Object detection method and device

Abstract

This disclosure provides an object detection method and apparatus, which can be applied to the fields of artificial intelligence and autonomous driving environmental perception technology. The method includes: classifying initial point cloud information of an initial object based on preset classification rules to generate initial non-ground point cloud information corresponding to the initial point cloud information; inputting the initial non-ground point cloud information into a target detection model to output target non-ground point cloud information, wherein the target non-ground point cloud information includes target non-ground point cloud feature information and target non-ground point cloud category information; generating initial detection box information based on the target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules; updating the initial detection box information based on the target point cloud information and preset update rules to determine the target detection box information; and determining the target object based on the target detection box information. This method can further improve the efficiency and accuracy of target object detection.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and autonomous driving environmental perception technology, and in particular to an object detection method and apparatus. Background Technology

[0002] In modern transportation driver assistance systems and autonomous driving technologies, obstacle detection and localization have become crucial components for ensuring road safety and enhancing the driving experience. Traditional technologies include point-based 3D obstacle detection network models, which detect obstacles by extracting and analyzing features from point clouds.

[0003] In the process of conceiving this disclosure, the inventors discovered the following defects in the related technologies for obstacle detection: the related detection models are designed for massive point cloud data, which results in huge data volume, high system resource consumption and long computation time, making it difficult to meet real-time requirements; traditional detection methods are limited by feature extraction capabilities, leading to unstable detection performance and difficulty in meeting the needs of actual detection operations. Summary of the Invention

[0004] In view of the above problems, this disclosure provides an object detection method, apparatus, device, storage medium, and program product.

[0005] According to the first aspect of this disclosure, an object detection method is provided, comprising: classifying initial point cloud information of an initial object based on a preset classification rule, and generating initial non-ground point cloud information corresponding to the initial point cloud information;

[0006] The above-mentioned initial non-ground point cloud information is input into the above-mentioned target detection model, and the target non-ground point cloud information is output. The above-mentioned target non-ground point cloud information includes target non-ground point cloud feature information and target non-ground point cloud category information.

[0007] Based on the above target non-ground point cloud feature information, the above target non-ground point cloud category information, and the preset processing rules, initial detection box information is generated;

[0008] The initial detection box information is updated based on the target point cloud information and preset update rules to determine the target detection box information, wherein the target point cloud information represents the point cloud information in the initial detection box information; and

[0009] The target object is determined based on the target detection box information mentioned above.

[0010] According to embodiments of this disclosure, the target detection model includes a first filtering layer and a second filtering layer; wherein, inputting the initial non-ground point cloud information into the target detection model and outputting target non-ground point cloud information includes: inputting the initial non-ground point cloud information into the first filtering layer of the target detection model and outputting initial filtered point cloud information corresponding to the initial non-ground point cloud information; and inputting the initial filtered point cloud information into the second filtering layer of the target detection model and outputting the target non-ground point cloud information.

[0011] According to embodiments of this disclosure, the preset processing rules include a first processing rule and a second processing rule; wherein, generating initial detection box information based on the target non-ground point cloud feature information, the target non-ground point cloud category information, and the preset processing rules includes: generating the center point of the initial object based on the target non-ground point cloud feature information, the target non-ground point cloud category information, and the first processing rule; and generating the initial detection box information based on the center point and the second processing rule.

[0012] According to an embodiment of this disclosure, the second processing rule includes a sphere query rule and a feature extraction rule; wherein, generating the initial detection box information based on the center point and the second processing rule includes: using the center point as the sphere center point corresponding to the sphere query rule; and extracting the non-ground point cloud feature information of the target based on the sphere center point and the feature extraction rule to obtain the initial detection box information.

[0013] According to embodiments of this disclosure, the above-mentioned classification processing of the initial point cloud information of the initial object based on preset classification rules to generate initial non-ground point cloud information corresponding to the initial point cloud information includes: processing the initial point cloud information in a first dimension to generate first grid information; classifying the first grid information according to the preset classification rules to generate ground grid information and non-ground grid information; and extracting the initial point cloud information corresponding to the non-ground grid information to generate the initial non-ground point cloud information.

[0014] According to embodiments of this disclosure, the second filtering layer is associated with the initial filtering point cloud category information and the initial filtering point cloud feature information in the initial filtering point cloud information; wherein, the input of the initial filtering point cloud information to the second filtering layer of the target detection model and the output of the target non-ground point cloud information includes: normalizing the initial confidence value corresponding to the initial filtering point cloud information according to the second filtering layer to generate a target confidence value corresponding to the initial filtering point cloud information; and determining the target non-ground point cloud information according to the target confidence value, the initial filtering point cloud category information and the initial filtering point cloud feature information.

[0015] According to an embodiment of this disclosure, the preset update rule includes a first update rule; wherein, the above-mentioned updating the initial detection box information based on the target point cloud information and the preset update rule to determine the target detection box information includes: performing voxelization processing on the target point cloud information to generate a voxel grid; and updating the feature information in the target point cloud information according to the center point corresponding to the voxel grid, the sphere query rule and the first update rule to generate first updated feature information.

[0016] According to embodiments of this disclosure, the preset update rule further includes a second update rule; wherein, the above-mentioned updating the initial detection box information based on the target point cloud information and the preset update rule to determine the target detection box information further includes: updating the first update feature information according to the second update rule to generate second update feature information; and updating the initial detection box information according to the second update feature information to generate the target detection box information.

[0017] According to embodiments of this disclosure, the method further includes: generating category information corresponding to the initial object based on the target non-ground point cloud feature information, the target non-ground point cloud category information, and preset processing rules.

[0018] A second aspect of this disclosure provides an object detection apparatus, comprising:

[0019] The initial non-ground point cloud information generation module is used to classify the initial point cloud information of the initial object based on preset classification rules and generate initial non-ground point cloud information corresponding to the initial point cloud information.

[0020] The target non-ground point cloud information output module is used to input the above-mentioned initial non-ground point cloud information into the above-mentioned target detection model and output the target non-ground point cloud information, wherein the above-mentioned target non-ground point cloud information includes target non-ground point cloud feature information and target non-ground point cloud category information;

[0021] The initial detection box information generation module is used to generate initial detection box information based on the above-mentioned target non-ground point cloud feature information, the above-mentioned target non-ground point cloud category information, and preset processing rules.

[0022] The target detection box information determination module is used to update the initial detection box information based on target point cloud information and preset update rules to determine the target detection box information, wherein the target point cloud information represents the point cloud information in the initial detection box information; and

[0023] The target object determination module is used to determine the target object based on the target detection box information mentioned above.

[0024] A third aspect of this disclosure provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the method described above.

[0025] A fourth aspect of this disclosure also provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the methods described above.

[0026] The fifth aspect of this disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0027] According to the object detection method, apparatus, device, storage medium, and program product provided in this disclosure, initial point cloud information is rapidly classified based on preset classification rules to generate initial non-ground point cloud information with richer semantic information, while eliminating ground point cloud information with low reference value. This reduces the amount of data while improving the quality of the data. The initial non-ground point cloud information is then input into the target detection model to obtain target non-ground point cloud feature information and target non-ground point cloud category information. Based on the target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules, initial detection box information is generated comprehensively. The initial detection box information is then updated using target point cloud information and preset update rules to obtain target detection box information with higher detection accuracy to determine the target object, further improving the efficiency and accuracy of target object detection. Attached Figure Description

[0028] The foregoing contents, as well as other objects, features, and advantages of this disclosure, will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0029] Figure 1 This diagram schematically illustrates an application scenario of the object detection method and apparatus according to embodiments of the present disclosure.

[0030] Figure 2 A flowchart illustrating an object detection method according to an embodiment of the present disclosure is shown schematically.

[0031] Figure 3 The diagram illustrates the initial point cloud information classification of a grid-based ground segmentation algorithm according to an embodiment of the present disclosure.

[0032] Figure 4A This illustration schematically shows an initial non-terrestrial point cloud feature extraction based on an object detection model according to an embodiment of the present disclosure;

[0033] Figure 4BThe illustration shows a schematic diagram of processing initial filtered point cloud information based on a local attention mechanism according to an embodiment of the present disclosure;

[0034] Figure 5 This illustration shows a flowchart of updating feature information of target point cloud information based on a feature fusion algorithm according to an embodiment of the present disclosure;

[0035] Figure 6 A schematic diagram illustrating the system architecture of an object detection method according to an embodiment of the present disclosure is shown.

[0036] Figure 7 A schematic block diagram of an object detection apparatus according to an embodiment of the present disclosure is shown; and

[0037] Figure 8 A block diagram schematically illustrates an electronic device suitable for implementing an object detection method according to an embodiment of the present disclosure. Detailed Implementation

[0038] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0039] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the above-described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0040] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0041] When using expressions such as "at least one of A, B, and C", they should generally be interpreted in accordance with the meaning that is commonly understood by a person skilled in the art (e.g., "a system having at least one of A, B, and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B, and C, etc.).

[0042] In the technical solution disclosed herein, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse.

[0043] In conceiving this disclosure, the inventors discovered that for obstacle detection, related technologies employ detection models that deal with massive amounts of point cloud data, resulting in high system resource consumption and long computation times, making it difficult to meet real-time requirements. Traditional detection methods, limited by feature extraction capabilities, suffer from unstable detection performance, failing to meet the needs of actual detection operations. Therefore, this disclosure rapidly classifies initial point cloud information based on preset classification rules, generating initial non-ground point cloud information with richer semantic information, while eliminating ground point cloud information with lower reference value. This reduces data volume while improving data quality. The initial non-ground point cloud information is then input into the target detection model to obtain target non-ground point cloud feature information and target non-ground point cloud category information. Based on these features, category information, and preset processing rules, initial detection box information is comprehensively generated. Finally, the initial detection box information is updated using target point cloud information and preset update rules, resulting in target detection box information with higher detection accuracy to identify target objects, further improving the efficiency and accuracy of target object detection.

[0044] Embodiments of this disclosure provide an object detection method and apparatus. The method includes: classifying initial point cloud information of an initial object based on preset classification rules to generate initial non-ground point cloud information corresponding to the initial point cloud information; inputting the initial non-ground point cloud information into a target detection model to output target non-ground point cloud information, wherein the target non-ground point cloud information includes target non-ground point cloud feature information and target non-ground point cloud category information; generating initial detection box information based on the target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules; updating the initial detection box information based on the target point cloud information and preset update rules to determine the target detection box information, wherein the target point cloud information represents the point cloud information in the initial detection box information; and determining the target object based on the target detection box information.

[0045] Figure 1 The illustration shows an application scenario of the object detection method and apparatus according to embodiments of the present disclosure.

[0046] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, a server 105, and a lidar sensor 106. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, the server 105, and the lidar sensor 106. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0047] Users can interact with server 105 via network 104 using at least one of the first terminal device 101, second terminal device 102, and third terminal device 103 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, second terminal device 102, and third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0048] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0049] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0050] The lidar sensor 106 and the server 105 can be connected via network 104. By scanning the surrounding environment, the lidar sensor 106 can detect and identify objects such as roads, vehicles, pedestrians, and obstacles, thereby helping the vehicle to perform real-time environmental perception and obstacle detection.

[0051] It should be noted that the object detection method provided in this embodiment can generally be executed by server 105. Correspondingly, the object detection device provided in this embodiment can generally be located in server 105. The object detection method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the object detection device provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0052] It should be understood that Figure 1 The number of terminal devices, networks, lidar sensors, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, lidar sensors, and servers can be included.

[0053] Figure 2 A flowchart illustrating an object detection method according to an embodiment of the present disclosure is shown schematically.

[0054] like Figure 2 As shown, the object detection method of this embodiment includes operations S210 to S250.

[0055] In operation S210, the initial point cloud information of the initial object is classified based on the preset classification rules to generate initial non-ground point cloud information corresponding to the initial point cloud information.

[0056] According to embodiments of this disclosure, it can be applied to the detection of target objects (obstacles) within a target range in autonomous driving. The initial object can be an object (obstacle) within the target range captured by a LiDAR sensor. The initial point cloud data can be unprocessed point cloud data acquired by the LiDAR sensor, such as the Kth frame of LiDAR data. The preset classification rule can be a grid-based ground segmentation algorithm. By classifying the initial point cloud information using a grid-based ground segmentation algorithm, initial ground point cloud information and initial non-ground point cloud information can be obtained.

[0057] In operation S220, the initial non-ground point cloud information is input into the target detection model, and the target non-ground point cloud information is output. The target non-ground point cloud information includes target non-ground point cloud feature information and target non-ground point cloud category information.

[0058] According to embodiments of this disclosure, the target detection model can be a detection model including a first filtering layer and a second filtering layer. Initial non-ground point cloud information can be input into the first filtering layer of the target detection model to obtain initial filtered point cloud information, and then the initial filtered point cloud information can be used as input to the second filtering layer to obtain target non-ground point cloud feature information (feature information of key points) and target non-ground point cloud category information (category information of key points).

[0059] Operation S230 generates initial detection box information based on the target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules.

[0060] According to embodiments of this disclosure, the preset processing rule may be a strategy rule for processing the non-ground point cloud feature information of the target using a neural network model. The neural network model may include a multilayer perceptron (MLP) model or a feature aggregation network model (PointNet). The preset processing rule may also include a detection head based on the intersection over union (IoU) score to further process the fused feature information to obtain initial detection box information.

[0061] In operation S240, the initial detection box information is updated based on the target point cloud information and the preset update rules to determine the target detection box information, wherein the target point cloud information represents the point cloud information in the initial detection box information.

[0062] According to embodiments of this disclosure, the preset update rule can be a rule that updates the information of the initial detection box by means of feature fusion algorithm (Region of Interest, RoI), intersection over union (IoU) score, and parameter correction of the initial detection box, so as to obtain a target detection box with higher detection accuracy than the initial detection box.

[0063] In operation S250, the target object is determined based on the target detection box information.

[0064] According to embodiments of this disclosure, the target object can be a target obstacle detected by a target detection box. In autonomous driving, by detecting the target object, the autonomous driving system can use various obstacle avoidance algorithms to perform path planning and decision-making, and select the best path to avoid the obstacle.

[0065] According to embodiments of this disclosure, initial point cloud information is rapidly classified based on preset classification rules to generate initial non-ground point cloud information with richer semantic information. Ground point cloud information with low reference value is eliminated, reducing the amount of data while improving the quality of data. The initial non-ground point cloud information is then input into the target detection model to obtain target non-ground point cloud feature information and target non-ground point cloud category information. Based on the target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules, initial detection box information is generated. The initial detection box information is then updated using target point cloud information and preset update rules to obtain target detection box information with higher detection accuracy to determine the target object, further improving the efficiency and accuracy of target object detection.

[0066] According to embodiments of this disclosure, the target detection model includes a first filtering layer and a second filtering layer; wherein, inputting initial non-ground point cloud information into the target detection model and outputting target non-ground point cloud information includes: inputting the initial non-ground point cloud information into the first filtering layer of the target detection model and outputting initial filtered point cloud information corresponding to the initial non-ground point cloud information; and inputting the initial filtered point cloud information into the second filtering layer of the target detection model and outputting target non-ground point cloud information.

[0067] According to embodiments of this disclosure, the first filtering layer may be a Set Abstraction (SA) layer, which can perform group sampling and feature extraction processing on the initial non-ground point cloud information to obtain the initial filtered point cloud information. The second filtering layer may obtain the target non-ground point cloud information by using a local attention algorithm and category-aware sampling, based on the first filtering layer.

[0068] In one feasible embodiment, the initial non-ground point cloud information is downsampled through a set abstraction layer (LA-SA layer) enhanced by two set abstraction layers (SA) and two local attention layers (LA) to obtain the target non-ground point cloud information.

[0069] In one feasible embodiment, the ensemble abstraction layer (SA) may include two steps: sampling grouping and feature extraction. Distance-based Farthest Point Sampling (D-FPS) is used to sample the target non-terrestrial point cloud (key point cloud). A ball query method is used to group the target non-terrestrial point cloud, and feature extraction is performed using a feature aggregation network model (PointNet). It should be noted that after the initial grid-based ground segmentation algorithm, key points are selected only from the initial non-terrestrial point cloud information for the first key point sampling. However, during feature extraction, feature aggregation extracts features indiscriminately from the entire initial point cloud information.

[0070] According to embodiments of this disclosure, the initial non-terrestrial point cloud, which has richer semantic information, is processed by a local attention layer, which reduces the amount of data to be detected, lowers the computational complexity of the model, and saves system resources.

[0071] According to embodiments of this disclosure, the preset processing rules include a first processing rule and a second processing rule; wherein, generating initial detection box information based on target non-ground point cloud feature information, target non-ground point cloud category information and preset processing rules includes: generating the initial object's center point position according to the target non-ground point cloud feature information, target non-ground point cloud category information and the first processing rule; and generating initial detection box information according to the center point position and the second processing rule.

[0072] According to embodiments of this disclosure, the first processing rule can be a rule for predicting the center point (geometric center of the target object) of the initial object based on a multilayer perceptron (MLP) algorithm; the second processing rule can be a rule for processing the fused key point feature information based on a ball query rule, a feature aggregation network model (PointNet), and a detection head based on the intersection over union (IoU) score, thereby obtaining the initial detection box information.

[0073] According to embodiments of this disclosure, based on the multilayer perceptual network in the first processing rule, the representation of input data (target non-ground point cloud feature information, target non-ground point cloud category information) is learned through multiple hidden layers. That is, the input data can be weighted and summed through neurons in multiple hidden layers, and nonlinear transformation is performed through activation functions to predict the center point of the initial object.

[0074] According to embodiments of this disclosure, the second processing rule includes a sphere query rule and a feature extraction rule; wherein, generating initial detection box information based on the center point and the second processing rule includes: using the center point as the sphere center point corresponding to the sphere query rule; and extracting the non-ground point cloud feature information of the target based on the sphere center point and the feature extraction rule to obtain the initial detection box information.

[0075] According to embodiments of this disclosure, the center point of the sphere can represent the geometric center point of the initial object (e.g., the center point of vehicle A), that is, the geometric center point of the initial object is used as the center point of the sphere query rule. The feature extraction rule can be a rule that extracts non-terrestrial point cloud feature information of the target based on the feature aggregation network model (PointNet), and then inputs the extracted feature information into the detection head based on the intersection over union (IoU) score to generate the rule for initial detection box information.

[0076] According to embodiments of this disclosure, initial point cloud information of an initial object is classified based on preset classification rules to generate initial non-ground point cloud information corresponding to the initial point cloud information, including: processing the initial point cloud information in a first dimension to generate first grid information; classifying the first grid information according to preset classification rules to generate ground grid information and non-ground grid information; and extracting the initial point cloud information corresponding to the non-ground grid information to generate initial non-ground point cloud information.

[0077] Figure 3 The diagram illustrates the initial point cloud information classification of a grid-based ground segmentation algorithm according to an embodiment of the present disclosure.

[0078] like Figure 3 As shown, based on the coordinate system of the lidar sensor 310, the initial point cloud data can be rasterized on the x-axis–y-axis plane of the lidar sensor 310 coordinate system to obtain multiple grids 320, which can then be sorted from near to far. Simultaneously, the ground height of the initial point cloud data can be initialized to the installation height of the lidar sensor.

[0079] Furthermore, after rasterizing the initial point cloud data, the distance between the grid cells can be determined by whether it exceeds the segmentation range. If it is outside the segmentation range, the grid cell is directly classified as a non-ground grid 320. For grid cells within the segmentation range, the extreme values ​​and mean values ​​of the z-axis of the point cloud within the grid cell can be calculated in the coordinate system of the LiDAR sensor 310. The grid cells are then classified based on the extreme values ​​and mean values ​​of the z-axis. The classification method can include: first, determining whether the extreme value of the z-axis within the grid cell is greater than a threshold (e.g., 1). If it is greater than 1, it is identified as a non-ground grid 320. If it is less than 1, the difference between the mean value of the z-axis and the ground height can be further determined. If the difference is greater than 2, it is also identified as a non-ground grid 320. If it is less than 2, the grid cell is identified as a ground grid 330, and the ground height is updated. Thus, the initial non-ground point cloud information can be extracted from the point clouds in all non-ground grid cells 320.

[0080] In one feasible embodiment, the segmented non-ground grid can be post-processed, including operations such as removing isolated ground fragments, filling holes, and smoothing edges, which can further improve the accuracy and robustness of point cloud extraction.

[0081] According to embodiments of this disclosure, a grid-based ground segmentation algorithm can intervene in the distribution of initial non-ground point cloud information (key points) in the initial point cloud information, thereby increasing the proportion of foreground points in the key points and improving the detection accuracy of objects.

[0082] According to embodiments of this disclosure, the second filtering layer is associated with the initial filtering point cloud category information and the initial filtering point cloud feature information in the initial filtering point cloud information; wherein, inputting the initial filtering point cloud information into the second filtering layer of the target detection model and outputting target non-ground point cloud information includes: normalizing the initial confidence value corresponding to the initial filtering point cloud information according to the second filtering layer to generate a target confidence value corresponding to the initial filtering point cloud information; and determining the target non-ground point cloud information according to the target confidence value, the initial filtering point cloud category information, and the initial filtering point cloud feature information.

[0083] According to embodiments of this disclosure, the initial confidence value corresponding to the initial filtered point cloud information can be used to characterize the detection accuracy of the initial filtered point cloud information. For example, the initial confidence value of the initial filtered point cloud B is 86%. The target confidence value can be a value obtained after normalizing the initial confidence value, and the range of the target confidence value is (0,1). The second filtering layer can replace the distance-based farthest point sampling (D-FPS) in the first filtering layer with class-aware sampling, and at the same time use a local attention mechanism to filter the initial filtered point cloud information, thereby obtaining the target non-ground point cloud information.

[0084] Figure 4A The illustration shows an initial non-terrestrial point cloud feature extraction based on an object detection model according to an embodiment of the present disclosure.

[0085] like Figure 4A As shown, after the grid-based ground segmentation algorithm obtains the initial non-ground point cloud information, it can use two set abstraction (SA) layers in the target detection model, followed by distance-based farthest point sampling (D-FPS), ball query, and feature aggregation network algorithm (PointNet), to obtain the extracted key points (initial filtered point cloud), key point feature information, and confidence scores. Then, it uses a set abstraction layer (LA-SA layer) enhanced by two local attention (LA) layers for downsampling, and further obtains the aggregated feature information of the target non-ground point cloud through ball query and feature aggregation network algorithm (PointNet).

[0086] Figure 4B The illustration shows a schematic diagram of processing initial filtered point cloud information based on a local attention mechanism according to an embodiment of the present disclosure.

[0087] like Figure 4BAs shown, the initial confidence value 410 of the initial selected point cloud information can be used to weight the initial selected point cloud feature information 420. This includes normalizing the initial confidence value 410 corresponding to the initial selected point cloud information to obtain a confidence score 430, and extracting the initial selected point cloud feature information 420 through a multilayer perceptron (MLP) algorithm to obtain extracted feature information 440. Then, the extracted feature information 440 is multiplied by the confidence score 430 to obtain the target non-ground point cloud information 450. The target non-ground point cloud information 450 may include target non-ground point cloud feature information and target non-ground point cloud category information. The specific method of normalizing the initial confidence value is shown in formulas (1)-(2):

[0088] S = {s1, ..., s} N}∈RN (1);

[0089]

[0090] Where S represents the confidence of all point clouds within the scope of a ball query rule, max(S) and min(S) are the maximum and minimum initial confidence values ​​in S, respectively, σ represents the sigmoid function, ω is the parameter of the initial confidence value, and s' i This represents the normalized confidence score.

[0091] According to embodiments of this disclosure, the target non-ground point cloud information is obtained by filtering the initial point cloud information through a local attention mechanism. This allows the computational model to focus only on a portion of the input data, rather than all information in the global scope, reducing computational costs and making the model more efficient when processing large-scale data. At the same time, the model can adaptively select the region to focus on based on the characteristics of the input data, enabling the model to better handle input data of different scales and resolutions, and improving the model's performance on complex tasks.

[0092] According to embodiments of this disclosure, the preset update rule includes a first update rule; wherein, updating the initial detection box information based on the target point cloud information and the preset update rule to determine the target detection box information includes: performing voxelization processing on the target point cloud information to generate a voxel grid; and updating the feature information in the target point cloud information according to the center point position, sphere query rule and the first update rule corresponding to the voxel grid to generate first updated feature information.

[0093] According to embodiments of this disclosure, voxelization can be a process of converting target point cloud information in continuous space into a discrete voxel (three-dimensional pixel) mesh to facilitate subsequent analysis and processing. The first update rule can be a rule for updating the feature information of the target point cloud information based on a feature fusion algorithm (Region of Interest, RoI).

[0094] Figure 5 The illustration shows a schematic diagram of updating feature information of target point cloud information based on a feature fusion algorithm according to an embodiment of the present disclosure.

[0095] like Figure 5 As shown, the feature information update method in this embodiment includes: firstly, voxelizing the initial detection box to obtain multiple voxel grids; then, using the center of each voxel grid as the center, aggregating the feature information of the target point cloud according to the ball query rule. For the feature information of the first and second layers, since there is no category information, the PointNet feature aggregation network algorithm can be used for feature extraction; while for the feature information of the third and fourth layers, which have category information, a local attention mechanism can be used to weight the feature information first, and then the PointNet feature aggregation network algorithm can be used for feature extraction. For the feature information of size 8×8×2×128 obtained from the first layer, convolution can be performed first in the z-axis direction to obtain 8×8×128 feature information, and then convolution can be performed in the xy direction to obtain 4×4×256 feature information. For the feature information of subsequent layers, it can be concatenated with the feature information output from the upper layer, and then downsampled through convolution to obtain feature information of size 1×1×512 for the correction of the initial detection box information.

[0096] According to embodiments of this disclosure, for large-scale image information in a target object detection scenario, the feature fusion algorithm is used to operate only on the region of interest (local region), which effectively reduces the amount of computation and further improves the efficiency and speed of object detection.

[0097] According to embodiments of this disclosure, the preset update rule further includes a second update rule; wherein, updating the initial detection box information based on the target point cloud information and the preset update rule to determine the target detection box information further includes: updating the first update feature information according to the second update rule to generate second update feature information; and updating the initial detection box information according to the second update feature information to generate target detection box information.

[0098] According to embodiments of this disclosure, the second update rule can be a rule that updates the initial detection box by modifying the parameters and the intersection-union ratio (IU) score of the initial detection box information. After feature fusion using a Region of Interest (RoI) algorithm, the modification parameters and the IU score of the initial detection box can be obtained through two convolutional layers, respectively. The modification parameters are then used to modify the initial detection box, and non-maximum suppression is performed based on the IU score to obtain the target detection box information.

[0099] According to embodiments of this disclosure, in target object detection, a large number of candidate boxes are usually generated. In order to avoid detecting the same target multiple times, non-maximum suppression (NMS) is used to filter the candidate boxes. The intersection-union ratio (IUU) score is used to measure the degree of overlap between two boxes to determine whether the two boxes need to be merged or retained, which further improves the accuracy of target object detection.

[0100] According to embodiments of this disclosure, the method further includes: generating category information corresponding to the initial object based on target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules.

[0101] According to embodiments of this disclosure, the initial object category information is diverse, including pedestrians, vehicles, trees, and animals. The category of the initial object can be predicted using target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules, providing information for subsequent path planning and obstacle avoidance strategies in autonomous driving.

[0102] Figure 6 A schematic diagram of the system architecture of an object detection method according to an embodiment of the present disclosure is shown.

[0103] like Figure 6 As shown, a grid-based ground segmentation algorithm classifies the initial point cloud information to generate initial non-ground point cloud information corresponding to the initial point cloud information. This initial non-ground point cloud information is then input into two first filtering layers (SA-layer) and two second filtering layers (LA-SA-layer) in the target detection model, outputting target non-ground point cloud information (including target non-ground point cloud feature information and target non-ground point cloud category information). Based on the target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules (including multilayer perceptron algorithm, feature aggregation network algorithm, and detection head), initial detection box information and category information can be generated. Furthermore, based on the point cloud information in the initial detection box information, a feature fusion algorithm (Region of Interest, RoI) is used to voxelize the initial detection box, collect feature information, and then update the initial detection box information using correction parameters and intersection-union scores to obtain the target detection box information.

[0104] According to embodiments of this disclosure, a comprehensive comparison is made between the object detection method of this disclosure and existing object detection methods (including Point R-CNN, Point-Voxel CNN, and Voxel-RCNN), with the average accuracy of each method calculated based on a recall of 40. Specifically, Table 1 below shows a comparison of the detection performance and detection speed according to embodiments of this disclosure. The comparison in Table 1 reveals that the object detection method of this disclosure has better detection performance and detection speed.

[0105] Table 1

[0106]

[0107] Based on the above object detection method, this disclosure also provides an object detection device. The following will be combined with... Figure 7 The device is described in detail.

[0108] Figure 7 A schematic block diagram of an object detection apparatus according to an embodiment of the present disclosure is shown.

[0109] like Figure 7 As shown, the object detection device in this embodiment includes an initial non-ground point cloud information generation module 710, a target non-ground point cloud information output module 720, an initial detection box information generation module 730, a target detection box information determination module 740, and a target object determination module 750.

[0110] The initial non-ground point cloud information generation module 710 is used to classify the initial point cloud information of the initial object based on a preset classification rule, and generate initial non-ground point cloud information corresponding to the initial point cloud information. In one embodiment, the initial non-ground point cloud information generation module 710 can be used to perform the operation S210 described above, which will not be repeated here.

[0111] The target non-ground point cloud information output module 720 is used to input initial non-ground point cloud information into the target detection model and output target non-ground point cloud information, wherein the target non-ground point cloud information includes target non-ground point cloud feature information and target non-ground point cloud category information. In one embodiment, the target non-ground point cloud information output module 720 can be used to perform the operation S220 described above, which will not be repeated here.

[0112] The initial detection box information generation module 730 is used to generate initial detection box information based on the target non-ground point cloud feature information, the target non-ground point cloud category information, and preset processing rules. In one embodiment, the initial detection box information generation module 730 can be used to perform the operation S230 described above, which will not be repeated here.

[0113] The target detection box information determination module 740 is used to update the initial detection box information based on the target point cloud information and a preset update rule to determine the target detection box information, wherein the target point cloud information represents the point cloud information in the initial detection box information. In one embodiment, the target detection box information determination module 740 can be used to perform the operation S240 described above, which will not be repeated here.

[0114] The target object determination module 750 is used to determine the target object based on the target detection box information. In one embodiment, the target object determination module 750 can be used to perform the operation S250 described above, which will not be repeated here.

[0115] According to embodiments of this disclosure, the object detection device comprises an initial non-ground point cloud information generation module 710, a target non-ground point cloud information output module 720, an initial detection box information generation module 730, a target detection box information determination module 740, and a target object determination module 750. The initial point cloud information is rapidly classified based on preset classification rules, generating initial non-ground point cloud information with richer semantic information. Ground point cloud information with lower reference value is eliminated, reducing the amount of data while improving data quality. This initial non-ground point cloud information is then input into the target detection model to obtain target non-ground point cloud feature information and target non-ground point cloud category information. Based on the target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules, initial detection box information is comprehensively generated. Finally, the initial detection box information is updated using target point cloud information and preset update rules, resulting in target detection box information with higher detection accuracy to determine the target object, further improving the efficiency and accuracy of target object detection.

[0116] According to embodiments of this disclosure, the target detection model includes a first filtering layer and a second filtering layer; wherein, the target non-ground point cloud information output module includes: an initial filtering point cloud information output submodule and a target non-ground point cloud information output submodule.

[0117] The initial filtering point cloud information output submodule is used to input the initial non-ground point cloud information into the first filtering layer of the target detection model and output the initial filtering point cloud information corresponding to the initial non-ground point cloud information.

[0118] The target non-ground point cloud information output submodule is used to input the initial filtered point cloud information into the second filtering layer of the target detection model and output the target non-ground point cloud information.

[0119] According to embodiments of this disclosure, the preset processing rules include a first processing rule and a second processing rule; wherein, the initial detection box information generation module includes: a center point generation submodule and an initial detection box information generation submodule.

[0120] The center point generation submodule is used to generate the initial center point of the object based on the target non-ground point cloud feature information, the target non-ground point cloud category information, and the first processing rule.

[0121] The initial detection box information generation submodule is used to generate initial detection box information based on the center point location and the second processing rule.

[0122] According to embodiments of this disclosure, the second processing rule includes a sphere query rule and a feature extraction rule; wherein, the initial detection box information generation submodule includes: a sphere center point as a unit and a target non-ground point cloud feature information extraction unit.

[0123] The center point is used as a unit to identify the center point as the corresponding center point in the ball query rule.

[0124] The target non-ground point cloud feature information extraction unit is used to extract the target non-ground point cloud feature information according to the sphere center point and feature extraction rules to obtain the initial detection box information.

[0125] According to embodiments of this disclosure, the initial non-ground point cloud information generation module includes: a first raster information generation submodule, a raster information generation submodule, and an initial non-ground point cloud information generation submodule.

[0126] The first raster information generation submodule is used to process the initial point cloud information in the first dimension to generate the first raster information.

[0127] The raster information generation submodule is used to classify the first raster information according to preset classification rules and generate ground raster information and non-ground raster information.

[0128] The initial non-ground point cloud information generation submodule is used to extract the initial point cloud information corresponding to the non-ground raster information and generate the initial non-ground point cloud information.

[0129] According to embodiments of this disclosure, the second filtering layer is associated with the initial filtering point cloud category information and the initial filtering point cloud feature information in the initial filtering point cloud information; wherein, the target non-ground point cloud information output submodule includes: a target confidence value generation unit and a target non-ground point cloud information determination unit.

[0130] The target confidence value generation unit is used to normalize the initial confidence value corresponding to the initial filtered point cloud information based on the second filtering layer, and generate the target confidence value corresponding to the initial filtered point cloud information.

[0131] The target non-ground point cloud information determination unit is used to determine the target non-ground point cloud information based on the target confidence value, the initial filtered point cloud category information, and the initial filtered point cloud feature information.

[0132] According to embodiments of this disclosure, the preset update rule includes a first update rule; wherein, the target detection box information determination module includes: a voxel mesh generation submodule and a first update feature information generation submodule.

[0133] The voxel mesh generation submodule is used to voxelize the target point cloud information and generate a voxel mesh.

[0134] The first updated feature information generation submodule is used to update the feature information in the target point cloud information according to the center point position corresponding to the voxel grid, the sphere query rule and the first update rule, and generate the first updated feature information.

[0135] According to embodiments of this disclosure, the preset update rule further includes a second update rule; wherein, the target detection box information determination module further includes: a second update feature information generation submodule and a target detection box information generation submodule.

[0136] The second updated feature information generation submodule is used to update the first updated feature information according to the second update rule and generate the second updated feature information.

[0137] The target detection box information generation submodule is used to update the initial detection box information based on the second updated feature information to generate target detection box information.

[0138] According to embodiments of this disclosure, the apparatus further includes: a category information generation module, used to generate category information corresponding to the initial object based on target non-ground point cloud feature information, target non-ground point cloud category information, and preset processing rules.

[0139] According to embodiments of this disclosure, any and multiple modules among the initial non-ground point cloud information generation module 710, the target non-ground point cloud information output module 720, the initial detection box information generation module 730, the target detection box information determination module 740, and the target object determination module 750 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this disclosure, at least one of the initial non-ground point cloud information generation module 710, the target non-ground point cloud information output module 720, the initial detection box information generation module 730, the target detection box information determination module 740, and the target object determination module 750 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or any other reasonable means of integrating or packaging circuits, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the initial non-ground point cloud information generation module 710, the target non-ground point cloud information output module 720, the initial detection box information generation module 730, the target detection box information determination module 740, and the target object determination module 750 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0140] Figure 8 A block diagram schematically illustrates an electronic device suitable for implementing an object detection method according to an embodiment of the present disclosure.

[0141] like Figure 8 As shown, an electronic device 800 according to an embodiment of this disclosure includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this disclosure.

[0142] RAM 803 stores various programs and data required for the operation of electronic device 800. Processor 801, ROM 802, and RAM 803 are interconnected via bus 804. Processor 801 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 802 and / or RAM 803. It should be noted that the programs may also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in said one or more memories.

[0143] According to embodiments of this disclosure, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to a bus 804. The electronic device 800 may also include one or more of the following components connected to the input / output (I / O) interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output (I / O) interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0144] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0145] According to embodiments of this disclosure, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this disclosure, the computer-readable storage medium may include ROM 802 and / or RAM 803 and / or one or more memories other than ROM 802 and RAM 803 described above.

[0146] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the object detection method provided in the embodiments of this disclosure.

[0147] When the computer program is executed by the processor 801, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0148] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 809, and / or installed from a removable medium 811. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0149] In such an embodiment, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by processor 801, it performs the functions defined in the system of this disclosure embodiment. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0150] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0151] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0152] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. In particular, the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0153] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of this disclosure is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. An object detection method, comprising: The initial point cloud information of the initial object is classified based on the preset classification rules to generate initial non-ground point cloud information corresponding to the initial point cloud information. The initial non-ground point cloud information is input into the target detection model, and target non-ground point cloud information is output. The target non-ground point cloud information includes target non-ground point cloud feature information and target non-ground point cloud category information. The process includes: inputting the initial non-ground point cloud information into the first filtering layer of the target detection model, and outputting initial filtered point cloud information corresponding to the initial non-ground point cloud information; normalizing the initial confidence value corresponding to the initial filtered point cloud information according to the second filtering layer of the target detection model to generate a target confidence value corresponding to the initial filtered point cloud information; determining the target non-ground point cloud information based on the target confidence value, the initial filtered point cloud category information, and the feature information of the initial filtered point cloud, wherein the second filtering layer is associated with the initial filtered point cloud category information and the initial filtered point cloud feature information in the initial filtered point cloud information. Based on the target non-ground point cloud feature information, the target non-ground point cloud category information, and preset processing rules, initial detection box information is generated, including: generating the center point of the initial object according to the target non-ground point cloud feature information, the target non-ground point cloud category information, and the first processing rule in the preset processing rules; using the center point as the center point of the sphere corresponding to the sphere query rule; and extracting the target non-ground point cloud feature information according to the center point and feature extraction rules to obtain the initial detection box information. The initial detection box information is updated based on target point cloud information and preset update rules to determine target detection box information. The target point cloud information represents the point cloud information in the initial detection box information, including: voxelizing the target point cloud information to generate a voxel grid; updating the feature information in the target point cloud information according to the center point corresponding to the voxel grid, the sphere query rule, and a first update rule in the preset update rules to generate first updated feature information; updating the first updated feature information according to a second update rule in the preset update rules to generate second updated feature information; and updating the initial detection box information according to the second updated feature information to generate the target detection box information; and... The target object is determined based on the target detection box information.

2. The method according to claim 1, wherein, The process of classifying the initial point cloud information of the initial object based on a preset classification rule to generate initial non-ground point cloud information corresponding to the initial point cloud information includes: The initial point cloud information is processed in the first dimension to generate the first raster information; The first grid information is classified according to the preset classification rules to generate ground grid information and non-ground grid information; The initial point cloud information corresponding to the non-ground raster information is extracted to generate the initial non-ground point cloud information.

3. The method according to claim 1, further comprising: Based on the target non-ground point cloud feature information, the target non-ground point cloud category information, and preset processing rules, category information corresponding to the initial object is generated.

4. An object detection device, comprising: The initial non-ground point cloud information generation module is used to classify the initial point cloud information of the initial object based on a preset classification rule, and generate initial non-ground point cloud information corresponding to the initial point cloud information. A target non-ground point cloud information output module is used to input the initial non-ground point cloud information into a target detection model and output target non-ground point cloud information. The target non-ground point cloud information includes target non-ground point cloud feature information and target non-ground point cloud category information. The module includes: inputting the initial non-ground point cloud information into a first filtering layer of the target detection model and outputting initial filtered point cloud information corresponding to the initial non-ground point cloud information; normalizing the initial confidence value corresponding to the initial filtered point cloud information according to a second filtering layer of the target detection model to generate a target confidence value corresponding to the initial filtered point cloud information; and determining the target non-ground point cloud information based on the target confidence value, the initial filtered point cloud category information, and the feature information of the initial filtered point cloud. The second filtering layer is associated with the initial filtered point cloud category information and the initial filtered point cloud feature information in the initial filtered point cloud information. The initial detection box information generation module is used to generate initial detection box information based on the target non-ground point cloud feature information, the target non-ground point cloud category information, and preset processing rules. This includes: generating the center point of the initial object according to the target non-ground point cloud feature information, the target non-ground point cloud category information, and a first processing rule in the preset processing rules; using the center point as the center point of a sphere query rule; and extracting the target non-ground point cloud feature information according to the center point and feature extraction rules to obtain the initial detection box information. A target detection box information determination module is used to update the initial detection box information based on target point cloud information and a preset update rule to determine the target detection box information. The target point cloud information represents the point cloud information in the initial detection box information, including: performing voxelization on the target point cloud information to generate a voxel grid; updating the feature information in the target point cloud information according to the center point corresponding to the voxel grid, the sphere query rule, and a first update rule in the preset update rules to generate first updated feature information; updating the first updated feature information according to a second update rule in the preset update rules to generate second updated feature information; and updating the initial detection box information according to the second updated feature information to generate the target detection box information; and... The target object determination module is used to determine the target object based on the target detection box information.

Images