Method for processing a point cloud and corresponding motor vehicle

Abstract

Method (10) for processing a point cloud (P), characterized by the following features: - the point cloud ( P ) is filtered (11), - any critical regions within the point cloud (P) are determined (12), - the filtered point cloud ( F ) is supplemented with points that are contained in previously processed point clouds and lie within the critical regions, and - Any clusters within the augmented point cloud are tracked back over a certain period of time against the previously processed point clouds (13) in order to identify reliable clusters (14). Figure 1 represents an exemplary snapshot in this process: The method generates from K 6 By retrospectively filtering, clustering and tracking the last n point clouds, a point cloud enriched with points from critical regions is created. F 6 ' , which results in a different clustering outcome. The tracking vector is also modified due to tracking across multiple frames. In the proposed framework, each point cloud is processed retrospectively in this way, including those generated in the future. P 7 and P 8

Description

[0001] The present invention relates to a method for processing a point cloud. The present invention further relates to a corresponding motor vehicle, a corresponding computer program, and a corresponding storage medium. State of the art

[0002] The processing of point clouds, captured by sensors such as ultrasonic, radar, or lidar sensors, is a key element of modern driver assistance systems and automated driving functions in motor vehicles. These sensors provide detailed information about the vehicle's surroundings and enable object detection as well as the assessment of traffic and driving situations.

[0003] Radar sensors are particularly widespread due to their ability to provide reliable data under varying weather and lighting conditions. They generate point clouds that reflect the position and movement of objects in the environment. Modern radar systems have high sensitivity and can therefore detect even small objects. However, these characteristics also result in very dense point clouds containing a large amount of data, including potential measurement errors or interference.

[0004] To make the raw sensor data usable for further processing, preprocessing is required. A common approach is filtering the point cloud to remove unwanted data such as noise or measurement errors. Various filter criteria are applied, for example, based on signal strength, distance, or the plausibility of the measured values.

[0005] After filtering, the remaining points are typically grouped into clusters. This so-called clustering combines points with similar properties or spatial proximity, making it possible to identify individual objects within the point cloud. These clusters are then tracked over time to detect movement patterns and make predictions about the future positions of the objects.

[0006] Well-known methods for clustering and tracking employ algorithms such as k-means clustering, DBSCAN, or probabilistic models like Kalman filters. These methods allow for the analysis of dynamic scenarios and are essential for the development of functions such as autonomous driving, collision avoidance, and adaptive cruise control.

[0007] DE 10 2022 212 898 A1, for example, relates to a sensor for detecting objects on the roadway of a vehicle with a filter that filters all data points of a point cloud section by section with a maximum distance to the roadway plane.

[0008] CN 1 06 802 668 A, WO 2022 / 81 432 A1 and US 2021 / 0 141 092 A1 disclose similar devices.

[0009] BRÜHL, T. [et al.]: Making Radar Detections Safe for Autonomous Driving: A Review, in: VEHITS, 2024, pp. 299-310, URL: https: / / www.scitepress.org / Papers / 2024 / 126304 / 126304.pdf [accessed online on 24.10.2025] provides an overview of the current state of research on the topic of making radar detections usable for safe perception functions.

[0010] DE 10 2023 123 012 B3 describes a method and device for rapid collision detection in highly automated testing of driving processes. Disclosure of the invention

[0011] A known problem with object detection using point clouds is the occurrence of false positive and false negative classifications of individual points or point groups. Despite the application of filters, unwanted points can remain in the point cloud, leading to incorrect object identifications. Conversely, filtering can cause relevant points to be removed, especially if they do not meet the defined filter criteria. This is critical if the removed points are actual objects that could pose a danger to the vehicle.

[0012] Furthermore, real-time processing of point clouds presents a challenge. The large data volume requires efficient algorithms and significant computing resources. Therefore, it is important to find a balance between reducing the data volume and retaining relevant information.

[0013] Another problem is that conventional methods often only consider the current point cloud without taking into account the history or temporal evolution of the data. However, movement patterns or the appearance of new objects over time can provide important information for better assessing the behavior of other road users. Without retrospective analysis, this information goes unused, which impairs the accuracy and reliability of the object detection and tracking system.

[0014] Furthermore, existing filter criteria are often rigid and do not consider the criticality of different areas within the point cloud. A differentiated assessment that identifies and appropriately addresses regions with a higher hazard potential is lacking. As a result, critical areas cannot be analyzed separately, increasing the risk of overlooking relevant objects.

[0015] The described problem is solved by a method for processing a point cloud, a corresponding motor vehicle, a corresponding computer program and a corresponding storage medium according to the independent claims.

[0016] The described method significantly reduces the probability of relevant points, which represent potentially critical objects, being lost during the point cloud filtering process. By identifying critical regions within the point cloud and supplementing the filtered point cloud with points from previously processed point clouds that lie within these regions, a more comprehensive and accurate analysis is enabled.

[0017] Unlike previous state-of-the-art methods, where filtering often results in the removal of significant data points, the present method takes the data history into account. By retrospectively tracking clusters across multiple iterations, movement patterns can be identified and the reliability of object detection increased. This is particularly important for detecting objects that, while small or difficult to see, nevertheless pose a significant risk.

[0018] The adaptive definition of critical regions, for example by elliptical shapes extending in the direction of movement of the points and increasing in size as the data points age, contributes to improved object tracking. This takes into account that objects can move between measurements, leading to a more realistic assessment of their position.

[0019] Furthermore, the method enables differentiated filtering within critical regions by requiring higher signal strengths at the edges than in the center. This helps to efficiently reduce the amount of data without impairing the detection of relevant objects.

[0020] Further advantageous embodiments of the invention are specified in the dependent patent claims. Brief description of the drawings Fig. Figure 1 shows retrospective clustering and tracking within a point cloud P using a sliding window. Fig. Figure 2 shows the processing steps of the method according to the invention in their entirety. Fig. Figure 3 shows real data of an unfiltered point cloud according to an application example. Fig. Figure 4 shows those points of the point cloud with a positive radar cross section (RCS). Embodiments of the invention

[0021] The figures, when viewed together, illustrate the process of retrospective clustering and tracking within a point cloud. P using a sliding window comprising n = 5 steps. The point cloud P represents a multitude of data points Pt, which were detected by sensors such as radar, lidar or ultrasound and map the surroundings of a vehicle.

[0022] First, the point cloud is P filtered (11) to remove irrelevant or erroneous data points, thus creating a filtered point cloud F to generate. This conventional filtering process (11) is based on criteria that take into account, for example, the signal strength, the quality of the reflection, or other characteristic features of the points.

[0023] In addition, for each of the points Pt It assesses whether a collision of the automated vehicle with objects potentially represented by this point would lead to a critical incident. In this case, the critical region is determined using the x- and y-coordinates of the point (12) to analyze the past point clouds. Kt−1,…,Kt−n in this area to subject it to a kind of "post-analysis" through retrospective clustering and tracking (13). The filtered point cloud F This will be done in accordance with the regulation. Ki=Fk∩r(Pk) supplemented with an unfiltered component in the critical regions.

[0024] Fig. 1 represents an exemplary snapshot in this process: The method generates from K6 By retrospectively filtering, clustering and tracking the last n point clouds, a point cloud enriched with points from critical regions is created. F6', which results in a different clustering outcome. The tracking vector is also modified due to tracking across multiple frames. In the proposed framework, each point cloud is processed retrospectively in this way, including those generated in the future. P7 and P8.

[0025] Advantageously, each point in the point cloud is evaluated according to criteria such as TTC (including reaction time), position within the travel path or potential for movement into the travel path, object class, intended speed of the planner upon reaching the point, etc. A criticality score is determined from this. The critical region is created when a specific criticality threshold is exceeded. The accuracy of the criticality metric is crucial for accurately identifying critical points and forming critical regions from them.

[0026] Fig. Figure 3 shows an unfiltered radar point cloud with such critical regions (represented as circles). The circles contain all critical points from the past five point clouds. The older the point cloud, the larger the region and the more transparent the circle appears. Advantageously, the shape and lifetime of this region can depend on the criticality value.

[0027] Fig. Figure 4 shows the same radar point cloud filtered according to the criterion RCS > 0. It is noticeable that the point cloud in Fig. 4 some critical points of the Fig. 3 does not actually contain it. The point cloud for reprocessing would contain all points of the Fig. 4 and additionally all points in the circles of the Fig. 3 included.

[0028] The region is expediently described by an ellipse, taking the Doppler velocity into account, whose major axis corresponds to the presumed direction of motion of the corresponding points. Furthermore, the ellipses are preferably made larger as the recorded points age, in order to account for the possible movement of the objects since their initial detection.

[0029] Adaptive filtering within critical regions is advantageous. This involves placing higher demands on the signal strength of points at the region's edges than in the center. This leads to more efficient data processing, as less relevant points are excluded, while centrally located, potentially critical points with lower signal strength are still considered.

[0030] The number of steps n, which determines the depth of the retrospective sliding window, is five in this example. However, this number can be adjusted depending on the specific requirements and dynamic conditions to achieve an optimal balance between data volume and responsiveness. Reference symbol list 10 procedures 11. Filtering (the point cloud) 12. Determine (critical regions within the point cloud) 13 Clustering and tracking of clusters 14 (Secure) Clusters Point cloud Filtered point cloud current frame n Number of steps (included in the review)

Claims

[1] Method (10) for processing a point cloud (P) in an iterative process, with the following characteristics: - the point cloud (P) is filtered (11), - any critical regions within the point cloud (P) are determined (12) and - the filtered point cloud (F) is supplemented with points that are contained in previously processed point clouds and lie within the critical regions, characterized by the following characteristic: - Any clusters within the augmented point cloud are tracked retrospectively over a period of time determined by a number (n) of five steps of the process, looking back at the previously processed point clouds (13), - the regions are determined using the coordinates of the points located within them, - Each of the regions is defined by an ellipse with a principal axis that corresponds to a presumed direction of movement of the points located in the respective region, - the regions are determined based on a point within them that potentially represents an object with which a collision is imminent, and - The region is defined as larger the earlier the point contained within it was recorded. [2] motor vehicle, characterized by the following characteristics: - the vehicle has sensors for capturing a point cloud (P) and - the vehicle is equipped to display the point cloud (P) to process according to a method according to claim 1. [3] Computer program configured to perform all steps of a method according to claim 1. [4] Machine-readable storage medium with a computer program stored thereon according to claim 3.

Images