A method and system for constructing occupancy grid maps based on dynamic environment

Abstract

This application discloses a method and system for constructing an occupancy grid map based on a dynamic environment, relating to the field of autonomous navigation and environmental perception for mobile robots. The method includes acquiring environmental point cloud data and mobile robot pose information; preprocessing the environmental point cloud data based on the mobile robot pose information to obtain preprocessed environmental point cloud data, which is then input into a deep neural network model for semantic classification to obtain semantic labels corresponding to the preprocessed environmental point cloud data; using a particle filter algorithm to track and predict dynamic objects to obtain their position and motion state information; and dividing and updating the preprocessed environmental point cloud data into regions based on the semantic labels and the position and motion state information of the dynamic objects to obtain a real-time occupancy grid map. This application enables the construction of an efficient and real-time occupancy grid map in a dynamic environment.

Description

Technical Field

[0001] This application relates to the field of autonomous navigation and environmental perception for mobile robots, and in particular to a method and system for constructing an occupancy grid map based on a dynamic environment. Background Technology

[0002] In the field of autonomous navigation and environmental perception for mobile robots, constructing accurate and real-time occupancy grid maps is fundamental to achieving efficient path planning and obstacle avoidance. Occupancy grid maps provide a spatial representation of the environment for mobile robots by dividing the environment into discrete grid cells and estimating the occupancy state of each grid cell (e.g., "occupied" or "idle"). However, as application scenarios expand from static to dynamic environments, traditional map-building methods face numerous challenges. Dynamic environments contain a large number of moving objects (such as pedestrians and vehicles), whose positions and motion states are constantly changing, making it difficult for traditional methods to accurately track and update map information. Furthermore, dynamic environments place higher demands on the system's real-time performance, computational efficiency, and resource utilization. Therefore, how to construct efficient and accurate occupancy grid maps in dynamic environments has become a current research hotspot and challenge.

[0003] Existing technologies for constructing occupancy grid maps in dynamic environments suffer from several drawbacks, including insufficient dynamic object recognition capabilities, low real-time performance and computational efficiency, inadequate utilization of semantic information, high storage and computational resource consumption, and unreasonable dynamic and static update strategies. These issues limit the effectiveness of traditional methods in complex dynamic environments. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for constructing occupied grid maps based on dynamic environments, which can realize the construction of efficient and real-time occupied grid maps in dynamic environments.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] Firstly, this application provides a method for constructing an occupied grid map based on a dynamic environment, including:

[0007] Acquire environmental point cloud data and mobile robot pose information;

[0008] Based on the pose information of the mobile robot, the environmental point cloud data is preprocessed to obtain preprocessed environmental point cloud data.

[0009] The preprocessed environmental point cloud data is input into a deep neural network model for semantic classification to obtain semantic labels corresponding to the preprocessed environmental point cloud data; the semantic labels are used to distinguish dynamic objects from static backgrounds.

[0010] Based on Bayesian filtering theory, a particle filtering algorithm is used to track and predict dynamic objects in the preprocessed environmental point cloud data, thereby obtaining the position information and motion state information of the dynamic objects.

[0011] Based on the semantic tags and the position and motion state information of the dynamic objects, the preprocessed environmental point cloud data is divided into regions and updated to obtain a real-time occupied grid map; the real-time occupied grid map is divided into dynamic regions and static regions; wherein the dynamic regions and the static regions are updated at different frequencies.

[0012] Secondly, this application provides a dynamic environment-based occupancy grid map construction system, including:

[0013] The data acquisition module is used to acquire environmental point cloud data and mobile robot pose information;

[0014] The data and processing module is used to preprocess the environmental point cloud data based on the location information of the mobile robot to obtain preprocessed environmental point cloud data.

[0015] The deep learning semantic segmentation module is used to input the preprocessed environmental point cloud data into a deep neural network model for semantic classification, and obtain the semantic labels corresponding to the preprocessed environmental point cloud data; the semantic labels are used to distinguish dynamic objects from static backgrounds.

[0016] The particle filtering module is used to track and predict dynamic objects in the preprocessed environmental point cloud data based on Bayesian filtering theory and particle filtering algorithm, so as to obtain the position information and motion state information of the dynamic objects.

[0017] The occupation grid map update module is used to divide and update the preprocessed environmental point cloud data into regions based on the semantic tags and the position and motion state information of the dynamic objects, so as to obtain a real-time occupation grid map; the real-time occupation grid map is divided into dynamic regions and static regions; wherein the dynamic regions and the static regions are updated at different frequencies.

[0018] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for constructing an occupancy grid map based on a dynamic environment.

[0019] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for constructing an occupancy grid map based on a dynamic environment.

[0020] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for constructing an occupancy grid map based on a dynamic environment.

[0021] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0022] This application provides a method and system for constructing an occupancy grid map based on a dynamic environment. The method includes acquiring environmental point cloud data and mobile robot pose information; preprocessing the environmental point cloud data based on the mobile robot pose information to obtain preprocessed environmental point cloud data; inputting the preprocessed environmental point cloud data into a deep neural network model for semantic classification to obtain semantic labels corresponding to the preprocessed environmental point cloud data; the semantic labels are used to distinguish between dynamic objects and static backgrounds; based on Bayesian filtering theory, a particle filtering algorithm is used to track and predict dynamic objects in the preprocessed environmental point cloud data to obtain the position information and motion state information of the dynamic objects; according to the semantic labels and the position and motion state information of the dynamic objects, the preprocessed environmental point cloud data is divided into regions and updated to obtain a real-time occupancy grid map; the real-time occupancy grid map is divided into dynamic regions and static regions; wherein the dynamic regions and static regions are updated at different frequencies. This application utilizes a deep neural network model for semantic classification, which can effectively classify dynamic objects and static backgrounds in point cloud data, improving recognition and processing capabilities and providing accurate basic information for subsequent dynamic object tracking and map updates. Furthermore, the particle filter algorithm enables real-time tracking of dynamic objects and prediction of their future positions, providing accurate information on the motion state of dynamic objects and thus improving the real-time performance and accuracy of the map. In addition, this application uses different update frequencies for dynamic and static regions, resulting in a lower update frequency for static regions and more frequent updates for dynamic regions, thereby optimizing the use of computational resources. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1This is an application environment diagram of an occupation grid map construction method based on a dynamic environment according to an embodiment of this application;

[0025] Figure 2 A flowchart illustrating a method for constructing an occupied grid map based on a dynamic environment, provided in an embodiment of this application;

[0026] Figure 3 A flowchart illustrating a method for constructing an occupied grid map based on a dynamic environment, provided in another embodiment of this application;

[0027] Figure 4 A flowchart illustrating a method for constructing an occupied grid map based on a dynamic environment, provided in another embodiment of this application;

[0028] Figure 5 A schematic diagram of the functional modules of an occupation grid map construction system based on a dynamic environment provided in an embodiment of this application;

[0029] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] While existing technologies have made significant progress in constructing occupied grid maps in static environments, they still have many shortcomings in dynamic environments. The main deficiencies of existing technologies will be detailed below, and relevant literature and patents will be cited to clarify the differences and deficiencies between these technologies and the solution proposed in this application.

[0032] I. Insufficient ability to recognize and process dynamic objects

[0033] Traditional methods for constructing occupancy grid maps primarily rely on LiDAR data and update the map using Bayesian filtering. However, these methods struggle to effectively distinguish between dynamic objects and static backgrounds in dynamic environments, resulting in inaccurate tracking of the position and motion of dynamic objects, thus affecting the overall accuracy and reliability of the map.

[0034] For example, traditional SLAM (Simultaneous Localization and Mapping) methods, such as GMAPping, perform poorly in dynamic environments. GMAPping assumes the environment is static, so when moving objects are present, they are mistakenly identified as part of the map, leading to the accumulation of localization and mapping errors. Another example is Karto SLAM, which builds maps using LiDAR data, but lacks an effective mechanism for distinguishing dynamic objects, making it unable to handle complex dynamic environments and limiting its applicability in practical applications.

[0035] II. Low real-time performance and computational efficiency

[0036] In dynamic environments, the frequent need for map updates places higher demands on the real-time performance of the system. Existing methods often employ a uniform update frequency, failing to adjust update strategies according to the dynamic nature of the environment. This results in high-frequency updates even in static areas, wasting computational resources and limiting the system's application in complex environments.

[0037] For example, while the traditional FastSLAM algorithm improves computational efficiency to some extent, it still requires frequent updates to the entire map in dynamic environments and cannot effectively distinguish between dynamic and static areas, resulting in a heavy overall computational burden. Furthermore, ORB-SLAM, as a visual SLAM method, performs excellently in static environments, but in dynamic environments, the computational efficiency drops significantly due to the need to process a large number of dynamic feature points, making it difficult to meet real-time requirements.

[0038] III. Insufficient utilization of semantic information

[0039] Traditional occupancy-based raster map construction methods focus primarily on processing geometric information, neglecting the utilization of semantic information about objects in the environment. This results in the system being unable to fully leverage the semantic features of objects to optimize the map construction and update process when dealing with complex scenes, thus limiting the map's expressive power and application scope.

[0040] For example, commonly used occupancy grid map building tools in ROS (such as the gmapping package) mainly rely on geometric information and cannot distinguish between different categories of objects (such as pedestrians, vehicles, buildings, etc.), making it difficult to effectively handle different types of dynamic objects in dynamic environments.

[0041] In recent years, some studies have attempted to incorporate semantic information into map building, such as Semantic SLAM, which integrates semantic information into the SLAM system by combining it with semantic segmentation networks. However, these methods often struggle to achieve a balance between real-time performance and accuracy, especially on resource-constrained mobile robot platforms, where efficient semantic segmentation and map building are difficult to achieve.

[0042] IV. High consumption of storage and computing resources

[0043] In large-scale environments, traditional occupying raster maps require storing a large amount of raster information, especially in high-resolution map construction, where storage and computational overhead increases significantly. Existing methods lack efficient storage and computational optimization techniques, making it difficult to meet the requirements of real-time performance and efficiency.

[0044] For example, OctoMap optimizes the storage of 3D occupied raster maps through an octree structure, but it still faces high computational and storage overhead when dealing with highly dynamic environments, especially when dynamic regions need to be updated frequently, where the performance bottleneck is obvious.

[0045] Furthermore, existing dense matrix representation methods are prone to memory overflow in large-scale environments, making it difficult to achieve efficient map updates and queries. For example, traditional two-dimensional raster maps require a large amount of memory and computing resources when extended to large-scale environments, limiting their scalability in practical applications.

[0046] V. The dynamic and static area update strategies are unreasonable.

[0047] Existing technologies typically employ a uniform map update strategy, failing to differentiate between the characteristics of dynamic and static areas. This approach fails to achieve efficient tracking and prediction in dynamic areas, while also being unable to effectively reduce update frequency in static areas, resulting in low overall system resource utilization.

[0048] For example, traditional Bayesian filtering methods apply the same update frequency to all regions when updating the occupied raster map, failing to differentiate based on the dynamics of objects. In environments with many dynamic objects, this leads to frequent updates of the entire map, increasing the computational burden; while in static regions, frequent updates are redundant and waste valuable computational resources.

[0049] In addition, existing patents have attempted to optimize map updates through different strategies. One such method is a map update method based on regional importance, which performs high-frequency updates by predefining important regions. However, it lacks a dynamic adjustment mechanism and cannot dynamically adjust the update frequency according to real-time environmental changes, resulting in low resource utilization.

[0050] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] The dynamic environment-based occupancy grid map construction method provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send acquired environmental point cloud data and mobile robot pose information to server 104. After receiving the data, server 104 preprocesses the environmental point cloud data based on the mobile robot pose information to obtain preprocessed environmental point cloud data. This preprocessed data is then input into a deep neural network model for semantic classification to obtain semantic labels corresponding to the preprocessed environmental point cloud data. A particle filter algorithm is then used to track and predict dynamic objects in the preprocessed environmental point cloud data to obtain the position and motion state information of the dynamic objects. Based on the semantic labels and the position and motion state information of the dynamic objects, the preprocessed environmental point cloud data is divided into regions and updated to obtain a real-time occupancy grid map. Server 104 can then feed back the obtained real-time occupancy grid map to terminal 102. In addition, in some embodiments, the method for constructing an occupancy grid map based on a dynamic environment can also be implemented by either the server 104 or the terminal 102. For example, the terminal 102 can directly process the environmental point cloud data and the mobile robot pose information, or the server 104 can obtain the environmental point cloud data and the mobile robot pose information from the data storage system and then process them.

[0052] Among them, terminal 102 can be, but is not limited to, various desktop computers, laptops and IoT devices, and server 104 can be implemented by independent server or server cluster composed of multiple servers, or it can be cloud server.

[0053] In one exemplary embodiment, such as Figure 2 As shown, a method for constructing an occupancy grid map based on a dynamic environment is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 205.

[0054] in:

[0055] Step 201: Obtain environmental point cloud data and mobile robot pose information.

[0056] Step 202: Based on the pose information of the mobile robot, preprocess the environmental point cloud data to obtain preprocessed environmental point cloud data.

[0057] Step 203: Input the preprocessed environmental point cloud data into a deep neural network model for semantic classification to obtain semantic labels corresponding to the preprocessed environmental point cloud data; wherein, the semantic labels are used to distinguish between dynamic objects and static backgrounds.

[0058] Step 204: Based on Bayesian filtering theory, the particle filtering algorithm is used to track and predict dynamic objects in the preprocessed environmental point cloud data to obtain the position information and motion state information of the dynamic objects.

[0059] Step 205: Based on semantic tags and the location and motion state information of dynamic objects, the preprocessed environmental point cloud data is divided into regions and updated to obtain a real-time occupied grid map; the real-time occupied grid map is divided into dynamic regions and static regions; the dynamic regions and static regions are updated at different frequencies.

[0060] Furthermore, the method for constructing occupied grid maps based on dynamic environments also includes:

[0061] Step 206: Use compressed sparse rows or octree structure to perform sparse representation on the real-time occupied grid map to obtain the sparsed occupied grid map, and store the sparsed occupied grid map.

[0062] Further, in step 202, based on the mobile robot's pose information, the environmental point cloud data is preprocessed to obtain preprocessed environmental point cloud data, specifically including:

[0063] Step 2021: Use a voxel grid filter to downsample the environmental point cloud data to obtain the first environmental point cloud data.

[0064] Step 2022: Use the statistical outlier removal algorithm to remove noise and outliers from the first environmental point cloud data to obtain the second environmental point cloud data.

[0065] Step 2023: Based on the mobile robot pose information, the second environmental point cloud data is transformed from the sensor coordinate system to the world coordinate system to obtain preprocessed environmental point cloud data.

[0066] Furthermore, the deep neural network model in step 203 includes a PointNet++ network and a GPU acceleration unit; the preprocessed environmental point cloud data is input into the deep neural network model for semantic classification to obtain the semantic labels corresponding to the preprocessed environmental point cloud data, specifically including:

[0067] The preprocessed environmental point cloud data is input into the PointNet++ network and accelerated using a GPU acceleration unit to obtain the semantic labels corresponding to the preprocessed environmental point cloud data.

[0068] Furthermore, the particle filtering algorithm in step 204 includes a state transition model and an observation model; based on Bayesian filtering theory, the particle filtering algorithm is used to track and predict dynamic objects in the preprocessed environmental point cloud data, obtaining the position information and motion state information of the dynamic objects, specifically including:

[0069] Step 2041: Based on the preprocessed environmental point cloud data, obtain the initial particle set and the initial weight set.

[0070] Step 2042: Based on the initial particle set and the state transition model, predict the new particle set. The expression for the state transition model is:

[0071]

[0072] in, Let be the state of the i-th particle at time t; Let u be the state of the i-th particle at time t-1; t For control input; Let f(·) be the process noise of the i-th particle from time t-1 to time t; and let |f(·)| be the state transition function.

[0073] Step 2043: Based on the new particle set and observation model, update the initial weight set to obtain the updated weight set. The expression for the observation model is:

[0074]

[0075] in, The noise of the i-th particle during the process from time t-2 to time t-1; z represents the observation probability. t For observational data.

[0076] Step 2044: Based on the updated weight set, resample the new particle set to obtain the resampled particle set.

[0077] Step 2045: Based on the resampled particle set, obtain the position and motion state information of the dynamic object.

[0078] Furthermore, in step 205, based on semantic tags and the location and motion state information of dynamic objects, the preprocessed environmental point cloud data is divided and updated into regions to obtain a real-time occupied grid map, specifically including:

[0079] Step 2051: Based on semantic tags and the position and motion state information of dynamic objects, the preprocessed environmental point cloud data is divided into dynamic and static regions using a spatial index structure; the spatial index structure is a quadtree or an octree.

[0080] Step 2052: The Bayesian update algorithm is used to update the dynamic and static regions to obtain a real-time occupied grid map. The update frequency of the dynamic region is higher than that of the static region; the update frequency range for the dynamic region can be selected from 30-50Hz based on accuracy requirements and environmental characteristics; the update frequency range for the static region is selected from 1-5Hz.

[0081] By implementing steps 201 to 206 above, this application, through combining deep learning semantic segmentation and particle filtering algorithms, can accurately identify and distinguish dynamic objects from static backgrounds. The deep neural network model utilizes the PointNet++ network to perform semantic classification of environmental point cloud data, ensuring that dynamic objects (such as pedestrians and vehicles) are accurately identified and separated. Furthermore, based on Bayesian filtering theory, the particle filtering algorithm is used to accurately track and predict these dynamic objects, significantly improving the accuracy and real-time performance of the map and reducing the interference of dynamic objects on localization and mapping. Simultaneously, the map update frequency is intelligently adjusted according to the dynamics of the environment, achieving resource optimization and energy saving; high-frequency updates are used in highly dynamic areas to ensure real-time tracking of dynamic objects and timely map reflection; while low-frequency updates are used in static areas to reduce unnecessary computational resource consumption. This differentiated update strategy not only improves the system's computational efficiency but also extends the energy lifespan of the mobile platform, adapting to complex and ever-changing application environments. Furthermore, this application significantly improves the storage and computation efficiency of the grid map by introducing sparse matrix operation optimization; it uses compressed sparse rows (CSR) or octree structure to represent map data, storing only information of non-idle or dynamic areas, which greatly reduces storage space and computational complexity, thereby enabling high frame rate real-time updates in highly dynamic environments and meeting the high real-time requirements of mobile platforms.

[0082] In another exemplary embodiment of this application, such as Figure 3 As shown, taking an autonomous vehicle as a mobile robot as an example, the occupation grid map construction method of this application is explained.

[0083] In urban road environments, autonomous vehicles need to perceive dynamic objects (such as pedestrians and other vehicles) in real time and construct accurate environmental maps to assist navigation and decision-making. The specific workflow is as follows: A 360-degree LiDAR and Inertial Measurement Unit (IMU) mounted on the vehicle's roof acquire point cloud data and pose information of the surrounding environment in real time. A data preprocessing module filters and denoises the acquired environmental point cloud data and transforms it to a global coordinate system. The preprocessed environmental point cloud data is input into a deep neural network model for semantic segmentation, accurately distinguishing pedestrians, vehicles, and static backgrounds (such as roads and buildings). Dynamic object data from the semantic segmentation results is passed to a particle filter module, which tracks the motion trajectories of these dynamic objects and predicts their future positions based on a Bayesian filtering algorithm. An occupancy grid map update module updates the occupancy grid information of dynamic areas at a high frequency based on the real-time positions of dynamic objects, while updating static areas at a lower frequency to ensure the map's real-time performance and stability. A sparse matrix operation optimization module converts the updated occupancy grid map into Compressed Sparse Row (CSR) format to optimize storage and query efficiency. The system control and scheduling module coordinates the operation of each module to ensure that autonomous vehicles can efficiently and in real time perceive dynamic objects and navigate safely in complex urban environments.

[0084] In another exemplary embodiment of this application, such as Figure 4 As shown, taking a drone as a mobile robot as an example, the occupation grid map construction method of this application is explained.

[0085] In complex aerial environments, UAVs need to perceive moving obstacles (such as other aircraft and birds) in real time and construct dynamic environmental maps to plan safe flight paths. The specific workflow is as follows: A combination of 3D LiDAR and GPS / IMU on the UAV collects point cloud data and pose information of the aerial environment in real time. A data preprocessing module filters and denoises the collected environmental point cloud data and transforms it to a global coordinate system. The preprocessed environmental point cloud data is input into a deep neural network model (PointNet++ network) for semantic segmentation, accurately distinguishing moving obstacles (such as other aircraft and birds) from static backgrounds (such as buildings and terrain). Dynamic obstacle data from the semantic segmentation results is passed to a particle filter module. The particle filter, based on a Bayesian filtering algorithm, tracks the trajectory of these dynamic obstacles and predicts their future positions. An occupancy grid map update module updates the occupancy grid information of dynamic areas at a high frequency based on the real-time positions of dynamic obstacles, while updating static areas at a lower frequency to ensure the map's real-time performance and stability. A sparse matrix operation optimization module uses an octree structure to represent the 3D occupancy grid map as a sparse matrix, optimizing storage and computational efficiency. The system control and scheduling module coordinates the operation of each module to ensure that the UAV can efficiently and in real time perceive moving obstacles and perform safe navigation and flight path planning in a dynamic aerial environment.

[0086] Based on the same inventive concept, this application also provides an occupation grid map construction system for implementing the occupation grid map construction method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more video tag processing device embodiments provided below can be found in the limitations of the video tag processing method above, and will not be repeated here.

[0087] In one exemplary embodiment, such as Figure 5 As shown, a dynamic environment-based occupancy grid map construction system is provided. Its modular design provides high scalability and adaptability, with each module independently optimized. This facilitates functional expansion and technical upgrades based on different application scenarios (such as autonomous vehicles and drone navigation), ensuring the system can flexibly respond to diverse environmental requirements. Specifically, it includes:

[0088] 301, Data Acquisition Module, is used to acquire environmental point cloud data and mobile robot pose information. Specifically, the data acquisition module mainly consists of a LiDAR sensor and pose acquisition devices (such as an inertial measurement unit (IMU) and an odometry system) mounted on the mobile robot. The LiDAR scans the environment at a fixed frequency (e.g., 10Hz), generating high-density two-dimensional or three-dimensional point cloud data in real time. Simultaneously, the IMU and odometry provide the mobile robot's current position and attitude information. The function of this module is to ensure that the system can continuously and accurately acquire environmental data and robot pose, providing reliable input for subsequent processing.

[0089] 302, the data preprocessing module, is used to preprocess environmental point cloud data based on the mobile robot's location information to obtain preprocessed environmental point cloud data. Specifically, the data preprocessing module includes a filter, a noise reduction algorithm unit, and a coordinate transformation unit. First, a voxel mesh filter downsamples the original environmental point cloud data to reduce the point cloud density and decrease the amount of subsequent computation. Next, a statistical outlier removal algorithm is used to remove noise and outliers from the environmental point cloud data, improving data quality. Finally, the coordinate transformation unit transforms the environmental point cloud data from the sensor coordinate system to the world coordinate system, and combined with the mobile robot's pose information, achieves accurate positioning. Its working principle is based on the geometric characteristics of the environmental point cloud data, improving data consistency and accuracy through filtering and noise reduction, ensuring that subsequent modules can efficiently process high-quality point cloud data.

[0090] 303, the deep learning semantic segmentation module, is used to input preprocessed environmental point cloud data into a deep neural network model for semantic classification, obtaining semantic labels corresponding to the preprocessed environmental point cloud data. Specifically, a deep neural network (i.e., the PointNet++ network) is used to perform semantic classification on the preprocessed environmental point cloud data. Simultaneously, this module utilizes GPU acceleration units to improve computational efficiency and ensure real-time requirements. In the specific implementation process, the preprocessed environmental point cloud data is input into the deep neural network, and the network outputs a semantic label for each point, distinguishing between dynamic objects (such as pedestrians and vehicles) and static backgrounds (such as buildings and roads). The core of semantic segmentation lies in using a convolutional neural network (CNN) to extract the spatial features of the point cloud and classifying them through fully connected layers. The formula is expressed as:

[0091] y = f(x; Θ);

[0092] Where x is the input environmental point cloud data, i.e., the preprocessed environmental point cloud data, Θ is the network parameter, and y is the output semantic label.

[0093] 304, Particle Filtering Module, is used to track and predict dynamic objects in preprocessed environmental point cloud data based on Bayesian filtering theory and particle filtering algorithms, obtaining the position and motion state information of the dynamic objects. This module includes a particle set, a state transition model, and an observation model. First, the particle state is predicted according to the state transition model (such as a uniform linear motion model):

[0094]

[0095] in, Let be the state of the i-th particle at time t; Let u be the state of the i-th particle at time t-1; t For control input; Let f(·) be the process noise of the i-th particle from time t-1 to time t; and let |f(·)| be the state transition function.

[0096] Then, the particle weights are updated based on the observation model:

[0097]

[0098] in, The noise of the i-th particle during the process from time t-2 to time t-1; z represents the observation probability. t For observational data.

[0099] Finally, a new particle set is generated through a resampling step to avoid particle degradation. The particle filtering module's function is to accurately estimate the position and motion state of dynamic objects, providing a reliable basis for updating the dynamic regions occupying the grid map.

[0100] 305. Occupancy Raster Map Update Module: This module is used to divide and update preprocessed environmental point cloud data into regions based on semantic tags and the location and motion information of dynamic objects, resulting in a real-time occupancy raster map. The specific implementation steps include: First, dividing the map into multiple sub-regions (such as dynamic and static regions) using a spatial index structure (e.g., a quadtree or octree). Then, evaluating the dynamism of each sub-region based on the number and motion status of dynamic objects output by the particle filter. Finally, adjusting the update frequency of each sub-region based on the dynamism evaluation results.

[0101] The specific update steps are as follows: For the dynamic region, apply the Bayesian update formula:

[0102]

[0103]

[0104] Wherein, the denominator p(z)t |x t ) is a normalization constant used to ensure that the final result is a valid probability (the sum of the probabilities of all states is 1), which can be further expanded to be the probability of m t Summing (or integrating) the entire space (occupied / free), i.e. m t This represents the occupied state of the grid at time t (discretely, it can take either "occupied" or "idle" states; continuously, it can be expanded into a corresponding state space); z t x represents the observations obtained by the sensor at time t (such as lidar measurements); t p(m) represents the robot's pose at time t (used to correlate observation and map coordinates); t p(z) represents the prior probability (an estimate of whether a grid cell is occupied before the update at time t, which can be obtained from the experiment at the previous time step); t |m t ,x t ) represents the observation model, i.e., if the grid state is m t And the robot's pose is x t , and obtain the observed z t The probability of.

[0105] Meanwhile, for static areas, the same Bayesian updates are performed at a lower frequency to ensure map stability. This module's function is to rationally allocate map update frequencies based on the dynamics of different areas, improving the system's real-time performance and computational efficiency. The map update frequency for each area is dynamically adjusted based on the results of semantic segmentation and particle filtering.

[0106] 306, the sparse matrix operation optimization module, is used to perform sparse representation of the real-time occupied raster map using compressed sparse rows or octree structures, obtaining a sparsed occupied raster map, and storing the sparsed occupied raster map, optimizing the storage and computation process. In specific implementation, only raster information of non-idle or dynamic areas is stored, significantly reducing storage space. The sparse matrix operation unit accelerates map updates and queries through optimized algorithms (such as sparse matrix multiplication and lookup), reducing computational complexity. The formula is expressed as:

[0107] A CSR ={values,columndices,rowointers};

[0108] The module stores non-zero elements in the `values` array, records the corresponding column indices in `column_indices`, and indicates the starting position of each row in `row_pointers`. Its function is to improve the efficiency of map building and querying through sparse matrix optimization, meeting real-time requirements.

[0109] Furthermore, the occupancy grid map construction system in this embodiment also includes a 307 system control and scheduling module, which is responsible for coordinating the collaborative work of various modules, managing data flow and resource allocation, and ensuring the overall real-time performance and stability of the system. In its implementation, this module employs a priority scheduling algorithm to ensure that critical modules (such as particle filtering and semantic segmentation) are processed first, rationally allocating computing resources and preventing resource contention and bottlenecks.

[0110] The scheduling process includes:

[0111] 1. Task scheduling: Assign processing tasks based on module priority and current system status.

[0112] 2. Resource Management: Dynamically monitor CPU, GPU, and memory usage, adjust resource allocation, and ensure efficient operation.

[0113] 3. Anomaly Handling: Monitor the system's operating status in real time, promptly handle module failures or resource shortages, and ensure system stability.

[0114] The function of this module is to coordinate the operation of all modules, optimize system performance, and ensure the real-time performance and reliability of the map building process.

[0115] This application provides a method and system for constructing an occupancy grid map based on a dynamic environment. It not only accurately distinguishes and tracks dynamic objects but also optimizes storage and computation efficiency through sparse matrix operations and dynamic update strategies. Whether in autonomous vehicles or drones, this system demonstrates significant advantages in efficiency, real-time performance, and resource optimization, significantly improving the environmental perception and navigation capabilities of mobile platforms in complex dynamic environments. It has broad application prospects and significant technological innovation.

[0116] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores and processes data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a dynamic environment-based occupancy grid map construction method.

[0117] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0118] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0119] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0120] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0121] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0122] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0123] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0124] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for constructing an occupied grid map based on a dynamic environment, characterized in that, The method for constructing an occupied grid map based on a dynamic environment includes: Acquire environmental point cloud data and mobile robot pose information; Based on the pose information of the mobile robot, the environmental point cloud data is preprocessed to obtain preprocessed environmental point cloud data. The preprocessed environmental point cloud data is input into a deep neural network model for semantic classification to obtain semantic labels corresponding to the preprocessed environmental point cloud data; the semantic labels are used to distinguish dynamic objects from static backgrounds. Based on Bayesian filtering theory, a particle filtering algorithm is used to track and predict dynamic objects in the preprocessed environmental point cloud data, obtaining the position and motion state information of the dynamic objects. The particle filtering algorithm includes a state transition model and an observation model. The expression of the state transition model is: ； in, Let be the state of the i-th particle at time t; Let i be the state of the i-th particle at time t-1; For control input; The noise of the i-th particle during the process from time t-1 to time t; This is the state transition function; The expression for the observation model is: ； in, The noise of the i-th particle during the process from time t-2 to time t-1; For observation probability; For observational data; Based on the semantic tags and the position and motion state information of the dynamic objects, the preprocessed environmental point cloud data is divided into regions and updated to obtain a real-time occupied grid map; the real-time occupied grid map is divided into dynamic regions and static regions; wherein the dynamic regions and the static regions are updated at different frequencies.

2. The method for constructing an occupied grid map based on a dynamic environment according to claim 1, characterized in that, Based on the mobile robot's pose information, the environmental point cloud data is preprocessed to obtain preprocessed environmental point cloud data, specifically including: The environmental point cloud data is downsampled using a voxel grid filter to obtain the first environmental point cloud data. The statistical outlier removal algorithm is used to remove noise and outliers from the first environmental point cloud data to obtain the second environmental point cloud data; Based on the mobile robot's pose information, the second environmental point cloud data is transformed from the sensor coordinate system to the world coordinate system to obtain preprocessed environmental point cloud data.

3. The method for constructing an occupied grid map based on a dynamic environment according to claim 1, characterized in that, The deep neural network model includes a PointNet++ network and a GPU acceleration unit; the preprocessed environmental point cloud data is input into the deep neural network model for semantic classification to obtain semantic labels corresponding to the preprocessed environmental point cloud data, specifically including: The preprocessed environmental point cloud data is input into the PointNet++ network and accelerated using a GPU acceleration unit to obtain semantic labels corresponding to the preprocessed environmental point cloud data.

4. The method for constructing an occupied grid map based on a dynamic environment according to claim 1, characterized in that, Based on Bayesian filtering theory, a particle filtering algorithm is used to track and predict dynamic objects in the preprocessed environmental point cloud data, obtaining the position and motion state information of the dynamic objects, specifically including: Based on the preprocessed environmental point cloud data, an initial particle set and an initial weight set are obtained; Based on the initial particle set and the state transition model, predict the new particle set; Based on the new particle set and the observation model, the initial weight set is updated to obtain the updated weight set. Based on the updated weight set, the new particle set is resampled to obtain the resampled particle set; Based on the resampled particle set, the position and motion state information of the dynamic object are obtained.

5. The method for constructing an occupied grid map based on a dynamic environment according to claim 1, characterized in that, Based on the semantic tags and the position and motion state information of the dynamic objects, the preprocessed environmental point cloud data is divided and updated into regions to obtain a real-time occupied grid map, specifically including: Based on the semantic tags and the position and motion state information of the dynamic objects, the preprocessed environmental point cloud data is divided into dynamic and static regions using a spatial index structure; the spatial index structure is a quadtree or an octree. The dynamic and static regions are updated using a Bayesian update algorithm to obtain a real-time occupied grid map.

6. The method for constructing an occupied grid map based on a dynamic environment according to claim 1, characterized in that, The update frequency of the dynamic region is greater than that of the static region; the update frequency of the dynamic region ranges from 30Hz to 50Hz; and the update frequency of the static region ranges from 1Hz to 5Hz.

7. The method for constructing an occupied grid map based on a dynamic environment according to claim 1, characterized in that, The method for constructing an occupied grid map based on a dynamic environment also includes: The real-time occupied grid map is sparsely represented using a compressed sparse row or octree structure to obtain a sparse occupied grid map, which is then stored.

8. A system for constructing an occupancy grid map based on a dynamic environment, characterized in that, The dynamic environment-based occupancy grid map construction system includes: The data acquisition module is used to acquire environmental point cloud data and mobile robot pose information; The data and processing module is used to preprocess the environmental point cloud data based on the pose information of the mobile robot to obtain preprocessed environmental point cloud data. The deep learning semantic segmentation module is used to input the preprocessed environmental point cloud data into a deep neural network model for semantic classification, and obtain the semantic labels corresponding to the preprocessed environmental point cloud data; the semantic labels are used to distinguish dynamic objects from static backgrounds. The particle filtering module is used to track and predict dynamic objects in the preprocessed environmental point cloud data based on Bayesian filtering theory and employing a particle filtering algorithm to obtain the position and motion state information of the dynamic objects. The particle filtering algorithm includes a state transition model and an observation model. The expression for the state transition model is: ； in, Let be the state of the i-th particle at time t; Let i be the state of the i-th particle at time t-1; For control input; The noise of the i-th particle during the process from time t-1 to time t; This is the state transition function; The expression for the observation model is: ； in, The noise of the i-th particle during the process from time t-2 to time t-1; For observation probability; For observational data; The occupation grid map update module is used to divide and update the preprocessed environmental point cloud data into regions based on the semantic tags and the position and motion state information of the dynamic objects, so as to obtain a real-time occupation grid map; the real-time occupation grid map is divided into dynamic regions and static regions; wherein the dynamic regions and the static regions are updated at different frequencies.

9. The dynamic environment-based occupancy grid map construction system according to claim 8, characterized in that, The dynamic environment-based occupancy grid map construction system also includes: The sparse matrix operation optimization module is used to perform sparse representation of the real-time occupied grid map using compressed sparse rows or octree structures to obtain a sparse occupied grid map, and to store the sparse occupied grid map.

Images