (6) in a position group (9), wherein the number of sequences can be adjusted as required for measurement data that are known to be highly noisy, wherein the number is increased for measurement data that are known to be highly noisy compared to less noisy measurement data, wherein the following are used for enrichment: - previous measurement data (7) which are used for the current task in step b.relevant, and- previous measurement data (7) which are learned as relevant by the neural network, wherein the current measurement data (6) and a number of previous measurement data (7) are superimposed by adding a presumed and / or measured change in position of the ego vehicle (2) to the previous data and / or subtracting it from the current measurement data (6), wherein only those measurement data (7) of the previous sequences are used for enrichment which have the identical and / or same speed and / or, in the case of a radar sensor, the same radar cross-section.

Description

The invention relates to a method for segmenting objects with their own motion using an ego-mobile vehicle. The invention further relates to a motor vehicle and a computer program, as well as a computer program product for executing such a method. Computer-aided methods and the measurement systems used to segment moving objects, such as other vehicles and pedestrians, are known in a wide variety of forms. Knowledge of moving objects in the vicinity of an autonomous vehicle (ego-vehicle) is essential for safe and reliable autonomous navigation. The interpretation of LiDAR and camera data yields outstanding results. However, this generally requires the accumulation and processing of temporal data sequences to extract motion information.In contrast, radar sensors, which are already installed in most newer vehicles, can overcome this limitation because they directly provide the Doppler velocity of the detections and thus contain the instantaneous motion information in a single measurement. With radar sensors, the measurement data (usually point clouds) is generated by comparing a transmitted signal and a corresponding echo signal, so that the shifted superposition (Doppler effect) leads to detection. A significant disadvantage, however, is that radar scans are highly noisy due to multipath propagation, self-motion, and sensor noise. These noisy measurements often lead to false readings and thus reduce the reliability of the detection signals. Compared to LiDAR data, radar sensors yield significantly fewer measurement points. DE 11 2010 004 163 B4 describes a cross-traffic collision warning system. DE 10 2018 220 114 A1 describes a device and a method for detecting a sensor blockage of an environmental sensor. From DE 10 2018 205 879 A1 a method, a device and a computer-readable storage medium with instructions for processing sensor data are known. DE 10 2017 206 396 A1 describes a control procedure, a control unit, an environment detection system and a mobile and / or stationary device. Autonomous (or self-)driving vehicles must be able to distinguish between moving and stationary objects to navigate safely in a dynamic, real-world driving environment. To enable redundancy and overcome the limitations of individual sensors, the sensor systems of autonomous vehicles are diverse and include cameras, LiDARs, and radars. The widely researched camera and LiDAR sensors use temporal sequences of input data to detect moving objects. Based on this, the present invention aims to overcome, at least partially, the disadvantages known from the prior art. The features of the invention are defined in the independent claims, for which advantageous embodiments are shown in the dependent claims. The features of the claims can be combined in any technically meaningful way, whereby the explanations in the following description and features from the figures, which comprise supplementary embodiments of the invention, can also be used. The invention relates to a method for segmenting objects with their own motion using an ego-vehicle for execution in a neural network, wherein the ego-vehicle comprises at least the following components: - at least one computer with a processor for processing digital data and with a data storage device for holding digital data for the processor; and - at least one measuring sensor for capturing the current vehicle environment of the ego-vehicle, wherein the method comprises at least the following continuously repeating steps: a. using the measuring sensor, generating a plurality of measurement data by capturing objects in a current vehicle environment of the ego-vehicle and providing the measurement data from the captured vehicle environment to the computer; and b.Using the processor, to generate segmented motion data, the current measurement data is enriched with a number of measurement data generated in previous sequences and position-corrected to the current measurement data in a position group, whereby the following are used for enrichment: - previous measurement data which are relevant for the current task in step b., and - previous measurement data which are learned as relevant by the neural network. Unless explicitly stated otherwise, ordinal numbers used in the preceding and following descriptions serve solely for unambiguous differentiation and do not indicate any order or ranking of the components referred to. An ordinal number greater than one does not necessarily imply the presence of another such component. It should be noted in advance that the procedure must be carried out with limited available resources, such as an on-board computer of a motor vehicle, and in real time, for example integrated into the frequency of a control loop to assist human-controlled driving or autonomous driving. The approach here is to address the problem of distinguishing between moving and stationary objects, for example, differentiating between parked and moving vehicles. This improves the understanding of the current scene. Furthermore, it is important to be able to recognize (i.e., differentiate) multiple moving objects as individual moving objects. This means it is crucial to segment the detected moving measurement data into different objects. This approach proposes utilizing the sensor information from the measuring sensor through machine learning to enable reliable segmentation of moving objects in the vehicle's environment. Previously acquired measurement data is used to improve the differentiation of objects located close together in the data, for example, a moving vehicle driving closely alongside another moving vehicle, or, as a redundant error exclusion method, next to a parked vehicle. Unlike previously known methods, this approach does not use all unprocessed measurement data. Instead, it utilizes both position-corrected (previous) measurement data and only those previous measurement data for which corresponding (current) measurement data exists in the current measurement.As a clear example, if a vehicle leaves the space in front of the ego-vehicle (for instance, by turning), the previously recorded measurement data for that vehicle is no longer used. This significantly reduces the amount of data to be processed, and therefore processor power and / or processing time. The advantage is that the measurement data is not processed by the entire neural network. It should be noted that while the amount of data is not necessarily less, less information is combined when the vehicle is moving. For example, a computer with a processor and data storage is conventionally designed, and reference is made to the following description, especially regarding the figures. In step a., the objects are recorded in their relative position and (at least via the sequence of their) movement to the ego vehicle. The ego vehicle is the vehicle with which the procedure is carried out.In step b, the measurement data generated in a sequence of repetitions is enriched. This enrichment is relevant for the current task, i.e., the motion data to be segmented according to the current repetition. In one embodiment, the relevance of the data is predetermined. Alternatively or additionally, the relevance is learned (possibly exclusively) by the neural network. In a simple example, the current measurement data and a number of previous measurement data points are superimposed, for example, by adding a presumed (or learned) and / or measured change in the position of the ego vehicle to the previous data or subtracting it from the current measurement data. In a relatively unchanging scene (for example, following at a constant distance), the measurement data would theoretically be identical from one repetition to the next.However, if an object in a scene undergoes a relative movement to the ego vehicle (for example, a vehicle turning), this measurement data will change in the sequence, and information about the relative movement of the object can be derived from it. The number of sequences compared for enrichment is preferably a predetermined static number. Alternatively, the number can be adjusted as needed; for example, the number can be increased for measurement data that is known to be very noisy (e.g., due to adverse weather conditions such as splashing water) compared to less noisy measurement data. For small amounts of data, it is proposed (as mentioned above) that not all previously generated measurement data be carried over to the calculation processes. Instead, only those measurement data points that correlate with the current measurement data should be included in the position group. One possibility is to use only those measurement data from the preceding sequences for enrichment which (position-corrected) have identical or similar coordinates and / or other identical properties (for example, the same speed and / or, in the case of a radar sensor, the same radar cross-section). Due to the short time intervals between the sequences, the measurement field used for enrichment is considered to be almost static. Therefore, any change is caused by a measurement error or relative motion. However, a measurement error will propagate identically across a multiple of sequences with a vanishingly small probability. Thus, such a measurement error has an intrinsically low weighting. This approach is referred to here as Sequential Attentive Feature Encoding [SAFE]. Therefore, if no corresponding value can be found in all previous sequences for a given value in the relative motion data (position-corrected), then movement has almost certainly occurred. Because this process is repeated repeatedly, it theoretically shows that an object moving relatively (over a given period) receives no data enrichment in each iteration. However, it has been shown that enrichment does occur at low relative velocities. But the network learns that if a displacement occurs, the object must have moved. These values ​​can then be linked, and relative motion can be resolved into space and time, i.e., detected.Together with other information, such as the location, the velocity recorded per sequence (for example, in the case of a radar sensor using the Doppler effect) and / or the radar cross-section in the case of a radar sensor, a very high degree of certainty arises that it is a single object, which can therefore be segmented. According to one aspect, the invention further comprises a measuring system for an ego-vehicle for detecting a vehicle environment, comprising at least the following components: - at least one measuring sensor for detecting objects in the vehicle environment of a ego-vehicle in question; - a computer with a processor for issuing recommendations for accelerating and / or controlling a drive motor and a deceleration device of the ego-vehicle; and - preferably at least one self-motion sensor for detecting the self-motion of the ego-vehicle, wherein the computer is configured, using data from the measuring sensor, and preferably from the self-motion sensor, to perform the above-mentioned and subsequently further supplemented method for segmenting objects with self-motion. In a further advantageous embodiment of the method, it is proposed that the measurement points are generated by a radar sensor of a motor vehicle for detecting objects in the current vehicle environment and comprise two spatial position coordinates, Doppler-based velocity data and / or radar cross-sections. The approach here is to address the problem of distinguishing between moving and stationary objects, for example, differentiating between parked and moving vehicles. This improves the understanding of the current scene. It is proposed here to leverage the previous detections of the radar sensor using the previously described learning-based approaches to enable reliable detection of moving objects in the vehicle's environment. Furthermore, the detected radar cross-section of the relevant detection point, which depends on the material properties and structure of the detection element, supports the differentiation of objects located close together in the detection data, such as a moving vehicle driving closely alongside another moving vehicle, or, as a redundant error exclusion, next to a parked vehicle.Because the nature of these two additional detection values ​​differs from that of the position coordinates, they are also weighted differently. In one embodiment, Doppler-based velocity data is also used. Each detection point of the radar sensor comprises two position coordinates in at least two-dimensional space and a velocity determined via the Doppler effect in a single scan. In one embodiment, this intrinsic velocity information is used as supplementary information for the learning-based approach. It should be noted that the use of radar detection data is not strictly necessary to perform the previously described procedure. However, when using a radar sensor, the previously generated relative motion data is preferably also detection data from the (same) radar sensor. Due to the additional intrinsically contained information, namely the Doppler velocity and / or the radar cross-section, further possibilities are provided for learning-based recognition to identify related objects. In an advantageous embodiment of the method, it is further proposed that the measurement data generated in step a. comprise individual points of a point cloud. It is proposed here that the measurement data be provided as points, each containing spatial (e.g., purely two-dimensional) information, and preferably, in the case of a radar sensor, also (Doppler) velocity information and / or material information via a radar cross-section. The majority of the points are arranged as a point cloud within a sensor field of a vehicle environment monitored by the radar sensors. Only detected signals are recorded as points in this point cloud, preferably only sufficiently reliable signals. Thus, the point cloud constitutes a multidimensional matrix for further processing. In a further advantageous embodiment of the method, it is proposed that ego position data of the ego vehicle, determined by a self-motion sensor, preferably a geolocation sensor, be used to correct the position of the preceding measurement data, preferably considering only longitudinal movement relative to the main measurement direction of the ego vehicle. The self-motion sensor is, for example, a geolocation sensor for detecting an absolute position or is simply configured to detect its own movement at a time without a fixed reference point. The geolocation sensor is, for example, a GPS sensor.: Global Positioning System] or comparable satellite-based systems (for motion detection with time measurement between two consecutive measuring points), a hodometer (detection of wheel rotations and steering angle of the Ego vehicle) and / or an accelerometer to detect longitudinal and lateral acceleration and to obtain a current movement, i.e., directionally correct speed, by means of simple temporal integration. Using the geolocation sensor, an (at least approximate) absolute position is calculated by adjusting the recorded measurement data with the Ego vehicle's own position, for example by adding the relative position vectors from the radar sensor's recording to the absolute geolocation data of the Ego vehicle's geolocation sensor. In an advantageous embodiment, position correction is performed solely using the detected longitudinal movement. This is because, when driving an ego-vehicle and when encountering objects (e.g., other vehicles) in the vehicle's vicinity, the longitudinal movement yields significantly larger values ​​than the lateral movement, especially over a (small) number of sequences, in a stable driving condition. The latter is practically negligible over short periods, and it is therefore advantageous to disregard this lateral movement during data processing. This reduces the computational effort for position correction by at least half, with negligible inaccuracy. According to a further aspect, a motor vehicle is proposed comprising at least the following components: - at least one drive wheel; - at least one drive motor and a deceleration device for accelerating the motor vehicle by means of the at least one drive wheel; - at least one measuring sensor, preferably a radar sensor, for detecting the vehicle's surroundings; and - an on-board computer for issuing recommendations for accelerating and / or controlling the drive motor and the deceleration device, wherein the on-board computer is configured to receive data from an external computer, which is configured to execute a method according to an embodiment as described above, and / or itself to execute at least part of the method for segmenting objects with their own motion. The external computer is located externally from the on-board computer and within the motor vehicle. The motor vehicle, preferably a semi-autonomous or fully autonomous vehicle, is the Ego vehicle for the method of segmenting objects with their own motion. It can be driven by means of one or more drive wheels, whereby an on-board computer can intervene in the driving process by means of a recommendation to a driver and / or by means of direct control. The drive unit is, for example, an internal combustion engine and / or an electric traction motor. The deceleration device is a braking system, a recuperation system, and / or a parking brake. Reference is made to the preceding description with regard to the method, at least as one possible embodiment, insofar as this concerns the motor vehicle and its components. For example, a computer with a processor and data storage is conventionally designed, and reference is made to the following description, especially regarding the figures. The measuring sensor is designed to detect objects in the vehicle's surroundings. For example, the measuring sensor may include camera sensors, LiDAR sensors, and / or a radar sensor. A radar sensor is configured to transmit a signal and receive an echo signal, thereby registering the Doppler effect and thus a reflective object moving relative to the vehicle. Typically, a measuring sensor in a vehicle comprises multiple sensor elements, such as LiDAR and / or radar sensors, which together generate measurement (or detection) data from the vehicle's surroundings (e.g., radar field) and provide this data to the computer. The self-motion sensor is, for example, a geolocation sensor for recording an absolute position or is simply configured to record the user's own movement at a point in time without a fixed reference point. The geolocation sensor could be, for example, a GPS sensor (Global Positioning System) or comparable satellite-based systems (for motion detection with time measurement between two consecutive measurement points), a hodometer (recording wheel rotations and steering angle of the ego vehicle), and / or an accelerometer for recording longitudinal and lateral acceleration and, through simple temporal integration, obtaining a current movement, i.e., directionally correct speed. According to a further aspect, a computer program is proposed, comprising a computer program code, wherein the computer program code is executable on at least one computer such that the at least one computer is induced to execute the method according to an embodiment as described above, wherein at least one unit of the computer is: - arranged in an on-board computer of a motor vehicle, preferably according to an embodiment as described above; and / or - configured to communicate with a cloud, on which preferably at least a part of the computer program code is provided. The method described here for segmenting objects with intrinsic motion is implemented in a computer according to this embodiment. The computer-implemented method is stored as computer program code, wherein the computer program code, when executed on a computer, for example comprising a data storage device and a processor, causes the computer to execute the method according to an embodiment as described above. The computer-implemented method is realized, for example, by a computer program, wherein the computer program comprises the computer program code, and wherein the computer program code, when executed on a computer, causes the computer to execute the method according to an embodiment as described above. Computer program code is synonymously defined as one or more instructions or commands that cause a computer to perform a series of operations, which, for example, represent an algorithm and / or other processing methods. The computer program is preferably executable, either partially or completely, on an onboard computer and / or on a server or server unit of a so-called cloud, a handheld device (e.g., a smartphone), and / or on at least one unit of the computer. The term server or server unit here refers to a computer that provides data and / or operational services or services for one or more other computer-based devices or computers, thus forming the cloud.Preferably, the computer on board a motor vehicle is able to execute the computer program independently without requiring additional resources, thus providing the data as input values ​​for downstream control processes in a short and (transmission-) secure manner for real-time operations, for example for collision avoidance of the Ego motor vehicle with another vehicle and / or for increasing the efficiency of the Ego motor vehicle's driving behavior in heavy traffic. The terms "cloud" and "computer" are used here synonymously with devices known from the prior art. A computer therefore comprises one or more general-purpose processors (CPUs) or microprocessors, RISC processors, GPUs, and / or DSPs. The computer also includes additional elements such as memory interfaces or communication interfaces. Alternatively or additionally, the terms refer to a device capable of executing a provided or integrated program, preferably using a standardized programming language (such as C++, JavaScript, or Python), and / or controlling and / or accessing data storage devices and / or other devices such as input and output interfaces.The term "computer" also refers to a multitude of processors or a multitude of (sub)computers that are interconnected via physical connections and / or other means of communication and may share one or more other resources, such as data storage. Data storage can be, for example, a hard disk drive (HDD) or non-volatile solid-state memory, such as ROM or flash memory (Flash EEPROM). Storage often comprises multiple individual physical units or is distributed across a multitude of separate devices, allowing access via data communication, such as a package data service. The latter is a decentralized solution, where storage and processors are used across multiple separate computers instead of, or in addition to, a single central server. According to a further aspect, a computer program product is proposed on which computer program code is stored, wherein the computer program code is executable on at least one computer in such a way that the at least one computer is caused to execute the method according to an embodiment as described above, wherein at least one unit of the computer is: - arranged in an on-board computer of a motor vehicle, preferably according to an embodiment as described above; and / or - set up for communication with a cloud, on which preferably at least part of the computer program code is provided. As a computer program product, comprising the computer program code described above, it is stored, for example, on a medium such as RAM, ROM, an SD card, a memory card, a flash memory card, or a disc, or on a server and can be downloaded. Once the computer program is made readable via a read unit, such as a drive and / or an installation, the contained computer program code and the method for segmenting objects with their own movement by a computer or in communication with a plurality of server units, for example, as described above, can be executed. The invention described above is explained in detail below against the relevant technical background with reference to the accompanying drawings, which show preferred embodiments. The invention is in no way limited by the purely schematic drawings, it should be noted that the drawings are not dimensionally accurate and are not suitable for defining size relationships. Figure 1 shows detection data as a point cloud, as well as two processing stages of the detection data; Figure 2 shows a diagram of Sequential Attentive Feature Encoding; Figure 3 shows a module for Sequential Attentive Feature Encoding; and Figure 4 shows an ego-vehicle with a radar sensor for detecting objects with their own motion. Figure 1 shows a schematic representation of detection data (measurement data) detected (acquired) by a radar sensor as detection points (measurement points 10) of a point cloud 11 in the topmost figure, and two processing stages based on the detection data in the figures below. The measurement points 10 show the spatial location of a detected reflection (from a material reflecting radar radiation), as well as the size of the point, the size of the detected (radar) cross-section. The arrows indicate detected relative movements and their direction from the detected points. Thus, (current as well as previous) measurement data 6, 7 are captured. In the figure below (i.e., the middle figure), circles and quadrilaterals are shown, which, for example, have been corrected for the relative movement of the Ego vehicle 2.The circles show stationary objects 1 or their respective reflections, and the quadrilaterals show objects 1 that are moving, for example, at least one other vehicle. In the bottom image, the processing of the middle image is further refined by grouping the measurement data 6, 7, which can be assigned to a coherent unit, into position groups 9 with points of the same type, namely hexagons, triangles, stars, and quadrilaterals, as well as triangles and a hexagon again (but representing separate position groups 9). These points are now segmented motion data 8, since they have already been optionally cleaned up by any movement of the ego vehicle 2. This is achieved by enriching the points shown here with points recorded at a previous time. Figure 2 shows a diagram of a Sequential Attentive Feature Encoding [SAFE 22] and its input values ​​(point clouds 11). The point cloud 11 above (for example, of a vehicle environment 5, see Figure 1) of a scan, which forms a direct input value for the SAFE module 22, shows a current set of detection data (measurement data) from a radar sensor 4. Below the current point cloud 11, in the leftmost column, the multitude of previously acquired point clouds 11 (previous scans) with previously acquired relative motion data are indicated. In this example, these are combined with a timestamp 24 corresponding to their respective acquisition times to form a point cloud 11 from the previous measurement data 7, resulting in superimposed relative motion data 23.The current point cloud 11 is then enriched with the enriched point cloud 11 as an additional input value for the SAFE module 22. This raw data, enriched in the SAFE module 22, is subsequently processed in a neural network using learned weights. Figure 3 shows a module for Sequential Attentive Feature Encoding [SAFE 22]. Temporal information is important for improving the interpretation of the scene in sparsely populated and noisy (e.g., radar) point clouds 11. In particular, temporal relationships help to reliably identify actors when segmenting moving objects 1. Furthermore, temporal dependencies support the differentiation between moving and static measurements, including noise, which, due to its changing appearance, can be directly defined in the temporal domain. In contrast to other approaches where multiple point clouds 11 are routed through the entire network, we propose the Sequential Attentive Feature Encoding module [SAFE 22] to efficiently enrich the features of a single point cloud 11 with temporal information. Therefore, only the current scan is routed through the entire network, while previous scans are processed only within the SAFE module 22.The aim is to adaptively combine the temporal information from previous scans with the features of the current point cloud 11, thereby reducing the computational effort while still obtaining important information. The module uses the current (or a plurality of) and N-1 previous scans of a sequence. First, we process the current and aggregated previous scans with a KPConv layer to extract higher-dimensional features. The KPConv layer is a deformable convolution that is learned so that the kernel points adapt to the local geometry. The points in the KPConv layer are processed using a convolution operation. KPConv coding 32 is known to be important for single-scan inputs in transformer-based networks. Because we use the self-attention mechanism, we extend KPConv coding 32 to the previous scans. The temporal coding processes the updated features and the corresponding positional information of the current and previous scans.The current point cloud 11 is the starting point cloud, and the idea is to enrich the information of each individual measurement point 10 to improve the accuracy of the downstream task. To adaptively combine the information, we extend the point transformer layer to an intra-attention module. We first encode the features of the previous scans XN (previous point coordinates 39) as value vectors 31v and query vectors 30q, and the features of the current scans XT (current point coordinates 38) as key vectors 29k, as follows: The matrices Wk∈ ℝD×D(key weighting matrix 26), Wq∈ ℝD×D(query weighting matrix 27) and Wv∈ ℝD×D(value weighting matrix 28) are the weights of multilayer perceptrons 33. For the relative position encoding 35, we determine the k-nearest neighbors [kNN algorithm 37] within the point cloud 11 PN for the point cloud 11 of the current time step PT(current point coordinates 38). Within the resulting local areas, we determine the relative position 34 between the two point clouds 11 (i.e., the current position data 40 and the previous position data 41), where pj∈ PN and pi∈ PT yields: To calculate the attention weights 42, we subtract the query vectors 30 from the key vectors 29 and add the relative position encoding 35 of the two point clouds 11 to include fine-grained position information in the attention weights 42.We use the Softmax function 25 to determine the final attention weights 42 as follows:. We process the attention weight 42 using a multilayer perceptron 33 with two batch-norm layers and two fully connected layers. Additionally, we add the values ​​and the relative position coding 35. To derive the weighted features y, we calculate the sum of the element-wise multiplications during aggregation 36: The weighted features and the features of our current scan are finally linked together in concatenation step 20. We enrich the information from the current scan with the inter-attentions from previous scans. We process the aggregated information using various backbone networks. This enriched temporal feature encoding enhances the information and improves overall performance. Linking is crucial so that the neural network can adaptively incorporate information from previous scans. It should be noted that the SAFE module 22 is model-agnostic and applicable to different backbones (i.e., different neural networks). Figure 4 shows a (self-driving) motor vehicle 2 with a radar sensor 4 for detecting moving objects 1. The motor vehicle 2 comprises a drive motor 16, by means of which the motor vehicle 2 can be accelerated (positively) via the drive wheel 15, and a deceleration device 17 (for example, a brake), by means of which the motor vehicle 2 can be accelerated (negatively). The motor vehicle 2 also includes an on-board computer 18, which here (purely optionally) is configured as a computer 3 for executing the method for detecting moving objects 1 in the vehicle environment 5 of the motor vehicle 2. These objects 1 are shown here purely as examples of vehicles.The Ego vehicle 2 comprises a self-motion sensor 12, for example an accelerometer and / or a geolocation sensor 13, for evaluating the Ego position data 14 or self-motion data, which are received via an antenna 21 for a satellite-based positioning system 19, so that self-motion of the Ego vehicle 2 can be detected. Furthermore, a radar sensor is provided, with which objects 1 with a relative velocity resulting from the Doppler effect between the transmitted signal and its echo signal in the vehicle environment 5 of the Ego vehicle 2 can be detected. The invention relates to a method for segmenting objects with their own motion using an ego-vehicle, wherein the method comprises at least the following steps, which are repeated in a continuous sequence: a. using the radar sensor, generating a plurality of measurement data by detecting objects in the current vehicle environment of the ego-vehicle and providing the measurement data from the detected vehicle environment to the computer; and b. using the self-motion sensor, generating relative motion data by detecting the ego-vehicle's own motion and providing the relative motion data from the detected motion to the computer. The method is characterized in particular by the fact that in step c.To generate segmented motion data, the current relative motion data is enriched in a position group with a number of relative motion data generated in previous sequences and corrected to the current relative motion data, whereby only those previous relative motion data which correspond to the current relative motion data are used for enrichment. The method proposed here allows objects in the vehicle environment of an autonomously driving vehicle to be segmented correctly with a high degree of certainty. Reference symbol list 1 Object 2 Ego vehicle 3 Computer 4 Measuring sensor 5 Vehicle environment 6 Current measurement data 7 Previous measurement data 8 Segmented motion data 9 Position group 10 Measurement points 11 Point cloud 12 Self-motion sensor 13 Geolocation sensor 14 Ego position data 15 Drive wheel 16 Drive motor 17 Deceleration device 18 On-board computer 19 Positioning system 20 Concatenation 21 Antenna 22 SAFE (module) 23 Superimposed relative motion data 24 Timestamp 25 Softmax function 26 Key weighting matrix 27 Query weighting matrix 28 Value weighting matrix 29 Key vector 30 Query vector 31 Value vector 32 KPConv coding 33 Multilayer perceptron [MLP] 34 Relative position 35 Position coding 36 Aggregation 37 kNN algorithm 38 current point coordinates 39 previous point coordinates 40 current position data 41 previous position data 42 attention weight

Claims

Method for segmenting objects (1) with their own motion using an ego-vehicle (2) for execution in a neural network, wherein the ego-vehicle (2) comprises at least the following components: - at least one computer (3) with a processor for processing digital data and with a data storage for holding digital data for the processor; and - at least one measuring sensor (4) for sensing the current vehicle environment (5) of the ego-vehicle (2), wherein the method comprises at least the following continuously repeating steps: a. using the measuring sensor (4), generating a plurality of measurement data (6, 7) by sensing objects (1) in a current vehicle environment (5) of the ego-vehicle (2) and providing the measurement data (6, 7) from the sensed vehicle environment (5) to the computer (3); and b.By means of the processor, to generate segmented motion data (8), the current measurement data (6) are enriched with a number of measurement data (7) generated in previous sequences and position-corrected to the current measurement data (6) in a position group (9), wherein the number of sequences can be adjusted as required for measurement data that are known to be highly noisy, wherein the number is increased for measurement data that are known to be highly noisy compared to less noisy measurement data, wherein the following are used for enrichment: - previous measurement data (7) which are used for the current task in step b.relevant, and- previous measurement data (7) which are learned as relevant by the neural network, wherein the current measurement data (6) and a number of previous measurement data (7) are superimposed by adding a presumed and / or measured change in position of the ego vehicle (2) to the previous data and / or subtracting it from the current measurement data (6), wherein only those measurement data (7) of the previous sequences are used for enrichment which have the identical and / or same speed and / or, in the case of a radar sensor, the same radar cross-section. Method according to claim 1, wherein the measuring points (10) are generated by a radar sensor of a motor vehicle (2) for detecting objects (1) in the current vehicle environment (5) and comprise two spatial position coordinates, Doppler-based velocity data and / or radar cross-sections. Method according to claim 1 or claim 2, wherein the measurement data (6,7) generated in step a. comprise individual points of a point cloud (11). Method according to one of the preceding claims, wherein Ego position data (14) of the Ego vehicle (2) determined by means of a self-motion sensor (12), preferably a geolocation sensor (13), are used to correct the position of the preceding measurement data (7), wherein preferably only a longitudinal movement relative to the main measurement direction of the Ego vehicle (2) is taken into account. Motor vehicle (2), comprising at least the following components: - at least one drive wheel (15); - at least one drive motor (16) and a deceleration device (17) for accelerating the motor vehicle (2) by means of the at least one drive wheel (15); and - at least one measuring sensor (4), preferably a radar sensor, for detecting the vehicle environment (5) of the motor vehicle (2); and - an on-board computer (18) for issuing recommendations for accelerating and / or controlling the drive motor (16) and the deceleration device (17), wherein the on-board computer (18) is configured to receive data from an external computer (3), which is configured to execute a method according to one of the preceding claims, and / or is itself configured to execute at least part of the method for segmenting objects (1) with intrinsic motion, wherein the external computer (3) is located externally from the on-board computer (18) and in the motor vehicle (2). Computer program comprising computer program code, wherein the computer program code is executable on at least one computer (3) such that the at least one computer (3) is induced to execute the method according to any one of claims 1 to 4, wherein at least one unit of the computer (3) is: - arranged in an on-board computer (18) of a motor vehicle (2), preferably according to claim 5; and / or - configured for communication with a cloud, on which preferably at least a part of the computer program code is provided. A computer program product on which a computer program code is stored, wherein the computer program code is executable on at least one computer (3) such that the at least one computer (3) is caused to execute the method according to any one of claims 1 to 4, wherein at least one unit of the computer (3) is: - arranged in an on-board computer (18) of a motor vehicle (2), preferably according to claim 5; and / or - configured for communication with a cloud on which preferably at least a part of the computer program code is provided.

Images