Precision evaluation method for updated layer of point cloud map

Abstract

The invention relates to a precision evaluation method for an updated layer of a point cloud map, and the method comprises the steps: obtaining a high-precision point cloud base map and the updated layer of the point cloud map, extracting semantic information from the high-precision point cloud base map and the updated layer, carrying out the instantiation operation of a static layer and the updated layer of the point cloud base map, and carrying out the precision evaluation of the updated layer of the point cloud map through employing the extracted semantic information. The method comprises the steps of finding a matched instance between a high-precision point cloud base map and an updated map layer by comparing features of geographic elements, performing point cloud registration of the updated map layer and a static map layer based on instance matching information, accurately aligning point cloud data in the updated map layer to the high-precision point cloud base map, and evaluating the precision of the updated map layer after point cloud registration is completed. According to the method, the semantic information and instance matching of the static layer and the updated layer are combined, the dynamic environment change is effectively adapted, the accuracy and reliability of quality inspection are improved, and the overall efficiency and accuracy of quality inspection are improved by utilizing the associated information between the updated layer and the base layer.

Description

Technical Field

[0001] The present invention belongs to the field of high-precision maps and autonomous driving technology, and specifically relates to a method for evaluating the accuracy of a point cloud map update layer. Background Art

[0002] In complex and dynamic autonomous driving operation scenarios such as ports and mines, high-precision point cloud updates are one of the key technologies. They can not only reflect changes in dynamic operation scenarios, but also enhance the vehicle's understanding of the environment, which is crucial to improving the safety and reliability of autonomous driving systems.

[0003] High-precision point cloud maps play a crucial role in autonomous driving, especially in complex and dynamic operating scenarios like ports and mines, which often involve numerous moving objects, irregular terrain, and changeable weather conditions. This places higher demands on the positioning, perception, and decision-making capabilities of autonomous driving systems. High-precision point cloud updates can reflect these dynamic changes in real time, enhancing the vehicle's understanding of the environment and ultimately improving the safety and reliability of autonomous driving systems.

[0004] However, existing methods for assessing the accuracy of high-precision point cloud maps rely primarily on measuring the geometric dimensions of key map elements, such as the thickness of the ground, walls, light poles, and signage. This method performs well in static or relatively static environments, but has poor adaptability in dynamic environments. Because point cloud data can contain noise, missing data, or inaccurate matching, these errors can accumulate in dynamic environments, severely impacting autonomous driving modules like positioning and perception.

[0005] Furthermore, as the real-world environment changes over time, point cloud maps require regular updates. To ensure the quality of the updated maps, quality inspection of the updated layers is necessary. However, existing methods are not specifically designed for evaluating the accuracy of updated layers. In autonomous driving systems, point cloud maps are continuously updated to reflect environmental changes. However, existing accuracy assessment methods often overlook the correlation between the updated layers and the basemap layer, which can introduce new errors during the update process or miss important change information.

[0006] Therefore, to further improve the safety and reliability of autonomous driving systems in complex and dynamic environments, it is necessary to develop more advanced, high-precision point cloud map accuracy assessment methods. These methods must fully account for issues such as noise, missing points, and matching accuracy in point cloud data, and be able to perform specialized accuracy assessments on update layers. Furthermore, they must fully utilize the correlation between update layers and basemap layers to ensure the accuracy and real-time nature of point cloud maps. Summary of the Invention

[0007] The purpose of the present invention is to address the problems existing in the prior art and propose a method for evaluating the accuracy of point cloud map update layers. By utilizing the semantic information in the static layer and the update layer in the high-precision point cloud base map, instance matching information is obtained, and point cloud registration of the update layer and the static layer is performed, providing effective reference information for the accuracy evaluation of the point cloud map update layer.

[0008] In order to achieve the above objectives, the present invention provides a method for evaluating the accuracy of a point cloud map update layer, comprising the following steps: Step 1: Obtain a high-precision point cloud base map and an updated layer of the point cloud map. The high-precision point cloud base map is a static point cloud dataset containing detailed geographic information. Step 2: Extract semantic information from the high-precision point cloud base map and update layer and instantiate the point cloud base map static layer and update layer. The semantic information includes the classification and location information of geographic elements, including but not limited to buildings, roads, vegetation, and traffic signs. Step 3: Using the extracted semantic information, find matching instances between the high-precision point cloud base map and the updated layer by comparing the features of geographic elements; Step 4: Based on the instance matching information, perform point cloud registration between the updated layer and the static layer, and accurately align the point cloud data in the updated layer to the high-precision point cloud base map; Step 5: After completing the point cloud registration, evaluate the accuracy of the updated layer.

[0009] The present invention evaluates the accuracy of updated layers by comparing semantic information in the updated layer with that in the static layer. This method performs static layer extraction on the point cloud basemap layer, label point cloud extraction on the point cloud update layer, and instantiation operations on the label point clouds in both the static and update layers. By completing static layer extraction and instantiation in one go, along with fully automatic label extraction for the update layer, a highly automated and reusable quality inspection process is achieved. Furthermore, the high-precision quality inspection method based on instance matching effectively reduces the interference of dynamic objects on subsequent point cloud registration. The core of this approach lies in its high degree of automation and efficient reuse. First, static layer extraction and instantiation can be completed in one go. This means that in the initial stage, the underlying layer undergoes in-depth analysis and processing to extract point cloud data for static objects, and then instantiation is performed to separately label different object instances. This process combines automated extraction with manual verification to ensure extraction accuracy. Once completed, this static layer and instantiation information can be reused in subsequent quality inspections of updated map layers, significantly improving efficiency.

[0010] Secondly, for update layers, fully automatic label extraction is now possible. This step relies on computer vision and machine learning technologies to automatically identify and extract point cloud data from update layers without manual intervention, which not only improves quality inspection efficiency but also reduces the possibility of errors.

[0011] Finally, a high-precision quality inspection method based on instance matching leverages semantic and geometric information to precisely match the point cloud of the static base map and the updated layer, effectively reducing the interference of dynamic objects on subsequent point cloud registration. This matching method not only improves quality inspection accuracy but also ensures the stability and consistency of the map layer.

[0012] Therefore, the present invention adopts a highly automated and efficiently reused quality inspection process combined with a high-precision quality inspection method based on instance matching, which greatly improves the quality inspection efficiency and accuracy.

[0013] The present invention further adopts the following technical solution: In step 1, the high-precision point cloud base map is a complete scene point cloud map layer, created for the first time for a specific work scenario and undergoing various automated and manual quality inspections. The updated layer of the point cloud map is a localized area within the work scenario. When the scene in that area changes, map data collection is triggered. The collected data is processed through point cloud mapping to create the updated layer. The map data collection process involves using sensors such as lidar and cameras to scan and capture the changed area to obtain the latest environmental data. The collected data is then processed through multiple steps such as point cloud registration, filtering, feature extraction, and gridding to produce the updated point cloud layer.

[0014] In step 2, the specific steps for instantiating the static layer and the update layer are as follows: Step 2.1: Extract static layers from the point cloud basemap layer. Automatically extract point clouds labeled as static objects based on the semantic information of the point cloud basemap layer. Manually verify the automatically extracted results to improve the accuracy of the static layer extraction. Step 2.2, performing label point cloud extraction on the updated layer of the point cloud map, and automatically extracting point clouds labeled as static objects based on the semantic information of the updated layer; Step 2.3: Instantiate the label point clouds in the static layer and the update layer respectively, and identify different object instances in the point cloud based on the region growing algorithm or the deep learning algorithm and label them separately.

[0015] In step 3, the specific steps for instance matching of the static layer and the update layer of the high-precision point cloud base map are as follows: Step 3.1: Record the instantiated point cloud set of the static layer as , It consists of multiple instance point clouds, namely , The first instance point cloud in the instanced point cloud set of the static layer, The second instance point cloud in the instanced point cloud set of the static layer, The instantiated point cloud set for the static layer Example point clouds ( ), The instantiated point cloud set for the static layer instance point cloud; the instanced point cloud set of the updated layer is recorded as , It consists of multiple instance point clouds, namely , To update the first instance point cloud in the instanced point cloud set of the layer, To update the second instance point cloud in the instanced point cloud set of the layer, To update the instantiated point cloud of the layer Example point clouds ( ), In the instantiated point cloud set of the updated layer, the nth instance point cloud is set; and As input for embodiment matching; Step 3.2, calculation Medium instance point cloud and Medium instance point cloud The similarity between the two is recorded as ; Step 3.3, set the instance similarity threshold as T, and judge and Similarity Is it greater than the instance similarity threshold? If , then Add to candidate match M, otherwise press give up ; Step 3.4, follow Order the candidate matches from largest to smallest Arrange the elements in to obtain the final successful matching set. The point cloud belonging to the static layer in the set is recorded as , the point cloud belonging to the updated layer is recorded as ; Step 3.5: Use point cloud matching algorithm to match the obtained instances and Perform the registration operation and output the pose change value as the accuracy evaluation index of the updated layer pose change value R * ,t * The calculation formula is as follows:

[0016] Where N is the number of point clouds to be registered, R is the rotation change matrix in the relative pose, t is the translation change in the relative pose, and pi is A point in A point in the.

[0017] In step 3.2, the similarity calculation between two instances is performed as follows: (1) k-nearest neighbor coverage (geometric overlap) For each point pi in instance A, search its k nearest neighbor points in instance B and calculate the points that meet the distance threshold d th The number of neighbors , count the number of matching points Nmatch and coverage Sknn of all points in instance A and instance B,

[0018] Sknn=Nmatch / ( NA+NB−Nmatch ) Where NA is the number of point clouds of instance A, and NB is the number of point clouds of instance B; (2) Euclidean distance between instance centers (spatial proximity) Assume that the centroid coordinates of instance A and instance B are CA and CB respectively. The centroid coordinates CA and CB are calculated by the following formula:

[0019] The Euclidean distance D between centroids is calculated by the following formula center , D center=∥ CA − CB ∥ 2 ; (3) Point cloud feature distance (semantic consistency) Extract the global feature vectors FA and FB of instance A and instance B, calculate the feature similarity, and calculate the cosine similarity S using the following formula feat , S feat= FA ⋅ Facebook / (∥ FA ∥∥ Facebook ∥) ; (4) Comprehensive similarity Fusion of multiple indicators, weighted to obtain comprehensive similarity S total , S total= αS knn+ βS center+ γS feat Where α is the weighted coefficient for k-nearest neighbor coverage, β is the weighted coefficient for the Euclidean distance to the instance center, and γ is the weighted coefficient for the point cloud feature distance. The value of α + β + γ = 1 is sufficient. The specific value depends on the scenario. For example, you can set α = 0.4, β = 0.3, and γ = 0.3.

[0020] In step 3.4, the elements in the sorted candidate matching M are and Do the following in order: a) If Already appeared in In, skip; if Not present in In join in ; b) If Already appeared in In, skip; if Not present in In join in .

[0021] The advantages of the present invention are that by combining the semantic information and instance matching of the static layer and the update layer, it can effectively adapt to dynamic environmental changes and improve the accuracy and reliability of quality inspection; accurate instance matching and alignment can reduce the impact of noise and omissions in point cloud data on the positioning and perception modules of the autonomous driving system; and by utilizing the association information between the update layer and the base map layer, the overall efficiency and accuracy of quality inspection can be improved.

[0022] This invention combines semantic information and instance matching to adapt to dynamic environmental changes. During the quality inspection process, not only the information of the static layer is considered, but also the information of the updated layer. By extracting semantic information, the categories and attributes of different objects can be identified. Then, through instance matching technology, the same or similar objects in the static layer and the updated layer can be accurately matched. This method of combining semantic information and instance matching can more effectively adapt to changes in the dynamic environment, maintaining the accuracy and reliability of quality inspection even when the position and shape of objects change.

[0023] This invention reduces the impact of noise and missing data on autonomous driving systems. Point cloud data is a crucial input for positioning and perception modules in autonomous driving systems. However, point cloud data often suffers from noise and missing data, impacting system performance and safety. This invention effectively reduces noise and missing data in point cloud data through precise instance matching and registration techniques. This allows the positioning and perception modules of autonomous driving systems to obtain more accurate and complete point cloud data, thereby improving system performance and safety.

[0024] This invention fully utilizes the correlation information between the update layer and the basemap layer to improve quality inspection efficiency and accuracy. By analyzing this correlation information, the autonomous driving system can more quickly identify the corresponding relationships between objects, thereby performing more accurate quality inspections. Furthermore, utilizing this correlation information can reduce unnecessary calculations and analysis processes, improving the overall efficiency of quality inspections. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present invention will be further described below with reference to the accompanying drawings.

[0026] Figure 1 This is the point cloud map update layer quality inspection flow chart of the present invention.

[0027] Figure 2 Schematic diagram of a point cloud for instantiation of the present invention. DETAILED DESCRIPTION Example 1

[0028] like Figure 1 As shown in the figure, a method for evaluating the accuracy of updating a point cloud map layer mainly includes two steps: instantiation of map features and point cloud matching. The specific process is as follows: Step 1: Obtain a high-precision point cloud base map and an updated layer of the point cloud map. The high-precision point cloud base map is a static point cloud dataset containing detailed geographic information.

[0029] High-precision point cloud base Figure 1 A point cloud map layer is typically created for the first time for a specific operational scenario and undergoes various automated and manual quality checks. It serves as a ground truth reference for evaluating the accuracy of subsequent updated layers. A high-precision point cloud basemap layer is a complete point cloud map layer created for the first time for a specific operational scenario. This layer not only contains a large amount of point cloud data, but also undergoes rigorous automated processing and manual quality checks to ensure its high precision and accuracy. Due to the high precision of the point cloud basemap layer, it is often used as a ground truth reference for evaluating the accuracy of subsequent updated layers. During subsequent updates, the new point cloud layer is compared and analyzed with the basemap layer to assess its precision and accuracy, ensuring that the updated layer meets specific application requirements. A point cloud map update layer typically represents a localized area within the operational scenario. Changes to this area trigger the map data collection process, which is then processed by the point cloud map production process to create the updated layer. The map data collection process involves scanning and capturing the changed area using sensors such as lidar and cameras to obtain the latest environmental data. The collected data is processed through multiple steps such as point cloud registration, filtering, feature extraction, and gridding to obtain the updated point cloud layer.

[0030] Step 2: Extract semantic information from the high-precision point cloud base map and update layer and instantiate the point cloud base map static layer and update layer. The semantic information includes the classification and location information of geographic elements. Geographic elements include but are not limited to buildings, roads, vegetation, and traffic signs.

[0031] The specific steps for instantiating static layers and update layers are as follows: Step 2.1: Extract static layers from the point cloud basemap layer. Automatically extract point clouds labeled as static objects based on the semantic information of the point cloud basemap layer. Manually verify the automatically extracted results to improve the accuracy of the static layer extraction. Step 2.2, performing label point cloud extraction on the updated layer of the point cloud map, and automatically extracting point clouds labeled as static objects based on the semantic information of the updated layer; Step 2.3: Instantiate the labeled point clouds in the static layer and the update layer respectively, and identify different object instances in the point cloud and label them separately based on the region growing algorithm or the 3D convolutional neural network and graph neural network algorithm.

[0032] In step 2.1, static layers are extracted from the point cloud basemap layer. This step only needs to be performed once for the basemap of a specific scene. The extracted static layers can then be used directly during subsequent layer quality checks. Static layer extraction typically combines automatic extraction and manual verification. The automatic extraction phase relies on the semantic information of the basemap layer. This assumes that label information for each point in the basemap layer has already been acquired in the previous steps. The automatic extraction phase specifically extracts point clouds labeled with static objects in the physical world, such as buildings, fixed guardrails, utility poles, and road signs. These static objects are relatively fixed in the scene and are resistant to change, making them suitable for inclusion in static layers. To further improve the accuracy of static layer extraction, manual verification is often employed. Manual verification reviews the automatically extracted results, correcting any errors or omissions, thereby ensuring the quality and accuracy of the static layers. Directly using the extracted and verified static layers in subsequent layer updates and quality checks improves efficiency and accuracy.

[0033] In step 2.2, labeled point cloud extraction is performed on the update layer. This step is consistent with the automatic extraction portion of the static layer extraction. To improve quality inspection efficiency and achieve automation, this step does not require manual verification. Specifically, it relies on the semantic information of the update layer. That is, the label information of each point in the update layer has been obtained in the previous step. The extracted point cloud is labeled as static objects in the physical world, such as buildings, fixed guardrails, utility poles, road signs, etc.

[0034] In step 2.3, instantiation is also called instance segmentation, which is to identify different object instances in the point cloud and mark them separately. The main methods of instantiation are divided into two categories, based on traditional algorithms and based on deep learning. Different object instances are identified in the point cloud and marked separately based on traditional algorithms or deep learning algorithms. Traditional algorithms are mainly region growing algorithms, and deep learning algorithms include algorithms based on 3D convolutional neural networks and graph neural networks. The schematic diagram of instantiated point cloud is as follows Figure 2 As shown, each instance is distinguished by a different color.

[0035] The specific steps of the region growing algorithm are as follows: a. Instance Growth (Seed Expansion) 1. Select the center point: Select a point from the unprocessed point cloud as the initial center (usually a boundary point or a feature salient point).

[0036] 2. Neighborhood search: With the point as the center, search for its neighboring points (such as Take the nearest 20 points).

[0037] 3. Similarity Determination: Neighboring points are included in the current instance if they meet the following conditions: geometric constraints (e.g., Euclidean distance from the center point ≤ 0.1m); surface continuity (e.g., normal angle < 30°); and attribute consistency (optional, e.g., reflection intensity difference < 10%). If a neighboring point does not meet these conditions, it is discarded and not included in the current instance.

[0038] 4. Iterative expansion: Add neighboring points that meet the conditions to the queue and repeat the above process until the current instance can no longer be expanded.

[0039] b. Loop through the remaining points 1. Mark complete: Mark the merged points as "processed".

[0040] 2. Repeated growth: Select a new center point from the remaining points and repeat step a until all points are processed.

[0041] c. Post-processing optimization 1. Over-segmentation merging: Two instances are merged into the same object if they meet the following conditions: geometric proximity (e.g., minimum distance between instances ≤ 0.5m) and feature similarity (e.g., average normal angle < 15°). If the two instances do not meet these conditions, they are not merged into the same object.

[0042] 2. Noise filtering: Eliminate instances with too few points (e.g., points < 30).

[0043] The steps of the collaborative algorithm of 3D convolutional neural network (CNN) and graph neural network (GNN) are as follows: a. Local geometric feature extraction based on 3D CNN 1. Input data encoding: Convert the original point cloud into structured input, including coordinates (x, y, z), normal vectors (nx, ny, nz), reflection intensity and other attributes.

[0044] 2. Local feature learning: Extract multi-scale local features through multi-layer 3D sparse convolution (such as SparseConvNet): output a high-dimensional feature vector for each point, encoding local geometric and semantic information.

[0045] b. Global topology modeling based on GNN 1. Graph structure construction: Build a point cloud graph structure based on the feature vector output by the 3D CNN.

[0046] 2. Topological relationship reasoning: Use graph convolutional networks (such as EdgeConv or Graph Transformer) to aggregate neighborhood features, output instance-aware global features, and distinguish topological boundaries between different objects.

[0047] c. Feature fusion and instance segmentation: The local features of 3D CNN are concatenated with the global features of GNN and input into the segmentation head: dynamic convolution or mask prediction is used to generate instance labels.

[0048] Step 3: Using the extracted semantic information, find matching instances between the high-precision point cloud base map and the updated layer by comparing the location, shape, size and other features of the geographic elements.

[0049] The specific steps for instance matching of the static layer and update layer of the high-precision point cloud base map are as follows: Step 3.1: Record the instantiated point cloud set of the static layer as , It consists of multiple instance point clouds, namely , The first instance point cloud in the instanced point cloud set of the static layer, The second instance point cloud in the instanced point cloud set of the static layer, The instantiated point cloud set for the static layer Example point clouds ( ), The instantiated point cloud set for the static layer instance point cloud; the instanced point cloud set of the updated layer is recorded as , It consists of multiple instance point clouds, namely , To update the first instance point cloud in the instanced point cloud set of the layer, To update the second instance point cloud in the instanced point cloud set of the layer, To update the instantiated point cloud of the layer Example point clouds ( ), To update the instantiated point cloud of the layer instance point cloud; and As input for embodiment matching; Step 3.2, calculation Medium instance point cloud and Medium instance point cloud The similarity between the two is recorded as ; Step 3.3, set the instance similarity threshold as T, and judge and Similarity Is it greater than the instance similarity threshold? If , then Add to candidate match M, otherwise press give up ; Step 3.4, follow Order the candidate matches from largest to smallest Arrange the elements in to obtain the final successful matching set. The point cloud belonging to the static layer in the set is recorded as , the point cloud belonging to the updated layer is recorded as .

[0050] In step 3.2, the similarity between the two instances is calculated using a combination of methods such as the k-nearest neighbor algorithm, the Euclidean distance between the center points of the point cloud instances, and the point cloud feature distance. The specific steps are as follows: (1) k-nearest neighbor coverage (geometric overlap) For each point pi in instance A, search for its k nearest neighbors (e.g., k=5) in instance B and calculate the distance threshold d th Number of neighboring points (such as 0.1m) , count the number of matching points Nmatch and coverage Sknn of all points in instance A and instance B,

[0051] Sknn=Nmatch / ( NA+NB−Nmatch ) in, Pi is a point in instance A, Pj is a point in instance B, NA is the number of point clouds in instance A, and NB is the number of point clouds in instance B; (2) Euclidean distance between instance centers (spatial proximity) Assume that the centroid coordinates of instance A and instance B are CA and CB respectively. The centroid coordinates CA and CB are calculated by the following formula:

[0052] The Euclidean distance D between centroids is calculated by the following formula center , D center=∥ CA − CB ∥ 2 ; (3) Point cloud feature distance (semantic consistency) Extract the global feature vectors FA and FB of instance A and instance B through PointNet or 3D CNN, calculate the feature similarity, and calculate the cosine similarity S using the following formula feat , S feat= FA ⋅ Facebook / (∥ FA ∥∥ Facebook ∥) ; (4) Comprehensive similarity Fusion of multiple indicators, weighted to obtain comprehensive similarity S total , S total= αS knn+ βS center+ γS feat Where α is the weighted coefficient of k-nearest neighbor coverage, β is the weighted coefficient of the Euclidean distance of the instance center point, and γ is the weighted coefficient of the point cloud feature distance.

[0053] In step 3.4, the elements in the sorted candidate matching M are and Do the following in order: a) If Already appeared in In, skip; if Not present in In join in ; b) If Already appeared in In, skip; if Not present in In join in .

[0054] Step 4: Based on the instance matching information, perform point cloud registration between the updated layer and the static layer, and accurately align the point cloud data in the updated layer to the high-precision point cloud base map.

[0055] Use point cloud matching algorithm to match the obtained instances and Perform the registration operation and output the pose change value as the accuracy evaluation index of the updated layer. The pose change value is recorded as , which is calculated as follows:

[0056] Where N is the number of point clouds to be registered, R is the rotation change matrix in the relative pose, t is the translation change in the relative pose, and pi is A point in A point in the.

[0057] Use point cloud matching algorithm to match instances and Perform registration to obtain the pose change value ， Point cloud registration uses the classic iterative closest point algorithm (ICP), the improved ICP method combined with point cloud covariance (GICP), or the probability distribution based registration method (NDT) 。

[0058] Step 5: After completing the point cloud registration, evaluate the accuracy of the updated layer.

[0059] In general, this invention uses the pose change value of the static layer to evaluate the accuracy of the updated layer. Generally, smaller pose change values ​​are considered better. In practical applications, when the pose change value is less than a set threshold, the accuracy of the evaluated map update layer is considered to meet the requirements. Otherwise, the evaluated map update layer does not meet the accuracy requirements.

[0060] This highly automated and highly reusable quality inspection process allows for a single-step extraction and instantiation of static layers, allowing for reuse during subsequent quality inspections of updated map layers. It also enables fully automatic label extraction for updated layers, minimizing manual intervention. Furthermore, a high-precision quality inspection method based on instance matching is employed to segment the point clouds of the static base map and updated layers. This method leverages semantic and geometric information for precise matching, minimizing interference with subsequent point cloud registration of dynamic objects.

[0061] By combining the semantic information and instance matching of static layers and update layers, the present invention can effectively adapt to dynamic environmental changes and improve the accuracy and reliability of quality inspection. Through precise instance matching and registration, it can effectively reduce the impact of noise and omissions in point cloud data on the positioning and perception modules of the autonomous driving system. The present invention fully utilizes the association information between the update layer and the base map layer to improve the overall efficiency and accuracy of quality inspection.

[0062] It should be noted that the execution order of the above steps is determined by their internal logic and functions. As long as the order of execution can achieve the desired results of the technical solution disclosed in this patent, it should not impose any restrictions or constraints on the implementation of the present invention and its embodiments. In addition to the above embodiments, the present invention can also have other implementation methods. Any technical solution formed by equivalent replacement or equivalent transformation falls within the scope of protection required by the present invention.

Claims

1. A method for evaluating the accuracy of a point cloud map update layer, characterized in that: The following steps are involved: Step 1: Obtain a high-precision point cloud base map and an updated layer of the point cloud map. The high-precision point cloud base map is a static point cloud dataset containing detailed geographic information. Step 2: Extract semantic information from the high-precision point cloud base map and update layer and instantiate the point cloud base map static layer and update layer. The semantic information includes the classification and location information of geographic elements, including but not limited to buildings, roads, vegetation, and traffic signs. Step 3: Using the extracted semantic information, find matching instances between the high-precision point cloud base map and the updated layer by comparing the features of geographic elements; Step 4: Based on the instance matching information, perform point cloud registration between the updated layer and the static layer, and accurately align the point cloud data in the updated layer to the high-precision point cloud base map; Step 5: After completing the point cloud registration, evaluate the accuracy of the updated layer.

2. The accuracy assessment method for updating a point cloud map layer according to claim 1, characterized in that: In step 1, the high-precision point cloud base map is a complete scene point cloud map layer that is produced for the first time in a specific operation scene and has undergone various automated and manual quality inspections. The update layer of the point cloud map is a local area in the operation scene. When the scene in the area changes, map data collection is triggered, and the collected data is processed by point cloud map production to obtain an update layer.

3. The accuracy assessment method for updating a point cloud map layer according to claim 2, characterized in that: In step 2, the specific steps for instantiating the static layer and the update layer are as follows: Step 2.1: Extract static layers from the point cloud basemap layer. Automatically extract point clouds labeled as static objects based on the semantic information of the point cloud basemap layer. Manually verify the automatically extracted results to improve the accuracy of the static layer extraction. Step 2.2, performing label point cloud extraction on the updated layer of the point cloud map, and automatically extracting point clouds labeled as static objects based on the semantic information of the updated layer; Step 2.3: Instantiate the label point clouds in the static layer and the update layer respectively, and identify different object instances in the point cloud based on the region growing algorithm or the deep learning algorithm and label them separately.

4. The accuracy assessment method for updating a point cloud map layer according to claim 3 is characterized in that: In step 3, the specific steps for instance matching of the static layer and the update layer of the high-precision point cloud base map are as follows: Step 3.1: Record the instantiated point cloud set of the static layer as , , The instantiated point cloud set for the static layer instance point cloud; the instanced point cloud set of the updated layer is recorded as , , To update the instantiated point cloud of the layer instance point clouds; Step 3.2, calculation Medium instance point cloud and Medium instance point cloud The similarity between the two is recorded as ; Step 3.3, set the instance similarity threshold as T, and judge and Similarity Is it greater than the instance similarity threshold? If , then Add to candidate match M, otherwise press give up ; Step 3.4, follow Order the candidate matches from largest to smallest Arrange the elements in to obtain the final successful matching set. The point cloud belonging to the static layer in the set is recorded as , the point cloud belonging to the updated layer is recorded as ; Step 3.5: Use point cloud matching algorithm to match the obtained instances and Perform registration operation and output the pose change value R * ,t * , as the accuracy evaluation index of the updated layer, the pose change value R * ,t * The calculation formula is as follows: ， Where N is the number of point clouds to be registered, R is the rotation change matrix in the relative pose, t is the translation change in the relative pose, and pi is A point in A point in the.

5. The accuracy assessment method for updating a point cloud map layer according to claim 4, characterized in that: In step 3.2, the similarity calculation between two instances is performed as follows: (1) k-nearest neighbor coverage For each point pi in instance A, search for its k nearest neighbors in instance B and calculate the number of neighboring points that meet the distance threshold n pi , count the number of matching points Nmatch and coverage Sknn of all points in instance A and instance B, Sknn=Nmatch / ( NA+NB−Nmatch ) Where NA is the number of point clouds of instance A, and NB is the number of point clouds of instance B; (2) Euclidean distance between instance centers Assume that the centroid coordinates of instance A and instance B are CA and CB respectively. The centroid coordinates CA and CB are calculated by the following formula: ， The Euclidean distance S between centroids is calculated by the following formula center , S center=∥ THAT − ? B ∥ 2 ; (3) Point cloud feature distance Extract the global feature vectors FA and FB of instance A and instance B, calculate the feature similarity, and calculate the cosine similarity S using the following formula feat , S feat= AGO ⋅ FB / (∥ AGO ∥∥ FB ∥) ; (4) Comprehensive similarity Fusion of multiple indicators, weighted to obtain comprehensive similarity S total , S total= αS knn+ βS center+ γS feat Where α is the weighted coefficient of k-nearest neighbor coverage, β is the weighted coefficient of the Euclidean distance of the instance center point, and γ is the weighted coefficient of the point cloud feature distance.

6. The accuracy assessment method for updating a point cloud map layer according to claim 4, characterized in that: In step 3.4, the elements in the sorted candidate matching M are and Do the following in order: a) If Already appeared in in, skip; like Not present in In join in ; b) If Already appeared in in, skip; like Not present in In join in .

Images