An artificial intelligence-based robot-assisted navigation device and method

Abstract

The application discloses a kind of robot-assisted navigation device and method based on artificial intelligence, it is related to the technical field of auxiliary navigation, comprising the following steps: real-time collection navigation environment data by multimodal perception device group;Through simultaneous localization and mapping technology, preliminary environment three-dimensional map is established;Unknown area is deduced back using preliminary environment three-dimensional map, and inference is carried out based on the physical and geometric constraints of environment, and the environment three-dimensional map is restored and perfected by reverse simulation model;Compensation and optimization are carried out on environment three-dimensional map using deep learning, and environment three-dimensional map is constantly updated and maintained using robot dynamic data;The optimal path of navigation is found by global path planning algorithm, and the optimal path is adjusted in real time to respond to dynamic environment changes, and the change of future navigation environment is predicted using model.The application collects environment data in real time by multimodal perception device, deduces back unknown area using preliminary three-dimensional map, and is optimized and updated, realizes dynamic navigation path planning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of assisted navigation, in particular to a robot-assisted navigation device and method based on artificial intelligence. BACKGROUND

[0002] With the wide application of robots in logistics, manufacturing, service and other fields, the demand for their autonomous navigation capabilities is increasing. Currently, common navigation technologies mainly rely on sensors such as laser radars and cameras to collect environmental information, and combine simultaneous localization and mapping technology to realize environmental modeling and path planning.

[0003] The prior art has the following disadvantages: the existing navigation methods still have some shortcomings in complex and variable environments, including:

[0004] 1. Insufficient adaptability to dynamic changes in the environment, making it difficult to update the environmental map in real time to respond to obstacles and environmental changes;

[0005] 2. Error accumulation in the process of constructing a three-dimensional map leads to positioning deviation, affecting navigation accuracy;

[0006] 3. It is difficult to achieve multi-scale and multi-level environmental understanding and analysis;

[0007] 4. Lack of inference and supplementing ability for unknown or unmapped areas, limiting autonomous navigation performance in unknown environments;

[0008] 5. Traditional path planning algorithms respond slowly to dynamic environments, making it difficult to achieve real-time optimization and predict future scenario changes.

[0009] The above information disclosed in the background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0010] The purpose of the present application is to provide a robot-assisted navigation device and method based on artificial intelligence to solve the problems in the background technology.

[0011] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: a robot-assisted navigation method based on artificial intelligence, specifically comprising:

[0012] Collecting navigation environment data in real time through a multi-modal perception device group;

[0013] Preprocessing the navigation environment data to establish a preliminary three-dimensional environmental map through simultaneous localization and mapping technology;

[0014] The unknown area is deduced by using the preliminary environmental three-dimensional map and inferring based on the physical and geometric constraints of the environment, and the environmental three-dimensional map is restored and perfected by reverse simulation model;

[0015] The environmental three-dimensional map is compensated and optimized by using deep learning, and the environmental three-dimensional map is continuously updated and maintained by fusing dynamic data of the robot;

[0016] An optimal navigation path is found by using a global path planning algorithm, and the optimal path is adjusted in real time to respond to dynamic environmental changes, and the change of the future navigation environment is predicted by using a model;

[0017] As a preferred scheme of the robot-assisted navigation method based on artificial intelligence, the multi-modal perception device group and the navigation environment data specifically include,

[0018] A laser radar for collecting spatial structure data of the navigation environment;

[0019] A camera for collecting dense depth image data;

[0020] An ultrasonic sensor for collecting distance information data;

[0021] An inertial measurement unit for collecting dynamic data of the robot;

[0022] The multi-modal perception device group is installed at different positions of the robot, the navigation environment coverage is increased by complementary view angles, a master clock source of the robot is selected as a time reference of the multi-modal perception device group, and the collection rate is adjusted according to dynamic requirements of the environment.

[0023] As a preferred scheme of the robot-assisted navigation method based on artificial intelligence, the multi-modal perception device group and the navigation environment data specifically include,

[0024] The preprocessing specifically includes,

[0025] A reference point is selected to initialize the position of the robot as an origin of the preliminary environmental three-dimensional map;

[0026] The navigation environment data are filtered, calibrated and registered, high-dimensional features are extracted, and feature matching is performed;

[0027] The preliminary environmental three-dimensional map specifically includes,

[0028] The pose point change of the current robot relative to the previous position is estimated according to the feature matching result;

[0029] The pose point estimation of the current robot is optimized by using extended Kalman filtering;

[0030] The current navigation environment data are fused into the existing map, and the preliminary three-dimensional environmental model is gradually constructed by using point cloud stitching.

[0031] Detecting whether the robot scans the known map area again through a closed loop to reduce cumulative errors, and performing global fusion and scan optimization when a closed loop is detected.

[0032] As a preferred scheme of the robot-assisted navigation method based on artificial intelligence, wherein:

[0033] The backstepping of the unknown area from the preliminary environmental three-dimensional map specifically includes:

[0034] Learning and updating the error state of the multi-modal perception device group in different navigation environments in real time through deep learning;

[0035] Using Bayesian analysis to analyze the uncertainty boundary of the error state, and outputting the confidence interval of the error;

[0036] Propagating the error as a dynamic variable in the navigation environment, and identifying the unknown area deviation of the preliminary environmental three-dimensional map;

[0037] Calculating the error boundary of each unknown area based on error propagation, specifically including:

[0038] Defining the expected value and standard deviation value of the error as the error boundary, each unknown area is affected by the error of the adjacent area, establishing a neighborhood influence model by accumulating the influence weight to propagate the error, and updating the expectation and covariance, and when any error upper bound exceeds the confidence interval of the error, marking the unknown area of the preliminary environmental three-dimensional map;

[0039] The reverse simulation model for restoring and perfecting the environmental three-dimensional map specifically includes:

[0040] Encoding the geometric relationship and scene layout of the preliminary environmental three-dimensional map;

[0041] Inputting the high-dimensional features of the existing map, and training the autoencoder to learn the latent space representation of the preliminary environmental three-dimensional map;

[0042] Using the decoder to reconstruct the preliminary environmental three-dimensional map, and backstepping the missing structure of the unknown area thereof through reverse reconstruction.

[0043] The identification of the unknown area deviation of the preliminary environmental three-dimensional map specifically includes:

[0044] By introducing error influence transmission, the error is "amplified" or "suppressed" between adjacent areas, and the influence degree is dynamically adjusted according to the uncertainty of the local error, when the errors in the neighborhood are significantly different, the influence is amplified, and the potential error mutation or missing area is timely marked, and the specific formula is:

[0045]

[0046] wherein, denotes an increment of the error of the unknown region i, denotes a neighborhood of the unknown region i, denotes an adjacency weight, denote error vectors of the unknown regions i and j, respectively, denotes an adaptive enhancement function.

[0047] As a preferred scheme of the robot-assisted navigation method based on artificial intelligence provided by the application, wherein:

[0048] The fusion robot dynamic data constantly updates and maintains the environment three-dimensional map, and specifically includes:

[0049] A unified point cloud coordinate system based on the origin point of the environment three-dimensional map;

[0050] According to the robot pose point of the last time period as an initial estimation, the nearest point of the target pose point is matched in the point cloud coordinate system, and a rigid transformation with the minimum distance between the two points is obtained;

[0051] The current relative pose point of the robot in the environment three-dimensional map is updated according to the rotation matrix and the translation vector through the deep network;

[0052] The update quality is judged according to the matching error, and the current updated relative pose point is taken as the next fusion reference.

[0053] As a preferred scheme of the robot-assisted navigation method based on artificial intelligence provided by the application, wherein:

[0054] The finding of the optimal navigation path specifically includes:

[0055] The environment three-dimensional map is divided into multi-scale situations by establishing a perception network, and specifically includes:

[0056] Macro scale: global scene of the environment three-dimensional map;

[0057] Medium scale: crossroads, obstacle boundary, road width of the global scene, output of route details and potential obstacle positions based on a deep learning model;

[0058] Micro scale: output of fine obstacle positions and types of the global scene according to the missing structure of the unknown region;

[0059] Multi-scale situation features are extracted, and a "path cost map" in the environment three-dimensional map is predicted by using a deep neural network, and the length of the navigation path and the degree of obstacle risk of each grid of the cost map are defined.

[0060] ​According to the time period utilization time series model of the multi-modal perception device group, a dynamic obstacle that may appear in the navigation process is predicted, and the current cost map is adjusted to dynamically increase the cost of a potential dangerous area.

[0061] By introducing a dynamic cost map, the A* algorithm is improved to select a starting point to initiate navigation search, generate a candidate path of a corresponding scale, and select an optimal navigation path according to an environment three-dimensional map.

[0062] In another aspect, the present application provides a robot-assisted navigation device based on artificial intelligence, which specifically comprises a multi-modal perception device group, a data processing unit, an analysis and decision unit, and a communication interface unit, the data processing unit specifically comprises an embedded processor for filtering, calibrating, and registering navigation environment data, and a memory for storing an environment three-dimensional map and high-level features; the analysis and decision unit specifically comprises a processing chip for real-time learning and compensation of environment error states, and a path planning module for generating and adjusting a navigation path according to an environment three-dimensional map; and the communication interface unit is a wireless module for realizing data transmission with other units.

[0063] In another aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of a robot-assisted navigation method based on artificial intelligence are realized.

[0064] In another aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of a robot-assisted navigation method based on artificial intelligence are realized.

[0065] In the above technical solution, the present application provides the following technical effects and advantages:

[0066] Through the cooperative work of the multi-modal perception device group, comprehensive and multi-perspective navigation environment information is obtained, multi-scale situation division makes the environment understanding more delicate and comprehensive, and the accuracy and robustness of navigation are improved, the cumulative error is significantly reduced through closed-loop detection and global fusion and scan optimization, unknown area deviations are detected in real time through deep learning and Bayesian analysis, potential environmental changes or deviations are identified in a timely manner, the accuracy and stability of the map are effectively improved, unknown areas with large deviation are dynamically identified by means of an error propagation model and neighborhood influence relationship, unknown or unmapped areas are strengthened in inference and supplement capacity by means of reverse simulation and autoencoder technology, the path cost map is adjusted according to the prediction to avoid potential dangerous areas in advance, and "active" obstacle avoidance is realized. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only represent some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained based on these drawings.

[0068] Figure 1 A method flowchart of a robot-assisted navigation method based on artificial intelligence. DETAILED DESCRIPTION

[0069] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive concept to those skilled in the art.

[0070] Embodiment 1, with reference to Figure 1 For the first embodiment of the present application, the embodiment provides a robot-assisted navigation device and method based on artificial intelligence, specifically comprising:

[0071] S1, collecting navigation environment data in real time through a multi-modal perception device group;

[0072] The multi-modal perception device group and the navigation environment data specifically comprise,

[0073] A laser radar for collecting navigation environment spatial structure data;

[0074] A camera for collecting dense depth image data;

[0075] An ultrasonic sensor for collecting distance information data;

[0076] An inertial measurement unit for collecting robot dynamic data;

[0077] The multi-modal perception device group is installed at different positions of the robot, the navigation environment coverage is increased through complementary view angles, the main clock source of the robot itself is selected as the time reference of the multi-modal perception device group, and the collection rate is adjusted according to the dynamic demand of the environment.

[0078] It should be further noted that complementary view angles of different perception devices can reduce shadow occlusion, avoid blind areas, and improve depth and spatial structure perception capabilities;

[0079] It should be further noted that the main clock source of the robot itself is selected as the synchronization reference of the multiple sensors, ensuring the time consistency of the collected data, thereby reducing the registration error caused by the time difference;

[0080] More should be explained that the collection rate is adjusted in real time according to environmental changes (such as speed, complexity, appearance of dynamic obstacles) to balance energy consumption and sensing effect.

[0081] S2, preprocessing the navigation environment data, and establishing a preliminary three-dimensional environment map through simultaneous localization and mapping technology;

[0082] The preprocessing specifically includes,

[0083] The reference point is selected to initialize the robot position as the origin of the preliminary three-dimensional environment map;

[0084] The navigation environment data is filtered, calibrated and registered, high-dimensional features are extracted, and feature matching is performed;

[0085] More should be explained that different filtering algorithms such as Kalman filtering, median filtering, mean filtering or outlier filtering are used for different sensing devices to optimize the sensing effect;

[0086] More should be explained that the different sensing devices are internally calibrated and externally calibrated to eliminate the deviation between the sensing devices;

[0087] More should be explained that feature matching matches continuous feature points by feature descriptors, and realizes accurate motion estimation and environment alignment.

[0088] The preliminary three-dimensional environment map specifically includes,

[0089] According to the feature matching result, the pose point change of the current robot relative to the previous position is estimated;

[0090] The pose point estimation of the current robot is optimized through extended Kalman filtering;

[0091] The current navigation environment data is fused into the existing map, and the point cloud splicing is used to gradually construct the preliminary three-dimensional environment model;

[0092] Through loop detection, whether the robot scans the known map area again is detected to reduce the cumulative error, and when the loop is detected, global fusion and scanning optimization are performed.

[0093] More should be explained that the extended Kalman filtering for optimizing the pose point estimation of the current robot includes the initialization, prediction and correction links of filtering, and ensures accurate estimation of the current position in a dynamic environment;

[0094] More should be explained that the closed loop detection presets unique landmark features (such as special structure, color label or persistent geometric features) in the navigation environment, builds a hash index based on the key features in the map area (such as local feature descriptor hash), quickly retrieves the potential matching area, establishes a closed loop constraint once the closed loop is detected, connects the current node with the identified corresponding map node, forms a closed loop edge; the closed loop constraint is taken as a global optimization target, panoramic optimization is carried out, error convergence is realized through multiple iterations, and the accuracy and reliability of the environment map are improved.

[0095] S3, the unknown area is deduced back from the preliminary environment three-dimensional map based on the physical and geometric constraints of the environment, and the environment three-dimensional map is restored and perfected through reverse simulation model;

[0096] The preliminary environment three-dimensional map is deduced back, specifically comprising,

[0097] The error state of the multi-modal perception device group in different navigation environments is learned and updated in real time through deep learning;

[0098] The uncertainty boundary of the error state is analyzed by using Bayesian analysis, and the confidence interval of the error is output;

[0099] The error is propagated in the navigation environment as a dynamic variable to identify the unknown area deviation of the preliminary environment three-dimensional map;

[0100] Based on the error propagation, the error boundary of each unknown area is calculated, and when any error upper limit exceeds the confidence interval of the error, the unknown area of the preliminary environment three-dimensional map is marked;

[0101] More should be explained that the propagation mechanism of the error in the environment specifically includes that the error is taken as a dynamic random variable, and the error is propagated according to the environmental features (such as the error state of the adjacent area, the path and the scene layout), the transmission strength of the error is dynamically adjusted through the neighborhood influence model, the error is diffused in the neighborhood, the process of error accumulation and infection is simulated, the error estimate of each area is updated in real time according to the global information, and a spatially distributed error field is formed;

[0102] More should be explained that the evolution of the error boundary is continuously monitored, the outputs of deep learning and Bayesian inference are combined, the error model parameters are dynamically adjusted, and the accuracy of error prediction is improved.

[0103] The preliminary environment three-dimensional map is restored and perfected through the reverse simulation model, specifically comprising,

[0104] The geometric relationship and scene layout of the preliminary environment three-dimensional map are encoded;

[0105] The high-dimensional features of the existing map are input, and the autoencoder is trained to learn the latent space representation of the preliminary environment three-dimensional map;

[0106] The preliminary environment three-dimensional map is reconstructed by a decoder, and missing structures of unknown regions thereof are inversely deduced by inverse reconstruction.

[0107] The identifying unknown region deviation of the preliminary environment three-dimensional map specifically comprises,

[0108] The error is "amplified" or "suppressed" between adjacent regions by introducing error influence transmission, the influence degree is dynamically adjusted according to the uncertainty of local error, the influence is amplified when the errors in the neighborhood are significantly different, and potential error mutations or missing regions are marked in time, and the specific formula is:

[0109]

[0110] Wherein, represents the increment of the error of the unknown region i, represents the neighborhood of the unknown region i, represents the adjacent weight, , respectively represent the error vectors of the unknown region i and the unknown region j, () represents an adaptive enhancement function.

[0111] S4, compensating and optimizing the environment three-dimensional map by using deep learning, and constantly updating and maintaining the environment three-dimensional map by fusing robot dynamic data;

[0112] The constantly updating and maintaining the environment three-dimensional map by fusing robot dynamic data specifically comprises,

[0113] Based on the origin unified point cloud coordinate system of the environment three-dimensional map;

[0114] According to the pose point of the robot in the last time period as an initial estimate, the nearest point of the target pose point is matched in the point cloud coordinate system, and the rigid transformation with the minimum distance between the two points is obtained;

[0115] The current relative pose point of the robot in the environment three-dimensional map is updated according to the rotation matrix and the translation vector by the deep network;

[0116] The update quality is judged according to the matching error, and the current updated relative pose point is taken as the reference for the next fusion.

[0117] S5, finding the optimal navigation path by a global path planning algorithm, and adjusting the optimal path in real time to respond to dynamic environmental changes, and predicting the future navigation environment changes by using a model;

[0118] The finding the optimal navigation path specifically comprises,

[0119] The environment three-dimensional map is divided into multiple scales by establishing a perception network, specifically comprising,

[0120] Macro scale: global scene of the environmental three-dimensional map;

[0121] Mid-scale: divide the global scene into intersections, obstacle boundaries, road widths, output route details and potential obstacle positions based on deep learning model;

[0122] Micro scale: output fine obstacle positions and types of the global scene according to the missing structure of the unknown area;

[0123] Extract multi-scale situation features, and use a deep neural network to predict a "path cost map" in the environmental three-dimensional map, define the length of the navigation path and the degree of obstacle risk corresponding to each grid of the cost map;

[0124] According to the time period collected by the multi-modal perception device group, use a time series model to predict dynamic obstacles that may appear in the navigation process, adjust the current cost map, and dynamically increase the cost of potential dangerous areas;

[0125] By introducing a dynamic cost map, the A* algorithm is improved to select the starting point to initiate navigation search, generate candidate paths of corresponding scales, and select the optimal navigation path according to the environmental three-dimensional map.

[0126] The computer device can be a terminal, which includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used for wired or wireless communication with external terminals. Wireless communication can be achieved through WIFI, operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. It can also be an external keyboard, touchpad or mouse, etc.

[0127] The embodiment also provides a storage medium on which a computer program is stored, the program being executed by a processor to implement steps of the method for robot-assisted navigation based on artificial intelligence proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk, or an optical disk.

[0128] Embodiment 2

[0129] The following is another embodiment of the application, which provides a method for robot-assisted navigation based on artificial intelligence. In order to verify the beneficial effects of the application, a simulation experiment is conducted for scientific demonstration.

[0130] In order to verify the effect of the application, a vehicle platform is selected as the background, and a multi-modal perception device group is constructed by laser radar, camera, ultrasonic sensor, and inertial measurement unit to scan the environment.

[0131] All perception devices collect data once per second, a total of 16.7 hours of data samples, covering an area of 5000 square meters. Within a range of 50 meters close to the vehicle, a higher resolution (10 cm / pixel) is used for detailed analysis.

[0132] In the far area (such as 300 meters away), a lower resolution is used to capture the overall environmental features. The navigation accuracy after multi-scale situation division is improved by 14%, from the original error of 1.6 meters to 1.4 meters.

[0133] Through the closed-loop detection mechanism and comparison with the global map, if a deviation (for example, an error exceeding 5%) is detected, an error correction process is immediately started, the probability distribution of the error is calculated using the Bayesian analysis method, and the optimal correction strategy is determined.

[0134] During the experiment, after 20 rounds of closed-loop detection and correction, the overall error accumulation is reduced from 7.8% to 3.2%, significantly improving the stability of navigation.

[0135] In a complex environment scene, the deep learning model accurately identifies 80% of unknown obstacles, which is significantly higher than the 60% of traditional methods, significantly improving the accuracy of detection.

[0136] By analyzing the error influence between different areas through the error propagation model, and combining the neighborhood influence relationship, the area related to the current deviation can be identified, and the path cost map can be updated in real time. For example, when a large error is detected in a certain road segment, the cost of this area is automatically adjusted according to the error propagation model, and the driving route is adjusted, so that the path optimization success rate is increased by 25%, and unnecessary detours are reduced by 15%.

[0137] By reverse simulation to speculate the unknown area, when a sudden obstacle appears, the possible obstacle trajectory is speculated through reverse simulation, and the inference ability to the unmapped area is enhanced by combining the autoencoder technology, and the active obstacle avoidance successfully avoids 90% of dangerous events, which is 25% higher than the traditional method of 65%, and improves the safety by 25%.

[0138] Through the cooperative work of multi-modal perception equipment group, comprehensive and multi-view navigation environment information is obtained, multi-scale situation division makes the environment understanding more delicate and comprehensive, and the accuracy and robustness of navigation are improved, the cumulative error is significantly reduced by using closed-loop detection and global fusion, scanning optimization, the unknown area deviation is detected in real time by deep learning and Bayesian analysis, potential environmental changes or deviations are identified in time, the accuracy and stability of the map are effectively improved, unknown areas with large deviation are dynamically identified by using the error propagation model and neighborhood influence relationship, and the inference and supplement ability of unknown or unmapped areas is enhanced by combining reverse simulation and autoencoder technology, the path cost map is adjusted according to the prediction, and potential dangerous areas are avoided in advance, realizing "active" obstacle avoidance.

[0139] The above only describes some exemplary embodiments of the present application in a descriptive manner, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

Claims

1. An artificial intelligence-based robot-assisted navigation method, characterized by, Specifically comprising: Collecting navigation environment data in real time through a multi-modal perception device group; Preprocessing the navigation environment data to establish a preliminary three-dimensional environment map through simultaneous localization and mapping technology; Using the preliminary three-dimensional environment map to deduce unknown areas and making inferences based on the physical and geometric constraints of the environment, and restoring and perfecting the three-dimensional environment map through a reverse simulation model; The use of preliminary three-dimensional environment map to deduce unknown areas specifically includes, Learning and updating the error state of the multi-modal perception device group in different navigation environments in real time through deep learning; Using Bayesian analysis to determine the uncertainty boundary of the error state, and outputting the confidence interval of the error; Spreading the error as a dynamic variable in the navigation environment to identify the deviation of the unknown areas of the preliminary three-dimensional environment map; Based on the error propagation, calculating the error boundary of each unknown area, and marking the unknown areas of the preliminary three-dimensional environment map when the upper bound of any error exceeds the confidence interval of the error; The restoration and perfection of the three-dimensional environment map through the reverse simulation model specifically includes, Encoding the geometric relationship and scene layout of the preliminary three-dimensional environment map; Inputting the high-dimensional features of the existing map and training the autoencoder to learn its latent space representation; Using the decoder to reconstruct the preliminary three-dimensional environment map; Reverse reconstruction to deduce the missing structure of unknown areas; Using deep learning to compensate and optimize the three-dimensional environment map, and continuously updating and maintaining the three-dimensional environment map by fusing dynamic data of robots; Finding the optimal navigation path through a global path planning algorithm, and adjusting the optimal path in real time to respond to dynamic environmental changes, and predicting future changes in the navigation environment using the model.

2. The method of claim 1, wherein: The multi-modal perception device group and the navigation environment data specifically include, A laser radar for collecting spatial structure data of the navigation environment; A camera for collecting dense depth image data; An ultrasonic sensor for collecting distance information data; An inertial measurement unit for collecting dynamic data of robots. 3.The method of claim 1, wherein: The preprocessing specifically includes, Selecting a reference point to initialize the robot position as the origin of the preliminary three-dimensional environment map; Filtering, calibrating and registering the navigation environment data, extracting high-dimensional features, and performing feature matching; The establishment of the preliminary three-dimensional environment map specifically includes, Estimating the pose point change of the current robot relative to the previous position according to the feature matching results; Optimizing the pose point estimation of the current robot through extended Kalman filtering; Fusing the current navigation environment data into the existing map, and gradually constructing the preliminary three-dimensional environment model by point cloud stitching; Detecting whether the robot scans the known map area again through closed-loop detection.

4. The method of claim 1, wherein: The identification of the deviation of the unknown areas of the preliminary three-dimensional environment map specifically includes, By introducing error influence transmission, the error is "amplified" or "suppressed" between adjacent areas, and the influence degree is dynamically adjusted according to the uncertainty of local error. When the error in the neighborhood is significantly different, the influence is amplified, and potential error mutations or missing areas are marked in time. Its specific formula is: ； where, denotes the increment of the error of unknown region i, denotes the neighborhood of unknown region i, denotes the adjacency weight, , denote the error vectors of unknown regions i and j, respectively, denotes the adaptive enhancement function.

5. The method of claim 1, wherein: The continuous updating and maintenance of the three-dimensional environment map by fusing dynamic data of robots specifically includes, A unified point cloud coordinate system based on the origin of the three-dimensional environment map; According to the pose point of the robot in the last time period as an initial estimate, the closest point of the target pose point is matched in the point cloud coordinate system, and a rigid transformation with the minimum distance between the two points is obtained; According to the rotation matrix and the translation vector, the current relative pose point of the robot in the environment three-dimensional map is updated through the deep network; According to the matching error, the update quality is judged, and the current updated relative pose point is taken as the reference for the next fusion.

6. The method of claim 1, wherein: The optimal navigation path finding specifically includes, The environment three-dimensional map is divided into multi-scale situations by establishing a perception network; Multi-scale situation features are extracted, and a "path cost map" in the environment three-dimensional map is predicted by using a deep neural network, and the length of the navigation path and the obstacle risk degree passed through each grid of the cost map are defined; According to the time period collected by the multi-modal perception device group, a time series model is used to predict the dynamic obstacles that may appear in the navigation process, adjust the current cost map, and dynamically increase the cost of the potential dangerous area; By introducing the dynamic cost map, the A* algorithm is improved to select the starting point to initiate navigation search, generate candidate paths of corresponding scales, and select the optimal navigation path according to the environment three-dimensional map.

7. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to realize the steps of the robot-assisted navigation method based on artificial intelligence according to any one of claims 1-6.

8. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the robot-assisted navigation method based on artificial intelligence according to any one of claims 1-6.

Images