Ground object detection method, device, equipment, storage medium and program product

Abstract

The present application provides a method, apparatus, device, storage medium and program product for detecting objects; the method can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, and assisted driving; the method comprises: obtaining point cloud data of a target geographical range, wherein the target geographical range includes objects to be detected; performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data; performing feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data; based on the point cloud features, predicting vectorized detection results of the objects to be detected to obtain vectorized detection results of the objects to be detected, wherein the vectorized detection results are used to describe the geometric representation of the objects to be detected; the present application can improve the detection efficiency of object detection.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment, storage medium and program product for detecting ground objects. Background Art

[0002] Artificial Intelligence (AI) is a comprehensive technology in computer science. By studying the design principles and implementation methods of various intelligent machines, machines are given the functions of perception, reasoning and decision-making. Artificial intelligence technology is an interdisciplinary subject that covers a wide range of fields, such as natural language processing technology and machine learning / deep learning. With the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0003] Object detection is also an important application direction of artificial intelligence. In related technologies, the detection of objects is achieved in the following way: using a neural network model to perform semantic prediction on the collected point cloud to obtain a point cloud carrying predicted semantic information, and then clustering, segmenting, denoising, fitting and other processing are performed on the point cloud carrying semantic information to obtain the vectorized detection result of the object. However, since the prediction result of the neural network model is a point cloud carrying semantic information, it is still necessary to go through multiple complex processing steps (including clustering, segmentation, denoising, fitting, etc.) to generate a vectorized result, resulting in low efficiency of object detection and waste of hardware processing resources. Summary of the invention

[0004] The embodiments of the present application provide a ground object detection method, device, electronic device, computer-readable storage medium and computer program product, which can improve the detection efficiency of ground object detection.

[0005] The technical solution of the embodiment of the present application is implemented as follows:

[0006] The present application provides a method for detecting ground objects, including:

[0007] Acquire point cloud data of a target geographic range, wherein the target geographic range includes a ground object to be detected;

[0008] Performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data;

[0009] Performing feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data;

[0010] Based on the point cloud features, a vectorized detection result of the object to be detected is predicted to obtain a vectorized detection result of the object to be detected, and the vectorized detection result is used to describe the geometric representation of the object to be detected.

[0011] The present application also provides a ground object detection device, including:

[0012] An acquisition module, used to acquire point cloud data of a target geographical range, wherein the target geographical range includes the ground object to be detected;

[0013] A normalization module, used for performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data;

[0014] A feature extraction module, used to perform feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data;

[0015] A prediction module is used to predict the vectorized detection result of the object to be detected based on the point cloud features to obtain the vectorized detection result of the object to be detected, and the vectorized detection result is used to describe the geometric representation of the object to be detected.

[0016] In the above scheme, the acquisition module is also used to transmit a detection signal to the target geographical range; receive a reflection signal reflected by the target within the target geographical range in response to the detection signal; determine the three-dimensional position information of the target within the target geographical range based on the detection signal and the reflection signal; and use the three-dimensional position information of the target within the target geographical range as point cloud data of the target geographical range.

[0017] In the above scheme, the normalization module is also used to determine the original coordinate system in which the point cloud data is located and construct a normalized coordinate system; obtain a coordinate transformation matrix for transforming from the original coordinate system to the normalized coordinate system; based on the coordinate transformation matrix, perform coordinate transformation processing on the point cloud data in the original coordinate system to obtain point cloud data in the normalized coordinate system; and use the point cloud data in the normalized coordinate system as the normalized point cloud data.

[0018] In the above scheme, the acquisition module is also used to perform multiple point cloud acquisitions on the target geographical range according to the point cloud acquisition frequency through the vehicle-mounted point cloud acquisition device of the vehicle during the movement of the vehicle, so as to obtain multiple frames of first sub-point cloud data; convert each frame of the first sub-point cloud data to the target coordinate system respectively, so as to obtain multiple frames of second sub-point cloud data under the target coordinates; and splice the multiple frames of the second sub-point cloud data to obtain the point cloud data of the target geographical range.

[0019] In the above scheme, the normalization module is also used to obtain the data volume of the point cloud data; when the data volume exceeds the data volume threshold, the driving trajectory of the vehicle is obtained, and according to the driving trajectory, based on the data slicing method, the point cloud data is sliced ​​to obtain multiple point cloud data blocks; the normalization module is also used to perform coordinate normalization processing on each point cloud data block respectively to obtain normalized point cloud data.

[0020] In the above scheme, the acquisition module is also used to acquire a ground object detection model for ground object detection; the feature extraction module is also used to call the ground object detection model, perform feature extraction processing on the normalized point cloud data, and obtain point cloud features of the normalized point cloud data; the prediction module is also used to call the ground object detection model, and based on the point cloud features, predict the vectorized detection results of the ground object to be detected, and obtain the vectorized detection results of the ground object to be detected.

[0021] In the above scheme, the acquisition module is also used to obtain an initial ground object detection model and point cloud data samples of ground object samples, wherein the point cloud data samples are marked with vectorized detection result labels of the ground object samples; coordinate normalization processing is performed on the point cloud data samples to obtain normalized point cloud data samples; the initial ground object detection model is called to perform feature extraction processing on the normalized point cloud data samples to obtain point cloud feature samples of the normalized point cloud data samples; the initial ground object detection model is called to predict the vectorized detection results of the ground object samples to obtain predicted vectorized detection results of the ground object samples; based on the difference between the vectorized detection result label and the predicted vectorized detection result, the model parameters of the initial ground object detection model are updated to obtain the ground object detection model used for ground object detection.

[0022] In the above scheme, the acquisition module is also used to obtain a first point cloud data sample and a first vectorization detection result of the ground object sample; perform point cloud filtering processing on the first point cloud data sample to obtain a second point cloud data sample; perform slicing processing on the second point cloud data sample to obtain multiple point cloud data blocks, wherein the multiple point cloud data blocks constitute the point cloud data sample of the ground object sample; and associate the first vectorization detection result with the point cloud data sample to mark the first vectorization detection result as the vectorization detection result label of the point cloud data sample.

[0023] In the above scheme, the prediction module is also used to call the object detection model, predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; the point cloud data sample is also marked with the category label to which the object sample belongs, and the acquisition module is also used to call the initial object detection model, predict the vectorized detection result of the object sample, and obtain the predicted category to which the object sample belongs; the acquisition module is also used to update the model parameters of the initial object detection model based on the difference between the vectorized detection result label and the predicted vectorized detection result, and the difference between the category label and the predicted category.

[0024] In the above scheme, the method is applied to an autonomous driving scenario; the prediction module is also used to predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; the prediction module is also used to obtain the vehicle position of the autonomous driving vehicle; when the vehicle position is within the target geographical range, an autonomous driving route is generated based on the vehicle position, combined with the vectorized detection result and the target category; and the autonomous driving vehicle is controlled to travel according to the autonomous driving route.

[0025] In the above scheme, the prediction module is also used to predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; the prediction module is also used to generate a regional map of the target geographical area based on the vectorized detection results and the target category.

[0026] In the above scheme, the prediction module is also used to extract target point cloud data associated with the vectorized detection result from the normalized point cloud data; obtain target point cloud features of the target point cloud data; and predict the object category of the object to be detected based on the target point cloud features to obtain the target category to which the object to be detected belongs.

[0027] In the above scheme, the feature extraction module is also used to determine the three-dimensional point cloud space where the normalized point cloud data is located, and divide the three-dimensional point cloud space into multiple sub-point cloud spaces; for each sub-point cloud space, perform feature extraction processing on the data of the sub-point cloud space to obtain the spatial point cloud features of the sub-point cloud space; perform feature integration processing on the multiple spatial point cloud features in the three-dimensional point cloud space to obtain the point cloud features of the normalized point cloud data.

[0028] In the above scheme, the prediction module is also used to convert the point cloud features into a point cloud feature map, which includes multiple feature points; for each of the feature points, a three-dimensional anchor point frame centered on the feature point is determined; for each of the three-dimensional anchor point frames, a regression prediction of the vectorization detection result of the three-dimensional anchor point frame is performed to obtain an intermediate vectorization detection result; and the intermediate vectorization detection results of each of the three-dimensional anchor point frames are spliced ​​to obtain the vectorization detection result of the object to be detected.

[0029] In the above scheme, the ground object to be detected is the tunnel entrance of the tunnel, the point cloud data is collected when the vehicle passes through at least one of the tunnels, and the vectorized detection result includes the vectorized detection result of each tunnel entrance of at least one of the tunnels; the prediction module is also used to predict the category of each tunnel entrance based on the point cloud features, and obtain the tunnel entrance category to which the tunnel entrance belongs, and the tunnel entrance category includes a tunnel exit and a tunnel entrance; according to the tunnel entrance category of each tunnel entrance, the tunnel entrance and exit of each tunnel entrance are matched to obtain at least one tunnel entrance and exit pair; and the tunnel indicated by each tunnel entrance and exit pair is determined.

[0030] The present application also provides an electronic device, including:

[0031] A memory for storing computer executable instructions;

[0032] The processor is used to implement the ground object detection method provided in the embodiment of the present application when executing the computer executable instructions stored in the memory.

[0033] An embodiment of the present application also provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the ground object detection method provided in the embodiment of the present application is implemented.

[0034] The embodiment of the present application also provides a computer program product, including computer executable instructions or a computer program. When the computer executable instructions or the computer program are executed by a processor, the ground object detection method provided in the embodiment of the present application is implemented.

[0035] The embodiments of the present application have the following beneficial effects:

[0036] Using the above-mentioned embodiment of the present application, firstly, coordinate normalization processing is performed on the point cloud data of the target geographic range to obtain normalized point cloud data, and then feature extraction processing is performed on the normalized point cloud data to obtain point cloud features of the normalized point cloud data, thereby predicting the vectorized detection results of the objects to be detected based on the point cloud features, and obtaining the vectorized detection results of the objects to be detected, which are used to describe the geometric representation of the objects to be detected. In this way, based on the point cloud features of the point cloud data of the target geographic range (including the objects to be detected), the vectorized detection results of the objects to be detected can be directly predicted, which reduces the complexity of the object detection processing and improves the detection efficiency of the object detection and the utilization rate of the hardware processing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1A is a schematic diagram of a tunnel scenario provided in an embodiment of the present application;

[0038] Figure 1B is a schematic diagram of point cloud data provided by an embodiment of the present application;

[0039] Figure 1C is a schematic diagram of a vectorized detection result provided in an embodiment of the present application;

[0040] Figure 2 is a schematic diagram of the architecture of a ground object detection system 100 provided in an embodiment of the present application;

[0041] Figure 3 is a structural schematic diagram of an electronic device 500 for implementing a ground object detection method provided in an embodiment of the present application;

[0042] Figure 4 It is a flowchart of the ground object detection method provided in the embodiment of the present application;

[0043] Figure 5 is a schematic diagram of the model structure of the ground object detection model provided in the embodiment of the present application;

[0044] Figure 6 is a schematic diagram of a point cloud carrying semantic information provided in the related art;

[0045] Figure 7 is a schematic diagram of the tunnel structure definition provided in an embodiment of the present application;

[0046] Figure 8 is a training flow chart of the ground object detection model provided in the embodiment of the present application;

[0047] Fig. 9 is an application flow chart of the ground object detection model provided in the embodiment of the present application;

[0048] Fig.10It is a simplified schematic diagram of a tunnel exit and a tunnel entrance provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0050] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0051] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0053] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0054] 1) Client: An application running in a terminal that provides various services, such as a client that supports ground object detection.

[0055] 2) In response, it is used to indicate the conditions or states on which the executed operations depend. When the dependent conditions or states are met, one or more operations executed may be in real time or have a set delay. Unless otherwise specified, there is no restriction on the order of execution of the multiple operations executed.

[0056] 3) Deep Learning (DL), a new research direction in the field of Machine Learning (ML). Deep learning is the study of the inherent laws and representation levels of sample data. The information obtained in the learning process is of great help in interpreting data such as text, images, and sounds. The ultimate goal of deep learning is to enable machines to have analytical learning capabilities like humans and to recognize data such as text, images, and sounds.

[0057] 4) Tunnel: A traffic passage that passes through the inside of a mountain, underground, underwater, or inside a large building and allows safe driving.

[0058] 5) Tunnel entrance and tunnel exit: The tunnel entrance is the first position where the tunnel wall or tunnel top appears, and the tunnel exit is the last position where the tunnel disappears. The tunnel entrance and exit are important elements for high-precision map production and are also important objects for autonomous driving to perceive the surrounding environment for safe driving. Figure 1A As shown, Figure 1A The middle (1) figure shows a tunnel that goes through the interior of the mountain. Figure 1A The middle figure (2) shows a tunnel that goes through the underground.

[0059] 6) LiDAR: It is a radar system that detects the position and speed of a target by emitting a laser beam. The working principle is to emit a detection signal (laser beam) to the target, and then compare the reflected signal (target echo) received from the target with the emitted signal to calculate the three-dimensional position information of the target.

[0060] 7) Point cloud (i.e. point cloud data): The laser radar continuously scans the target and obtains a series of three-dimensional position information, which is the point cloud. Point clouds generally have anti-color intensity information (i.e. the anti-color ratio formed by the laser scanning on the target, which is related to the material of the target. The visualization effect of the point cloud is that the target of different materials has different colors). Figure 1B As shown, the left part is a picture of a certain geographical range, and the right part is the point cloud collected within the geographical range.

[0061] 8) Vectorization result (i.e. vectorization detection result) refers to the geometric shape composed of geometric primitives such as points, line segments, polylines, curves, rectangles, bounding boxes, etc. It is a simplified geometric representation of two-dimensional and three-dimensional objects (including point clouds), with the characteristics of small number of points, simple structure, simple storage, and easy to understand. Figure 1C As shown, this is the vector detection result of the feature “tunnel”.

[0062] 9) Land features are defined as fixed objects on the earth’s surface, including those formed naturally and artificially. They generally refer to relatively fixed objects on the earth’s surface, such as bridges, tunnels, caves, openings, gantries, etc.

[0063] The embodiments of the present application provide a method, device, electronic device, computer-readable storage medium, and computer program product for detecting ground objects, which can improve the detection efficiency of ground object detection.

[0064] It should be noted that when the embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0065] The following describes the ground object detection system provided by the embodiment of the present application. Figure 2 , Figure 2 It is a schematic diagram of the architecture of the ground object detection system 100 provided in an embodiment of the present application. To support an exemplary application, a terminal (terminal 400-1 is shown as an example) is connected to a server 200 via a network 300. The network 300 may be a wide area network or a local area network, or a combination of the two, and data transmission is achieved using wireless or wired links.

[0066] The terminal (e.g., 400 - 1 ) is configured to send a request for detecting a ground object within a target geographical range to the server 200 in response to a ground object detection instruction for the target geographical range;

[0067] The server 200 is used to receive a ground object detection request for a target geographical range sent by a terminal; in response to the ground object detection request, obtain point cloud data of the target geographical range, where the target geographical range includes the ground object to be detected; perform coordinate normalization processing on the point cloud data to obtain normalized point cloud data; perform feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data; based on the point cloud features, predict the vectorized detection result of the ground object to be detected to obtain the vectorized detection result of the ground object to be detected, where the vectorized detection result is a geometric representation of the ground object to be detected; and return the vectorized detection result to the terminal;

[0068] The terminal (eg, 400 - 1 ) is further configured to receive the vectorization detection result returned by the server and display the vectorization detection result.

[0069] In some embodiments, the ground object detection method provided in the embodiments of the present application can be implemented by various electronic devices, for example, it can be implemented by a terminal alone, it can also be implemented by a server alone, or it can be implemented by a terminal and a server in collaboration. The ground object detection method provided in the embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, maps, etc.

[0070] In some embodiments, the electronic device for implementing the method for detecting objects provided in the embodiments of the present application may be various types of terminals or servers. Among them, the server (e.g., server 200) may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. The terminal (e.g., terminal 400-1) may be a smart phone, a tablet computer, a laptop computer, a desktop computer, an intelligent voice interaction device (e.g., an intelligent speaker), a smart home appliance (e.g., a smart TV), a smart watch, a vehicle-mounted terminal, a wearable device, a virtual reality (VR) device, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, and the embodiments of the present application do not limit this.

[0071] In some embodiments, the object detection method provided in the embodiments of the present application can be implemented with the aid of cloud technology (Cloud Technology). Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and network in a wide area network or a local area network to realize the calculation, storage, processing, and sharing of data. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, and application technology based on the cloud computing business model application, which can form a resource pool, which is used on demand and is flexible and convenient. Cloud computing technology will become an important support. The background services of the technical network system require a large amount of computing resources and storage resources. As an example, a server (such as server 200) can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution networks (Content Delivery Network, CDN), and big data and artificial intelligence platforms.

[0072] In some embodiments, multiple servers may form a blockchain, and the servers are nodes on the blockchain. Information connections may exist between each node in the blockchain, and information may be transmitted between nodes through information connections. Among them, data related to the ground object detection method provided in the embodiment of the present application (such as point cloud data, vectorized detection results, ground object detection models, etc.) can be stored on the blockchain.

[0073] In some embodiments, the terminal or server can implement the ground object detection method provided in the embodiments of the present application by running a computer program. For example, the computer program can be a native program or software module in the operating system; it can be a native application (Application, APP), that is, a program that needs to be installed in the operating system to run; it can also be a small program, that is, a program that can be run only by downloading it to a browser environment; it can also be a small program that can be embedded in any APP. In short, the above-mentioned computer program can be an application, module or plug-in in any form.

[0074] The following describes an electronic device for implementing a method for detecting ground objects provided in an embodiment of the present application. Figure 3 , Figure 3 It is a schematic diagram of the structure of an electronic device 500 for implementing a ground object detection method provided in an embodiment of the present application. The electronic device 500 provided in an embodiment of the present application may be a terminal or a server. The electronic device 500 provided in an embodiment of the present application includes: at least one processor 510, a memory 550, at least one network interface 520 and a user interface 530. The various components in the electronic device 500 are coupled together via a bus system 540. It is understandable that the bus system 540 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 540 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, in Figure 3 Various buses are labeled as bus system 540 .

[0075] The processor 510 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0076] The user interface 530 includes one or more output devices 531 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 530 also includes one or more input devices 532, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0077] The memory 550 may be removable, non-removable, or a combination thereof. The memory 550 may include one or more storage devices physically located away from the processor 510. The memory 550 includes a volatile memory or a non-volatile memory, and may also include both a volatile memory and a non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 550 described in the embodiments of the present application is intended to include any suitable type of memory.

[0078] In some embodiments, the memory 550 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0079] Operating system 551, including system programs for processing various basic system services and performing hardware-related tasks, such as framework layer, core library layer, driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0080] A network communication module 552, used to reach other electronic devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 include: Bluetooth, Wireless Compatibility Certification (WiFi), and Universal Serial Bus (USB), etc.;

[0081] a presentation module 553 for enabling presentation of information via one or more output devices 531 (e.g., display screen, speaker, etc.) associated with the user interface 530 (e.g., a user interface for operating peripherals and displaying content and information);

[0082] The input processing module 554 is used to detect one or more user inputs or interactions from one of the one or more input devices 532 and translate the detected inputs or interactions.

[0083] In some embodiments, the ground object detection device provided in the embodiments of the present application can be implemented in a software manner. Figure 3 A ground object detection device 555 stored in a memory 550 is shown, which may be software in the form of a program and a plug-in, including the following software modules: an acquisition module 5551, a normalization module 5552, a feature extraction module 5553 and a prediction module 5554. These modules are logical, and therefore may be arbitrarily combined or further split according to the functions implemented. The functions of each module will be described below.

[0084] The following describes the ground object detection method provided by the embodiment of the present application. In some embodiments, the ground object detection method provided by the embodiment of the present application can be implemented by various electronic devices, for example, it can be implemented by a terminal alone, it can also be implemented by a server alone, or it can be implemented by a terminal and a server in collaboration. Figure 4 , Figure 4 : is a flow chart of a method for detecting ground objects provided in an embodiment of the present application. The method for detecting ground objects provided in an embodiment of the present application includes:

[0085] Step 101: The terminal obtains point cloud data of a target geographic range.

[0086] The target geographical range includes the objects to be detected.

[0087] In step 101, the terminal may be provided with a client, such as a client supporting ground object detection. When the terminal receives an operation instruction for the client, it may perform ground object detection by running the client. In actual application, the terminal may be a vehicle-mounted terminal; the target geographical range is the range to be detected for ground object detection, and the target geographical range includes the ground objects to be detected. When performing ground object detection, the terminal first obtains point cloud data of the target geographical range, and then processes the point cloud data to detect the objects to be detected within the target geographical range. In actual application, the point cloud data may be collected by the terminal in real time, or it may be existing point cloud data obtained from other parties (such as a server).

[0088] In some embodiments, the terminal can obtain point cloud data of the target geographic range in the following manner: transmit a detection signal into the target geographic range; receive a reflection signal reflected by the target within the target geographic range in response to the detection signal; determine the three-dimensional position information of the target within the target geographic range based on the detection signal and the reflection signal; and use the three-dimensional position information of the target within the target geographic range as point cloud data of the target geographic range.

[0089] Here, when the terminal collects point cloud data, it can first transmit a detection signal to the target geographical range. In actual implementation, the terminal can control the detection device (such as a laser device) to transmit a detection signal (such as a laser signal) to the target geographical range. The detection device can be separate from the terminal or integrated into the terminal. Since the detection signal will be reflected when it encounters the target (i.e., the object within the target geographical range), the terminal detects the reflected signal of the detection signal after transmitting the detection signal to receive the reflected signal reflected by the target within the target geographical range in response to the detection signal. Based on this, the three-dimensional position information of the target within the target geographical range can be calculated by comparing the detection signal and the reflected signal, so that the three-dimensional position information of the target within the target geographical range can be used as the point cloud data of the target geographical range.

[0090] In some embodiments, the terminal may also obtain point cloud data of the target geographic range in the following manner: during the movement of the vehicle, point cloud acquisition equipment on the vehicle is used to perform multiple point cloud acquisitions on the target geographic range according to the point cloud acquisition frequency to obtain multiple frames of first sub-point cloud data; each frame of the first sub-point cloud data is converted to the target coordinate system to obtain multiple frames of second sub-point cloud data under the target coordinates; and multiple frames of the second sub-point cloud data are spliced ​​to obtain point cloud data of the target geographic range.

[0091] Here, the terminal may be a vehicle-mounted terminal. During the movement of the vehicle, the vehicle-mounted point cloud acquisition device (such as a vehicle-mounted laser device) may be used to perform multiple point cloud acquisitions on the target geographic range according to the point cloud acquisition frequency to obtain multiple frames of first sub-point cloud data. In practical applications, the point cloud acquisition frequency may be pre-set, for example, during the movement of the vehicle, point cloud data is collected once for each target distance (such as 10 meters). Since multiple acquisitions are performed during the movement of the vehicle, the coordinate systems of the collected first sub-point cloud data of each frame are different. At this time, each frame of the first sub-point cloud data needs to be converted to the target coordinate system respectively, thereby obtaining multiple frames of second sub-point cloud data under the target coordinates. To facilitate data processing, multiple frames of second sub-point cloud data under the target coordinates may be spliced ​​to obtain spliced ​​point cloud data, which is the point cloud data of the target geographic range.

[0092] In some embodiments, before the terminal performs coordinate normalization processing on the point cloud data to obtain normalized point cloud data, it may also perform data preprocessing on the point cloud data in the following manner: obtain the data volume of the point cloud data; when the data volume exceeds the data volume threshold, obtain the driving trajectory of the vehicle, and slice the point cloud data according to the driving trajectory based on the data slicing method to obtain multiple point cloud data blocks; accordingly, the terminal can perform coordinate normalization processing on the point cloud data in the following manner to obtain normalized point cloud data: perform coordinate normalization processing on each point cloud data block separately to obtain normalized point cloud data.

[0093] Here, when processing the point cloud data (such as coordinate normalization), since the data processing capacity of the hardware device is limited, it is necessary to determine whether the data volume of the point cloud data exceeds the data volume threshold, and the data volume threshold can be adapted to the data processing capacity. When it is determined that the data volume of the point cloud data exceeds the data volume threshold, the point cloud data can be sliced, for example, the point cloud data can be segmented according to the driving trajectory of the vehicle. Specifically, the driving trajectory of the vehicle is obtained, and the driving trajectory can be understood as consisting of a series of trajectory points. Each trajectory point can correspond to a frame of point cloud data, so that it can be sliced ​​according to the number of points of the trajectory point, such as slicing once every 1000 trajectory points; it can also be sliced ​​according to the length of the driving trajectory, such as slicing once every 10 meters. In this way, multiple point cloud data blocks can be obtained. Based on this, when processing the point cloud data, each point cloud data block can be processed separately, which can improve the processing efficiency of the point cloud data, thereby improving the efficiency of ground object detection, and can balance the existing hardware resources and the resources required for data processing, and improve the utilization of hardware resources.

[0094] Step 102: performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data.

[0095] In step 102, the acquired point cloud data may be subjected to coordinate normalization processing to obtain normalized point cloud data. In practical applications, the coordinate normalization processing is to normalize the point cloud data to a normalized coordinate system. For example, when the terminal is a vehicle-mounted terminal, the normalized coordinate system may be a vehicle body coordinate system, which is not limited here. In this way, the complexity of point cloud data processing may be reduced and the efficiency of ground object detection may be improved through coordinate normalization processing.

[0096] In some embodiments, the terminal can perform coordinate normalization processing on the point cloud data to obtain normalized point cloud data in the following manner: determine the original coordinate system in which the point cloud data is located, and construct a normalized coordinate system; obtain a coordinate transformation matrix that transforms from the original coordinate system to the normalized coordinate system; based on the coordinate transformation matrix, perform coordinate transformation processing on the point cloud data in the original coordinate system to obtain point cloud data in the normalized coordinate system; use the point cloud data in the normalized coordinate system as normalized point cloud data.

[0097] Here, when performing coordinate normalization processing, a normalized coordinate system can be pre-constructed. Then, the original coordinate system of the point cloud data is determined, and based on the original coordinate system and the normalized coordinate system, a coordinate transformation matrix for transforming from the original coordinate system to the normalized coordinate system is determined. Furthermore, based on the coordinate transformation matrix, coordinate transformation processing is performed on the point cloud data in the original coordinate system to obtain point cloud data in the normalized coordinate system, and the point cloud data in the normalized coordinate system is the normalized point cloud data.

[0098] Step 103: performing feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data.

[0099] In step 103, after the normalized point cloud data is obtained, feature extraction processing is performed on the normalized point cloud data to obtain point cloud features of the normalized point cloud data. In practical applications, the feature extraction processing process can be implemented by a feature extraction model, and the feature extraction model can be pre-trained and can be built based on a neural network (such as a convolutional neural network, a deep neural network, etc.).

[0100] In some embodiments, the terminal can perform feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data in the following manner: determine the three-dimensional point cloud space where the normalized point cloud data is located, and divide the three-dimensional point cloud space into multiple sub-point cloud spaces; for each sub-point cloud space, perform feature extraction processing on the sub-point cloud space to obtain spatial point cloud features of the sub-point cloud space; perform feature integration processing on multiple spatial point cloud features in the three-dimensional point cloud space to obtain point cloud features of the normalized point cloud data.

[0101] Here, the point cloud data is actually three-dimensional point cloud data, so the normalized point cloud data is also three-dimensional point cloud data. When extracting features from normalized point cloud data, the three-dimensional point cloud space where the normalized point cloud data is located can be first determined, and the three-dimensional point cloud space can be divided into multiple sub-point cloud spaces; then, for each sub-point cloud space, feature extraction processing can be performed on the sub-point cloud space to obtain spatial point cloud data of the sub-point cloud space. In actual implementation, for each sub-point cloud space, a data volume threshold for the data included in the sub-point cloud space can be set. If the data volume of the sub-point cloud space does not reach the data volume threshold, the data volume can be supplemented by filling 0 to ensure that the feature extraction processing of the data in the sub-point cloud space can be performed normally. In this way, the spatial point cloud features of each sub-point cloud space are obtained. The spatial point cloud features of each sub-point cloud space can be understood as local features of the point cloud data, that is, by extracting local features, the precision of feature extraction is improved, thereby improving the feature extraction effect.

[0102] Continuing, after obtaining the spatial point cloud features of each sub-point cloud space, the spatial point cloud features of each sub-point cloud space can be integrated to obtain the overall features of the point cloud data. Specifically, a plurality of the spatial point cloud features in the three-dimensional point cloud space can be subjected to three-dimensional convolution processing, thereby integrating the local relationship between the spatial point cloud features through three-dimensional convolution processing to expand the receptive field to obtain richer point cloud features of the point cloud data. In this way, the accuracy of feature extraction can be improved, so that the extracted point cloud features can better express the point cloud data, thereby improving the detection accuracy of ground object detection based on point cloud features.

[0103] Step 104: Based on the point cloud features, predict the vectorized detection result of the object to be detected to obtain the vectorized detection result of the object to be detected.

[0104] The vectorized detection result is used to describe the geometric representation of the object to be detected.

[0105] In step 104, after the point cloud features of the point cloud data are obtained, the vector detection result of the object to be detected can be predicted based on the point cloud data, thereby obtaining the vectorized detection result of the object to be detected. The vectorized detection result can be composed of geometric primitives such as points, line segments, polylines, curves, rectangles, and bounding boxes. Therefore, the vectorized detection result can be used to describe the geometric representation of the object to be detected. The vectorized detection result can also describe the position of the object to be detected, and based on the vectorized detection result, the shape, size and other information of the object to be detected can also be obtained.

[0106] In some embodiments, the above-mentioned feature extraction process and the prediction process of the vectorized detection result can be implemented by a pre-trained object detection model. That is, before the terminal performs feature extraction processing on the normalized point cloud data to obtain the point cloud features of the normalized point cloud data, it can also obtain the object detection model used for object detection; accordingly, the terminal can call the object detection model to perform feature extraction processing on the normalized point cloud data to obtain the point cloud features of the normalized point cloud data; and the terminal can also call the object detection model to predict the vectorized detection results of the object to be detected based on the point cloud features to obtain the vectorized detection results of the object to be detected.

[0107] In some embodiments, the terminal can train the ground object detection model in the following manner to obtain the ground object detection model used for ground object detection: obtain the initial ground object detection model, and the point cloud data samples of the ground object samples, the point cloud data samples are annotated with the vectorized detection result labels of the ground object samples; perform coordinate normalization processing on the point cloud data samples to obtain normalized point cloud data samples; call the initial ground object detection model, perform feature extraction processing on the normalized point cloud data samples, and obtain point cloud feature samples of the normalized point cloud data samples; call the initial ground object detection model, predict the vectorized detection results of the ground object samples, and obtain the predicted vectorized detection results of the ground object samples; based on the difference between the vectorized detection result labels and the predicted vectorized detection results, update the model parameters of the initial ground object detection model to train the initial ground object detection model to obtain the ground object detection model used for ground object detection.

[0108] Here, when training the object detection model, first obtain the constructed initial object detection model and the training data for training, that is, the point cloud data sample of the object sample, and the point cloud data sample is annotated with the vectorized detection result label of the object sample. When training the initial object detection model through the point cloud data sample, firstly perform coordinate normalization processing on the point cloud data sample to obtain the normalized point cloud data sample, and then call the initial object detection model to perform feature extraction processing on the normalized point cloud data sample to obtain the point cloud feature sample of the normalized point cloud data sample; and call the initial object detection model to predict the vectorized detection result of the object sample to obtain the predicted vectorized detection result of the object sample. Based on this, the model parameters of the initial object detection model can be updated according to the difference between the vectorized detection result label and the predicted vectorized detection result to train the initial object detection model.

[0109] In practical applications, the value of the loss function of the initial object detection model can be determined based on the difference between the vectorized detection result label and the predicted vectorized detection result, thereby updating the model parameters of the initial object detection model based on the value of the loss function. Specifically, when the value of the loss function exceeds a preset threshold, the error signal of the initial object detection model is determined based on the loss function, so that the error signal is back-propagated in the initial object detection model, and the model parameters of each layer in the initial object detection model are updated during the propagation process. In practical applications, the object detection model provided in the embodiment of the present application can be constructed based on a neural network. The type and structure of the neural network are not limited here. For example, a VoxelNet network structure can be used to implement the object detection model.

[0110] In some embodiments, before obtaining the initial ground object detection model and the point cloud data sample of the ground object sample, the terminal can obtain training data in the following manner: obtain a first point cloud data sample and a first vectorization detection result of the ground object sample; perform point cloud filtering processing on the first point cloud data sample to obtain a second point cloud data sample; perform slicing processing on the second point cloud data sample to obtain multiple point cloud data blocks, wherein the multiple point cloud data blocks constitute the point cloud data sample of the ground object sample; and associate the first vectorization detection result with the point cloud data sample to mark the first vectorization detection result as the vectorization detection result label of the point cloud data sample.

[0111] Here, when constructing training data, the terminal first obtains the first point cloud data sample of the ground object sample and the first vectorization detection result corresponding to the first point cloud data sample. Then, the first point cloud data sample is subjected to point cloud filtering processing to obtain a second point cloud data sample. The point cloud filtering processing can be to filter point cloud data that does not meet preset requirements, such as point cloud data located at the point cloud boundary and point cloud data with a point count lower than a quantity threshold. The second point cloud data sample is then sliced ​​to obtain a plurality of point cloud data blocks, which are the point cloud data samples of the ground object sample. Finally, the first vectorization detection result is associated with the point cloud data sample to label the point cloud data sample, that is, the first vectorization detection result is labeled as the vectorization detection result label of the point cloud data sample. In this way, a point cloud data sample carrying a vectorization detection result label is obtained.

[0112] In some embodiments, the terminal may also call the object detection model to predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; accordingly, the above-mentioned point cloud data samples may also be marked with the category label to which the object sample belongs. Based on this, the terminal may call the initial object detection model to predict the vectorized detection results of the object sample, and obtain the predicted category to which the object sample belongs; at this time, the terminal may update the model parameters of the initial object detection model based on the difference between the vectorized detection result label and the predicted vectorized detection result in the following manner: based on the difference between the vectorized detection result label and the predicted vectorized detection result, and the difference between the category label and the predicted category, update the model parameters of the initial object detection model.

[0113] Here, the object detection model can also be used to predict the object category of the object to be detected. That is, the object detection model can be called to predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs. Based on this, the point cloud data sample also needs to be labeled with the category label to which the object sample belongs. When training the object detection model, the initial object detection model can be called to predict the vectorized detection results of the object samples to obtain the predicted category to which the object samples belong. Correspondingly, the model parameters of the initial object detection model are updated based on the difference between the vectorized detection result label and the predicted vectorized detection result, as well as the difference between the category label and the predicted category.

[0114] In actual implementation, the value of the first loss function of the initial object detection model can be determined based on the difference between the vectorized detection result label and the predicted vectorized detection result, and the value of the second loss function of the initial object detection model can be determined based on the difference between the category label and the predicted category, and the sum of the value of the first loss function and the value of the second loss function is used as the value of the loss function of the initial object detection model. Therefore, the model parameters of the initial object detection model are updated based on the value of the loss function. Specifically, when the value of the loss function exceeds a preset threshold, the error signal of the initial object detection model is determined based on the loss function, so that the error signal is back-propagated in the initial object detection model, and the model parameters of each layer in the initial object detection model are updated during the propagation process.

[0115] In some embodiments, the terminal can perform category detection on the object to be detected in the following manner: extract target point cloud data associated with the vectorized detection result from the normalized point cloud data; obtain target point cloud features of the target point cloud data; and predict the object category of the object to be detected based on the target point cloud features to obtain the target category to which the object to be detected belongs.

[0116] Here, the object detection model can also be used to predict the object category of the object to be detected. In practical applications, 1) the object category of the object to be detected can be predicted directly based on the point cloud features to obtain the target category to which the object to be detected belongs. 2) The target point cloud data associated with the vectorized detection result can be extracted from the normalized point cloud data, and the target point cloud features of the target point cloud data can be obtained, so as to predict the object category of the object to be detected based on the target point cloud features, and obtain the target category to which the object to be detected belongs. The target point cloud data associated with the vectorized detection result is at least one of the following data: the point cloud data around the vectorized detection result, and the point cloud data covered by the vectorized detection result. For example, if the vectorized detection result is a line segment, the target point cloud data can be at least one of the following data: the point cloud data around the line segment, and the point cloud data covered by the line segment. In this way, the point cloud data of the object to be detected can be refined so that the extracted target point cloud features can better represent the object to be detected, thereby making the result of the category prediction more accurate. 3) First, based on the point cloud features, the object category of the object to be detected can be predicted to obtain the first category to which the object to be detected belongs; then, the target point cloud data associated with the vectorized detection result can be extracted from the normalized point cloud data, and the target point cloud features of the target point cloud data can be obtained, so as to predict the object category of the object to be detected based on the target point cloud features and obtain the second category to which the object to be detected belongs; the first category and the second category are compared, if the two are the same, then the object to be detected belongs to the first category, if the two are different, then the object to be detected belongs to the second category. In this way, the point cloud data of the object to be detected can also be refined, so that the extracted target point cloud features can better represent the object to be detected, thereby making the category prediction result more accurate.

[0117] In some embodiments, the terminal can predict the vectorized detection results of the objects to be detected based on the point cloud features in the following manner to obtain the vectorized detection results of the objects to be detected: convert the point cloud features into a point cloud feature map, which includes multiple feature points; for each feature point, determine a three-dimensional anchor point frame centered on the feature point; for each three-dimensional anchor point frame, perform regression prediction of the vectorized detection results of the three-dimensional anchor point frame to obtain an intermediate vectorized detection result; splice the intermediate vectorized detection results of each three-dimensional anchor point frame to obtain the vectorized detection result of the objects to be detected.

[0118] Here, when converting the point cloud features into a point cloud feature map, the point cloud features can be first converted into a first point cloud feature map, and then the first point cloud feature map is subjected to a multi-layer cascade downsampling operation to obtain a downsampled point cloud feature map outputted by each layer, and then the downsampled point cloud feature maps of each layer are respectively converted into target point cloud feature maps having a target size, and finally the target point cloud feature maps of each layer are spliced ​​to obtain the point cloud feature map. When determining a three-dimensional anchor point frame (with a size of preset length, width, and height), for each feature point, a plurality of candidate three-dimensional anchor point frames centered on the feature point can be determined, and then for each candidate three-dimensional anchor point frame, the candidate three-dimensional anchor point frame is predicted to obtain a prediction result for characterizing whether the candidate three-dimensional anchor point frame includes the object to be detected, and the confidence of the prediction result is outputted at the same time, so that the candidate three-dimensional anchor point frame characterized by the prediction result including the object to be detected and having the highest confidence is used as the three-dimensional anchor point frame of the feature point. Then, for each 3D anchor point box, a regression prediction of the vectorized detection result is performed on the 3D anchor point box to obtain an intermediate vectorized detection result. Finally, the intermediate vectorized detection results of each 3D anchor point box are spliced ​​to obtain the vectorized detection result of the object to be detected.

[0119] See also Figure 5 , Figure 5 It is a schematic diagram of the model structure of the object detection model provided in the embodiment of the present application. Here, the object detection model includes a local feature extraction layer, a feature integration layer and a feature prediction layer, and the feature prediction layer includes a vectorized detection result prediction layer and an object category prediction layer. Among them, 1) the local feature extraction layer is used to determine the three-dimensional point cloud space where the normalized point cloud data is located, and divide the three-dimensional point cloud space into multiple sub-point cloud spaces. For each sub-point cloud space, the sub-point cloud space is subjected to feature extraction processing to obtain the spatial point cloud features of the sub-point cloud space. 2) the feature integration layer is used to perform feature integration processing on multiple spatial point cloud features in the three-dimensional point cloud space to obtain the point cloud features of the normalized point cloud data. 3) the vectorized detection result prediction layer is used to predict the vectorized detection results of the objects to be detected based on the point cloud features, and obtain the vectorized detection results of the objects to be detected. 4) the object category prediction layer is used to predict the object category of the objects to be detected based on the point cloud features, and obtain the category to which the objects to be detected belong.

[0120] In some embodiments, the object detection method provided in the embodiments of the present application can be applied to autonomous driving scenarios. In practical applications, the terminal can also predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; based on this, the terminal can control the autonomous driving in the following ways: obtain the vehicle position of the autonomous driving vehicle; when the vehicle position is within the target geographical range, based on the vehicle position, combined with the vectorization detection results and the target category, generate an autonomous driving route; according to the autonomous driving route, control the autonomous driving vehicle to travel. Here, when applied to the autonomous driving scenario, an autonomous driving route can be generated based on the vehicle position, combined with the vectorization detection results and the target category, so as to control the autonomous driving vehicle to travel according to the autonomous driving route. In this way, based on the vectorization detection results, the scene understanding of autonomous driving can be achieved and the safety of autonomous driving can be improved.

[0121] In some embodiments, the object detection method provided in the embodiments of the present application can be applied to map generation scenarios. In practical applications, the terminal can also predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; based on this, the terminal can generate a map in the following manner: based on the vectorized detection results and the target category, generate a regional map of the target geographical area. Here, the vectorized detection results can be used for the element representation of the corresponding object elements in the map (such as a high-precision map, a standard-precision map), thereby realizing map production based on the vectorized detection results.

[0122] In some embodiments, the ground feature to be detected is the tunnel entrance of a tunnel, the point cloud data is collected when a vehicle passes through at least one tunnel, and the vectorized detection result includes the vectorized detection results of each tunnel entrance of at least one tunnel; based on this, the terminal can also predict the category of each tunnel entrance based on the point cloud features, and obtain the tunnel entrance category to which the tunnel entrance belongs, and the tunnel entrance category includes the tunnel exit and the tunnel entrance; according to the tunnel entrance category of each tunnel entrance, the tunnel entrance and exit of each tunnel entrance are matched to obtain at least one tunnel entrance and exit pair; and the tunnel indicated by each tunnel entrance and exit pair is determined.

[0123] Here, when the object to be detected is the tunnel entrance of a tunnel, the point cloud data is collected during the process of the vehicle passing through at least one tunnel. Based on this, the predicted vector detection results include: the vector detection results of each tunnel entrance of at least one tunnel. At the same time, the terminal can also predict the category of each tunnel entrance based on the point cloud features, and obtain the tunnel entrance category to which the tunnel entrance belongs, that is, predict whether the tunnel entrance is a tunnel exit or a tunnel entrance. After determining the tunnel entrance category of each tunnel entrance, the tunnel entrance and exit of each tunnel entrance can also be matched, that is, for the tunnel entrance, the tunnel exit closest to the tunnel entrance is determined, and the tunnel exit and the tunnel exit closest to the tunnel entrance can form a tunnel entrance and exit pair. Match each tunnel entrance in turn to obtain at least one tunnel entrance and exit pair. Each tunnel entrance and exit pair corresponds to a section of a tunnel. In this way, tunnel identification is achieved by detecting tunnel entrances and exits.

[0124] Using the above-mentioned embodiment of the present application, firstly, coordinate normalization processing is performed on the point cloud data of the target geographic range to obtain normalized point cloud data, and then feature extraction processing is performed on the normalized point cloud data to obtain point cloud features of the normalized point cloud data, thereby predicting the vectorized detection results of the objects to be detected based on the point cloud features, and obtaining the vectorized detection results of the objects to be detected, which are used to describe the geometric representation of the objects to be detected. In this way, based on the point cloud features of the point cloud data of the target geographic range (including the objects to be detected), the vectorized detection results of the objects to be detected can be directly predicted, which reduces the complexity of the object detection processing and improves the detection efficiency of the object detection and the utilization rate of the hardware processing resources.

[0125] The following will take a tunnel as an example to illustrate the exemplary application of the present invention in a practical application scenario. In the related art, the detection of tunnels is achieved in the following way: first, each point in the point cloud sample collected for the tunnel is annotated to generate a point cloud sample carrying semantic information (such as Figure 6 As shown in the figure, different semantic information means different point cloud colors), then the point cloud samples carrying semantic information are input into the deep learning network model for learning, and after the training is completed, the trained model is used to predict the collected point cloud to obtain a point cloud carrying predicted semantic information, and then the point cloud carrying semantic information is clustered, segmented, denoised, fitted, etc., to obtain the vectorized detection result of the tunnel. However, the related technology has the following disadvantages: (1) Each point in the point cloud sample is labeled, which is difficult, time-consuming and costly. (2) After the deep learning model is learned, the prediction result of the model is a point cloud carrying semantic information, so it is necessary to go through multiple complex processing steps (including clustering, segmentation, denoising, fitting, etc.) to generate a vectorized result, resulting in low efficiency of tunnel detection.

[0126] Based on this, an embodiment of the present application provides a method for detecting an object to at least solve the above-mentioned problems. In an embodiment of the present application, a point cloud marked with vectorization results is used as training input data for an object detection model (such as a deep learning model) to train the object detection model. After training and learning, the trained object detection model can directly output the vectorization results of the tunnel based on the point cloud. In this way, through the object detection method provided by an embodiment of the present application, 1) deep learning is used to directly process the point cloud of the tunnel, and the vectorization results of the tunnel are output, thereby achieving the effect of direct geometric vectorization output of large objects (such as tunnels). 2) The tunnel is described as a tunnel entrance and a tunnel exit, and the detection of the large object as a whole is achieved by detecting the positions of the tunnel entrance and the tunnel exit. 3) It can be applied to the case of few samples. First, a preliminary prediction is made in the first stage (for example, through deep learning) to obtain a vectorized result and a preliminary category prediction result. Then, in the second stage, the point cloud around the vectorized result is extracted, and then the extracted point cloud is finely classified and predicted (for example, using machine learning) to obtain the final category prediction result, thereby filtering out falsely detected objects and making the category prediction result more accurate.

[0127] In practical applications, the object detection method provided by the embodiments of the present application can be applied to map production, intelligent driving projects and product applications, and can optimize the annotation process of point cloud data, the map production automation process, and the intelligent driving perception process. For example, in map production, 1) when annotating data in a geometric shape, or directly reusing the map operation result data, and training the deep learning model with the point cloud, it has the characteristics of simplified annotation, and if the map operation results are reused, even no annotation is required, which further saves annotation costs and simplifies the annotation method. 2) The output result of the object detection model is directly the vectorization result of the map, which simplifies the map production process. For another example, in intelligent driving, the deep learning annotation method is also simplified, and the method of directly outputting the vectorization result is characterized by high efficiency, data-driven, and accurate driving.

[0128] Next, the training process and application process of the ground object detection model involved in the ground object detection method provided in the embodiment of the present application are described. Here, the ground object detection model is used for tunnel detection. In the embodiment of the present application, the tunnel structure is defined as follows: Figure 7 As shown, along the driving direction, the tunnel exit and tunnel entrance are defined (such as Figure 7 As shown in (1), for the tunnel exit and tunnel entrance, the tunnel is represented by n segments of broken lines (as shown in Figure 7 As shown in (2), the size of n depends on the accuracy. In this way, large objects with variable sizes such as tunnels are decomposed into exits and entrances, and each entrance is accurately expressed by n segments of polylines. This structural definition is the basis for subsequent processing.

[0129] See also Figure 8 , Figure 8 is a training flow chart of the ground object detection model provided in the embodiment of the present application. Here, the training process of the ground object detection model includes:

[0130] (1) Collect the point cloud within the sample detection range and obtain the vectorization result corresponding to the point cloud.

[0131] (2) Data preprocessing of point clouds and vectorization results. Here, data preprocessing may include: a) association: associating point clouds with vectorization results to achieve the annotation of vectorization results of point cloud data; b) filtering: removing point clouds that do not meet the requirements (for example, the number of points is less than the point threshold or located at the boundary); c) slicing: slicing the point cloud according to the vehicle's driving direction to obtain a point cloud block sequence.

[0132] (3) performing coordinate transformation on the point cloud coordinates and the annotation data coordinates (i.e., the vectorization results annotated on the point cloud) to transform the point cloud coordinates and the annotation data coordinates into the vehicle body coordinate system.

[0133] (4) Determine the true value category of the tunnel entrance. Tunnels are divided into two types: exit and entrance. According to the direction of vehicle travel, the entrance to the tunnel is defined as the tunnel entrance, and the exit from the tunnel is defined as the tunnel exit.

[0134] (5) Determine the true vectorization result of the tunnel entrance. The tunnel exit and tunnel entrance are represented by geometric vectors, that is, the annotation data is converted into a polyline of n points, where the value of n depends on the required accuracy.

[0135] (6) The point cloud is predicted and processed by the object detection model to obtain a predicted value. The object detection model can be constructed based on a convolutional neural network. When the point cloud is processed by the object detection model, the predicted vectorization result of the tunnel entrance is predicted in the regression prediction branch, and the predicted category of the tunnel entrance (including the predicted entrance category and the predicted exit category) is predicted in the classification prediction branch.

[0136] (7) Calculate the loss. Based on the difference between the true value vectorization result and the predicted vectorization result, calculate the regression loss. Based on the difference between the true value category and the predicted category, calculate the category loss. Then add the regression loss and category loss to get the total loss of the object detection model. Thus, the object detection model is trained based on the total loss of the object detection model.

[0137] At this point, the next round of data iteration and model optimization is carried out until the training end conditions are met to obtain the final trained object detection model.

[0138] See also Fig. 9 , Fig. 9 is an application flow chart of the ground object detection model provided in the embodiment of the present application. Here, the application flow of the ground object detection model includes:

[0139] (1) Collect the point cloud of the range to be detected.

[0140] (2) Performing data preprocessing on the point cloud. Here, data preprocessing includes slicing, that is, the point cloud is sliced ​​according to the vehicle's driving direction to obtain a point cloud block sequence.

[0141] (3) Convert the point cloud coordinates to the vehicle body coordinate system.

[0142] (4) Processing the point cloud by using a ground object detection model. The ground object detection model can be constructed based on a convolutional neural network. When processing the point cloud using the ground object detection model, the predicted vectorization result of the tunnel entrance (i.e., the vectorization detection result) is predicted in the regression prediction branch, and the predicted category of the tunnel entrance (including the predicted entrance category and the predicted exit category) is predicted in the classification prediction branch.

[0143] like Figure 1C As shown, the line segment in the figure is the vectorization result of the tunnel (the figure is the most concise expression of two points). The vectorization result can be used to identify tunnel scenes and high-precision locations, such as the width of the tunnel entrance and the location of entering and exiting the tunnel in the driving direction, so as to help the autonomous driving vehicle understand the scene and drive safely when applied to autonomous driving.

[0144] (5) Matching of tunnel exits and entrances. After predicting the category of each tunnel entrance, match the tunnel entrances and exits to obtain the starting and ending positions of each tunnel section. Fig.10 , Fig.10 This is a simplified schematic diagram of a tunnel exit and a tunnel entrance provided by an embodiment of the present application. Here, multiple tunnel entrances may be detected in the point cloud data collected during one trip. Each tunnel section has a tunnel entrance and a tunnel exit. Therefore, the principle of matching entrances and exits is: one tunnel entrance matches the nearest tunnel exit. Iterate sequentially until all tunnel exits and tunnel entrances are matched, so that different sections of tunnels can be distinguished. In this way, tunnel detection can be completed.

[0145] (6) Refined classification. If the training data is insufficient, in order to improve the overall recognition ability of the model, high-recall extraction can be performed in the first stage (the above step (5)), and then machine learning methods (such as classification networks or statistical methods) can be used to filter approximate tunnels. When predicting in step (5), the approximate tunnel features may be mistakenly identified as tunnel entrances. Since the first-stage detection focuses on the existence of the tunnel entrance, it is not enough to distinguish some complex tunnels or suspected tunnels, so it is necessary to add a second stage of detection. The second stage of detection (i.e., refined classification) is to re-extract the point cloud around the tunnel entrance, and then perform secondary sub-classification based on the extracted point cloud to achieve the purpose of accurate distinction.

[0146] In practical applications, the object detection model provided in the embodiment of the present application can be implemented using the VoxelNet network structure; however, the embodiment of the present application does not specifically limit the model type and topological structure of the object detection model, and it can be replaced by various other effective model structures, such as other convolutional neural network structures.

[0147] By applying the above embodiments of the present application, 1) the sample method can be simplified, and the training efficiency of the model can be improved, thereby improving the detection efficiency. 2) It supports the detection of super-large objects (such as tunnels). For large and curved objects like tunnels (such as those with a length of about 30m or more), a single bounding box may not be able to accurately describe them. The vectorization results in the present application are composed of geometric primitives such as line segments, points, broken lines, and curves, which can accurately describe the geometric representation of super-large objects; and for large objects, detection can be completed in real time and efficiently. 3) The vectorization results of the objects can be accurately and directly output based on the point cloud, without the need for multiple complex processing steps (including clustering, segmentation, denoising, fitting, etc.) on the point cloud, and the running speed is faster, further improving the detection efficiency. 4) Through refined classification, tunnels can be accurately identified with high precision.

[0148] The following is a description of an exemplary structure of the ground object detection device 555 provided in the embodiment of the present application implemented as a software module. In some embodiments, Figure 3 As shown, the software modules stored in the object detection device 555 of the memory 550 may include: an acquisition module 5551, used to acquire point cloud data of a target geographical range, wherein the target geographical range includes the objects to be detected; a normalization module 5552, used to perform coordinate normalization processing on the point cloud data to obtain normalized point cloud data; a feature extraction module 5553, used to perform feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data; a prediction module 5554, used to predict the vectorized detection results of the objects to be detected based on the point cloud features to obtain the vectorized detection results of the objects to be detected, wherein the vectorized detection results are used to describe the geometric representation of the objects to be detected.

[0149] In some embodiments, the acquisition module 5551 is also used to transmit a detection signal to the target geographical range; receive a reflection signal reflected by the target within the target geographical range in response to the detection signal; determine the three-dimensional position information of the target within the target geographical range based on the detection signal and the reflection signal; and use the three-dimensional position information of the target within the target geographical range as point cloud data of the target geographical range.

[0150] In some embodiments, the normalization module 5552 is also used to determine the original coordinate system in which the point cloud data is located and construct a normalized coordinate system; obtain a coordinate transformation matrix for transforming from the original coordinate system to the normalized coordinate system; based on the coordinate transformation matrix, perform coordinate transformation processing on the point cloud data in the original coordinate system to obtain point cloud data in the normalized coordinate system; and use the point cloud data in the normalized coordinate system as the normalized point cloud data.

[0151] In some embodiments, the acquisition module 5551 is also used to perform multiple point cloud acquisitions on the target geographic range according to the point cloud acquisition frequency through the vehicle's on-board point cloud acquisition equipment during the vehicle's movement, so as to obtain multiple frames of first sub-point cloud data; convert each frame of the first sub-point cloud data to the target coordinate system, so as to obtain multiple frames of second sub-point cloud data under the target coordinates; and splice multiple frames of the second sub-point cloud data to obtain the point cloud data of the target geographic range.

[0152] In some embodiments, the normalization module 5552 is also used to obtain the data volume of the point cloud data; when the data volume exceeds a data volume threshold, the driving trajectory of the vehicle is obtained, and according to the driving trajectory, based on a data slicing method, the point cloud data is sliced ​​to obtain a plurality of point cloud data blocks; the normalization module 5552 is also used to perform coordinate normalization processing on each point cloud data block respectively to obtain normalized point cloud data.

[0153] In some embodiments, the acquisition module 5551 is also used to acquire a ground object detection model for ground object detection; the feature extraction module 5553 is also used to call the ground object detection model, perform feature extraction processing on the normalized point cloud data, and obtain point cloud features of the normalized point cloud data; the prediction module 5554 is also used to call the ground object detection model, predict the vectorized detection results of the ground object to be detected based on the point cloud features, and obtain the vectorized detection results of the ground object to be detected.

[0154] In some embodiments, the acquisition module 5551 is also used to acquire an initial ground object detection model, and point cloud data samples of ground object samples, wherein the point cloud data samples are annotated with vectorized detection result labels of the ground object samples; coordinate normalization is performed on the point cloud data samples to obtain normalized point cloud data samples; the initial ground object detection model is called to perform feature extraction on the normalized point cloud data samples to obtain point cloud feature samples of the normalized point cloud data samples; the initial ground object detection model is called to predict the vectorized detection results of the ground object samples to obtain predicted vectorized detection results of the ground object samples; based on the difference between the vectorized detection result label and the predicted vectorized detection result, the model parameters of the initial ground object detection model are updated to obtain the ground object detection model used for ground object detection.

[0155] In some embodiments, the acquisition module 5551 is also used to acquire a first point cloud data sample and a first vectorization detection result of the ground object sample; perform point cloud filtering processing on the first point cloud data sample to obtain a second point cloud data sample; perform slicing processing on the second point cloud data sample to obtain multiple point cloud data blocks, wherein the multiple point cloud data blocks constitute the point cloud data sample of the ground object sample; and associate the first vectorization detection result with the point cloud data sample to mark the first vectorization detection result as the vectorization detection result label of the point cloud data sample.

[0156] In some embodiments, the prediction module 5554 is also used to call the object detection model, predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; the point cloud data sample is also marked with the category label to which the object sample belongs, and the acquisition module 5551 is also used to call the initial object detection model, predict the vectorized detection result of the object sample, and obtain the predicted category to which the object sample belongs; the acquisition module 5551 is also used to update the model parameters of the initial object detection model based on the difference between the vectorized detection result label and the predicted vectorized detection result, and the difference between the category label and the predicted category.

[0157] In some embodiments, the method is applied to an autonomous driving scenario; the prediction module 5554 is further used to predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; the prediction module 5554 is further used to obtain the vehicle position of the autonomous driving vehicle; when the vehicle position is within the target geographical range, an autonomous driving route is generated based on the vehicle position, combined with the vectorized detection result and the target category; and the autonomous driving vehicle is controlled to travel according to the autonomous driving route.

[0158] In some embodiments, the prediction module 5554 is also used to predict the object category of the object to be detected based on the point cloud features, and obtain the target category to which the object to be detected belongs; the prediction module 5554 is also used to generate a regional map of the target geographical area based on the vectorized detection results and the target category.

[0159] In some embodiments, the prediction module 5554 is also used to extract target point cloud data associated with the vectorized detection result from the normalized point cloud data; obtain target point cloud features of the target point cloud data; and predict the object category of the object to be detected based on the target point cloud features to obtain the target category to which the object to be detected belongs.

[0160] In some embodiments, the feature extraction module 5553 is also used to determine the three-dimensional point cloud space where the normalized point cloud data is located, and divide the three-dimensional point cloud space into multiple sub-point cloud spaces; for each sub-point cloud space, perform feature extraction processing on the data of the sub-point cloud space to obtain the spatial point cloud features of the sub-point cloud space; perform feature integration processing on the multiple spatial point cloud features in the three-dimensional point cloud space to obtain the point cloud features of the normalized point cloud data.

[0161] In some embodiments, the prediction module 5554 is also used to convert the point cloud features into a point cloud feature map, which includes multiple feature points; for each of the feature points, determine a three-dimensional anchor point frame centered on the feature point; for each of the three-dimensional anchor point frames, perform regression prediction on the vectorization detection results of the three-dimensional anchor point frames to obtain intermediate vectorization detection results; and splice the intermediate vectorization detection results of each of the three-dimensional anchor point frames to obtain the vectorization detection results of the object to be detected.

[0162] In some embodiments, the ground feature to be detected is the tunnel entrance of a tunnel, the point cloud data is collected when a vehicle passes through at least one of the tunnels, and the vectorized detection result includes the vectorized detection result of each of the tunnel entrances of at least one of the tunnels; the prediction module 5554 is also used to predict the category of each of the tunnel entrances based on the point cloud features, and obtain the tunnel entrance category to which the tunnel entrance belongs, and the tunnel entrance category includes a tunnel exit and a tunnel entrance; according to the tunnel entrance category of each of the tunnel entrances, the tunnel entrances and exits of each of the tunnel entrances are matched to obtain at least one tunnel entrance and exit pair; and the tunnel indicated by each of the tunnel entrance and exit pairs is determined.

[0163] Using the above-mentioned embodiment of the present application, firstly, coordinate normalization processing is performed on the point cloud data of the target geographic range to obtain normalized point cloud data, and then feature extraction processing is performed on the normalized point cloud data to obtain point cloud features of the normalized point cloud data, thereby predicting the vectorized detection results of the objects to be detected based on the point cloud features, and obtaining the vectorized detection results of the objects to be detected, which are used to describe the geometric representation of the objects to be detected. In this way, based on the point cloud features of the point cloud data of the target geographic range (including the objects to be detected), the vectorized detection results of the objects to be detected can be directly predicted, which reduces the complexity of the object detection processing and improves the detection efficiency of the object detection and the utilization rate of the hardware processing resources.

[0164] The embodiment of the present application also provides a computer program product, which includes computer executable instructions or computer programs, and the computer executable instructions or computer programs are stored in a computer readable storage medium. The processor of the electronic device reads the computer executable instructions or computer program from the computer readable storage medium, and the processor executes the computer executable instructions or computer program, so that the electronic device performs the ground object detection method provided in the embodiment of the present application.

[0165] An embodiment of the present application also provides a computer-readable storage medium, which stores computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the ground object detection method provided in the embodiment of the present application.

[0166] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or may be various devices including one or any combination of the above memories.

[0167] In some embodiments, computer executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0168] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0169] As an example, computer executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed at multiple sites and interconnected by a communication network.

[0170] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A method for detecting ground objects, characterized in that: The method comprises: Acquire point cloud data of a target geographic range, wherein the target geographic range includes a ground object to be detected; Performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data; Performing feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data; Based on the point cloud features, predicting a vectorized detection result of the object to be detected, to obtain a vectorized detection result of the object to be detected, wherein the vectorized detection result is used to describe a geometric representation of the object to be detected; Based on the point cloud features, predicting the category of the to-be-detected object, and obtaining a first category to which the to-be-detected object belongs; Extracting target point cloud data associated with the vectorized detection result from the normalized point cloud data, and predicting the object category of the object to be detected based on the target point cloud features of the target point cloud data to obtain a second category to which the object to be detected belongs; If the first category and the second category are the same, it is determined that the to-be-detected ground object belongs to the first category; if the first category and the second category are different, it is determined that the to-be-detected ground object belongs to the second category.

2. The method according to claim 1, characterized in that The step of obtaining point cloud data of a target geographical range includes: Transmitting a detection signal within the target geographical range; receiving a reflection signal reflected by a target within the target geographical range in response to the detection signal; Determine three-dimensional position information of a target within the target geographic range based on the detection signal and the reflection signal; The three-dimensional position information of the target within the target geographical range is used as the point cloud data of the target geographical range.

3. The method according to claim 1, characterized in that The step of performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data includes: Determine the original coordinate system of the point cloud data and construct a normalized coordinate system; Obtaining a coordinate transformation matrix from the original coordinate system to the normalized coordinate system; Based on the coordinate transformation matrix, coordinate transformation processing is performed on the point cloud data in the original coordinate system to obtain point cloud data in the normalized coordinate system; The point cloud data in the normalized coordinate system is used as the normalized point cloud data.

4. The method according to claim 1, characterized in that The step of obtaining point cloud data of a target geographical range includes: During the movement of the vehicle, the on-board point cloud acquisition device of the vehicle performs multiple point cloud acquisitions on the target geographical range according to the point cloud acquisition frequency to obtain multiple frames of first sub-point cloud data; Convert each frame of the first sub-point cloud data to the target coordinate system to obtain multiple frames of second sub-point cloud data under the target coordinate system; Multiple frames of the second sub-point cloud data are spliced ​​to obtain the point cloud data of the target geographical range.

5. The method according to claim 4, characterized in that Before performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data, the method further includes: Acquire the data volume of the point cloud data; When the data volume exceeds the data volume threshold, the driving trajectory of the vehicle is obtained, and According to the driving trajectory, based on the data slicing method, the point cloud data is sliced ​​to obtain a plurality of point cloud data blocks; The performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data includes: performing coordinate normalization processing on each point cloud data block respectively to obtain normalized point cloud data.

6. The method according to claim 5, characterized in that According to the driving trajectory, based on the data slicing method, the point cloud data is sliced ​​to obtain a plurality of point cloud data blocks, including: Acquire a plurality of trajectory points included in the driving trajectory, each of the trajectory points corresponding to a frame of the point cloud data; According to the number of trajectory points, the point cloud data corresponding to the multiple trajectory points are sliced ​​to obtain multiple point cloud data blocks; or according to the length of the driving trajectory, the point cloud data corresponding to the multiple trajectory points are sliced ​​to obtain multiple point cloud data blocks.

7. The method according to claim 1, characterized in that Before performing feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data, the method further includes: Obtaining a ground object detection model for ground object detection; The step of performing feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data includes: Calling the ground object detection model to perform feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data; The predicting of the vectorized detection result of the to-be-detected object based on the point cloud feature to obtain the vectorized detection result of the to-be-detected object includes: The ground object detection model is called, and based on the point cloud features, a vectorized detection result of the ground object to be detected is predicted to obtain a vectorized detection result of the ground object to be detected.

8. The method according to claim 7, characterized in that The obtaining of the ground object detection model for ground object detection includes: Acquire an initial ground object detection model and a point cloud data sample of a ground object sample, wherein the point cloud data sample is annotated with a vectorized detection result label of the ground object sample; Performing coordinate normalization processing on the point cloud data samples to obtain normalized point cloud data samples; Calling the initial ground object detection model, performing feature extraction processing on the normalized point cloud data sample, and obtaining a point cloud feature sample of the normalized point cloud data sample; Calling the initial ground object detection model to predict the vectorized detection result of the ground object sample to obtain the predicted vectorized detection result of the ground object sample; Based on the difference between the vectorized detection result label and the predicted vectorized detection result, the model parameters of the initial ground object detection model are updated to obtain the ground object detection model used for ground object detection.

9. The method according to claim 8, characterized in that Before obtaining the initial ground object detection model and the point cloud data samples of the ground object samples, the method further includes: Acquire a first point cloud data sample and a first vectorization detection result of the ground object sample; Performing point cloud filtering processing on the first point cloud data sample to obtain a second point cloud data sample; Slicing the second point cloud data sample to obtain a plurality of point cloud data blocks, wherein the plurality of point cloud data blocks constitute the point cloud data sample of the ground object sample; The first vectorization detection result is associated with the point cloud data sample to mark the first vectorization detection result as a vectorization detection result label of the point cloud data sample.

10. The method according to claim 8, characterized in that The point cloud data sample is also marked with a category label to which the ground object sample belongs. After the feature extraction process is performed on the normalized point cloud data sample to obtain a point cloud feature sample of the normalized point cloud data sample, the method further includes: Calling the initial ground object detection model, predicting the vectorized detection result of the ground object sample, and obtaining the prediction category to which the ground object sample belongs; The updating of the model parameters of the initial object detection model based on the difference between the vectorized detection result label and the predicted vectorized detection result includes: Based on the difference between the vectorized detection result label and the predicted vectorized detection result, and the difference between the category label and the predicted category, the model parameters of the initial ground object detection model are updated.

11. The method according to claim 1, characterized in that The method is applied to an autonomous driving scenario; the method further includes: Obtaining the vehicle position of the autonomous vehicle; When the vehicle position is within the target geographical range, generating an automatic driving route based on the vehicle position, in combination with the vectorized detection result and the category to which the object to be detected belongs; Control the autonomous driving vehicle to travel according to the autonomous driving route.

12. The method according to claim 1, characterized in that The method further comprises: Based on the vectorized detection result and the category to which the to-be-detected object belongs, a regional map of the target geographical area is generated.

13. The method according to claim 1, characterized in that The step of performing feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data includes: Determine the three-dimensional point cloud space where the normalized point cloud data is located, and divide the three-dimensional point cloud space into a plurality of sub-point cloud spaces; For each of the sub-point cloud spaces, performing feature extraction processing on the sub-point cloud space to obtain spatial point cloud features of the sub-point cloud space; Feature integration processing is performed on the plurality of spatial point cloud features in the three-dimensional point cloud space to obtain point cloud features of the normalized point cloud data.

14. The method according to claim 1, wherein: The predicting of the vectorized detection result of the to-be-detected object based on the point cloud feature to obtain the vectorized detection result of the to-be-detected object includes: Converting the point cloud features into a point cloud feature map, wherein the point cloud feature map includes a plurality of feature points; For each of the feature points, determine a three-dimensional anchor point frame centered on the feature point; For each of the three-dimensional anchor point boxes, performing regression prediction on the vectorized detection result of the three-dimensional anchor point box to obtain an intermediate vectorized detection result; The intermediate vectorized detection results of each of the three-dimensional anchor point frames are spliced ​​to obtain the vectorized detection result of the object to be detected.

15. The method according to claim 1, wherein: The ground object to be detected is a tunnel entrance of a tunnel, the point cloud data is collected when a vehicle passes through at least one of the tunnels, the vectorized detection result includes the vectorized detection result of each tunnel entrance of at least one of the tunnels, and the categories to which the tunnel entrance belongs include a tunnel exit and a tunnel entrance; The method further comprises: Matching the tunnel entrances and exits of the tunnel entrances according to the categories to which the tunnel entrances belong, to obtain at least one tunnel entrance and exit pair; The tunnels indicated by the tunnel entrance and exit pairs are determined.

16. The method according to claim 15, characterized in that The vectorized detection result of the tunnel entrance includes: a geometric figure composed of n segments of broken lines and used to indicate the tunnel entrance; The matching of tunnel entrances and exits of each tunnel entrance according to the category to which each tunnel entrance belongs, to obtain at least one tunnel entrance and exit pair, comprises: According to the driving direction of the vehicle and based on the categories to which the tunnel entrances belong, the target tunnel entrances that the vehicle passes through are determined in sequence; For each of the determined target tunnel entrances, determine the tunnel entrance closest to the target tunnel entrance as the target tunnel exit corresponding to the target tunnel entrance, and construct the tunnel entrance-exit pair based on the target tunnel entrance and the target tunnel exit; The determining of the tunnels indicated by the tunnel entrances and exits comprises: For each tunnel entrance and exit pair, the following processing is performed: Based on the target tunnel entrance and the target tunnel exit included in the tunnel entrance-exit pair, the starting position and the ending position of the tunnel indicated by the tunnel entrance-exit pair are determined to detect the tunnel indicated by the tunnel entrance-exit pair.

17. A ground object detection device, characterized in that: The device comprises: An acquisition module, used to acquire point cloud data of a target geographical range, wherein the target geographical range includes the ground object to be detected; A normalization module, used for performing coordinate normalization processing on the point cloud data to obtain normalized point cloud data; A feature extraction module, used to perform feature extraction processing on the normalized point cloud data to obtain point cloud features of the normalized point cloud data; A prediction module, used for predicting a vectorized detection result of the object to be detected based on the point cloud features, to obtain a vectorized detection result of the object to be detected, wherein the vectorized detection result is used to describe a geometric representation of the object to be detected; The prediction module is also used to predict the object category of the object to be detected based on the point cloud features, and obtain the first category to which the object to be detected belongs; extract the target point cloud data associated with the vectorized detection result from the normalized point cloud data, and predict the object category of the object to be detected based on the target point cloud features of the target point cloud data, and obtain the second category to which the object to be detected belongs; if the first category and the second category are the same, it is determined that the object to be detected belongs to the first category; if the first category and the second category are different, it is determined that the object to be detected belongs to the second category.

18. An electronic device, characterized in that: The electronic device comprises: A memory for storing computer executable instructions; A processor, used to implement the ground object detection method described in any one of claims 1 to 16 when executing the computer executable instructions stored in the memory.

19. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that: When the computer executable instructions or computer program are executed by a processor, the ground object detection method described in any one of claims 1 to 16 is implemented.

20. A computer program product comprising computer executable instructions or a computer program, characterized in that When the computer executable instructions or computer program are executed by a processor, the ground object detection method described in any one of claims 1 to 16 is implemented.

Images