Laser radar environment modeling and navigation system based on artificial intelligence

Abstract

The invention relates to a laser radar environment modeling and navigation system based on artificial intelligence, and the system comprises the steps: obtaining three-dimensional point cloud data in a complex environment, and obtaining an initial data set; preprocessing the initial data set by adopting a data enhancement mode to obtain a target data set; denoising and complementing the initial data set by adopting a deep learning model based on a generative adversarial network to obtain a target data set; dividing the target data set into a training set and a test set; distinguishing static obstacles and dynamic objects in the training set by adopting an attention enhancement algorithm; identifying a state space in a complex environment by adopting a path planning algorithm; based on the combination of the convolutional neural network and the recurrent neural network, the environmental complexity is evaluated and the control parameters are dynamically adjusted, so that the environmental modeling precision of the system is improved, the path planning efficiency and real-time performance are improved, the navigation control adaptive capability is improved, and the engineering and generalization effects of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of environmental modeling and navigation technology, and in particular to an artificial intelligence-based laser radar environmental modeling and navigation system. Background Art

[0002] LiDAR navigation systems and traditional path planning methods have been widely used and have achieved considerable success in environmental perception and navigation. LiDAR, with its high precision and interference immunity, has become a key sensor for environmental modeling and perception. It can generate high-resolution 3D point clouds, providing environmental information to mobile devices. Traditional path planning methods, based on static graph structures or predefined cost maps, provide basic path planning support for mobile device navigation.

[0003] Currently, researchers are attempting to introduce cutting-edge AI algorithms such as deep learning, reinforcement learning, and graph neural networks to advance environmental perception and navigation technologies. For example, patent CN118999529A proposes a 3D lidar-based semantic mapping and positioning system for mobile robots. This system uses deep learning to semantically segment laser point clouds and extract key features, making maps more semantically informative and improving the accuracy and robustness of the navigation system. Another example is patent CN108107459A, which proposes a robot orientation detection method based on a navigation system. This method utilizes multi-module collaborative positioning technology to improve the robot's initial positioning accuracy and operational efficiency.

[0004] However, in complex environments (such as urban roads, forests, and warehouse logistics), LiDAR is subject to dynamic obstacles, unstructured terrain, and inclement weather, and point cloud data often suffers from noise, occlusion, and uneven sparsity. This reduces the accuracy of environmental modeling and leads to errors in obstacle identification and traffic area extraction. Furthermore, traditional path planning methods lack the ability to perceive and adapt to dynamic environments in real time. They suffer from low computational efficiency and delayed response in large-scale, highly dynamic scenarios, making it difficult to meet millisecond-level response requirements. Furthermore, existing technologies suffer from low integration, poor environmental adaptability, and insufficient real-time performance, making them unable to fully meet the needs of complex and diverse application scenarios.

[0005] Therefore, the present invention proposes a lidar environment modeling and navigation system based on artificial intelligence. Summary of the Invention

[0006] In view of the above problems existing in the prior art, the purpose of the embodiments of the present invention is to provide a lidar environment modeling and navigation system based on artificial intelligence.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an artificial intelligence-based laser radar environment modeling and navigation system, comprising: S1, obtain 3D point cloud data in a complex environment to obtain the initial data set; S2, uses a deep learning model based on generative adversarial networks to denoise and complete the initial dataset to obtain the target dataset; S3, dividing the target data set into a training set and a test set; S4, uses an attention enhancement algorithm to distinguish static obstacles and dynamic objects in the training set; S5, uses path planning algorithms to identify state spaces in complex environments; S6, based on convolutional neural networks combined with recurrent neural networks, evaluates environmental complexity and dynamically adjusts control parameters to form an adaptive navigation strategy update mechanism; S7, the test set is sent into the constructed network model to complete the modeling and perform navigation.

[0008] Furthermore, in S2, a deep learning model based on a generative adversarial network is used to denoise and complete the initial dataset, and the target dataset obtained includes: S21, the initial dataset is used as multi-source data input, and multi-frame time series point cloud (including timestamp) + IMU attitude data (acceleration / angular velocity) are input for spatiotemporal alignment (extended Kalman filter), dynamic occlusion holes (such as pedestrians occluding vehicles) are repaired in the multi-frame time series, and IMU attitude assists in spatial positioning to reduce completion errors; An IMU is a sensor module that can measure an object's acceleration, angular velocity, and (sometimes) magnetic field.

[0009] Furthermore, S22, the initial dataset is processed with point cloud features, and geometric features (curvature / density) are extracted through the PointConv layer. The PointConv layer supports unordered point clouds. The PointConv layer is used for deep hierarchical feature learning networks of point sets in metric space; FPS is a technology that reduces the amount of video data or reduces the processing load by lowering the frame rate (Frames Per Second, FPS).

[0010] Furthermore, in S4, an attention enhancement algorithm is used to distinguish static obstacles and dynamic objects in the training set, including: An attention-enhanced PointNet++ based on PointNet++, a point cloud deep learning framework that improves the accuracy and efficiency of point cloud analysis through hierarchical structure and local feature extraction; The variable radius is grouped by multi-scale grouping. The small scale is used to capture small targets such as pedestrians; the large scale is used to extract global features such as road boundaries.

[0011] Furthermore, the dynamic attention module is adjusted by inputting temporal difference data (including data of the current frame and the previous frame), and the dynamic probability weight is calculated based on the temporal difference data. When the weight of dynamic points (such as pedestrians) increases, the weight of static points decreases, and when the weight of dynamic points decreases, the weight of static points increases, and the characteristics of radio targets are enhanced preferentially.

[0012] Furthermore, the generator generates point clouds, predicts dynamic occlusion missing point clouds, filters noise points, and learns the laws of dynamic scenes to repair large areas of missing images.

[0013] Furthermore, in S5, a path planning algorithm is used to identify the state space in a complex environment, including: optimizing the path through a bidirectional search strategy, expanding the search tree simultaneously at the starting point and the end point, terminating when they meet, and prioritizing the expansion of high-threat areas (such as nodes near pedestrians).

[0014] Furthermore, in S6, based on the convolutional neural network combined with the recurrent neural network, the environmental complexity is evaluated and the control parameters are dynamically adjusted to form an adaptive navigation strategy update mechanism, including: inputting a training set, which includes lidar projection images and visual images, outputting an environmental complexity score, and dynamically adjusting the speed threshold.

[0015] The embodiment of the present invention further provides a network-side server, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned artificial intelligence-based lidar environment modeling and navigation system.

[0016] An embodiment of the present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned artificial intelligence-based lidar environment modeling and navigation system.

[0017] The beneficial effects of the present invention are: The present invention provides an artificial intelligence-based laser radar environment modeling and navigation system, which obtains three-dimensional point cloud data in a complex environment to obtain an initial data set; pre-processes the initial data set in a data enhancement manner to obtain a target data set; S2, uses a deep learning model based on a generative adversarial network to denoise and complete the initial data set to obtain a target data set; divides the target data set into a training set and a test set; uses an attention enhancement algorithm to distinguish static obstacles and dynamic objects in the training set; uses a path planning algorithm to identify the state space in a complex environment; based on a convolutional neural network combined with a recurrent neural network, evaluates the complexity of the environment and dynamically adjusts the control parameters to form an adaptive navigation strategy update mechanism; sends the test set to the constructed network model to complete the modeling and perform navigation. The artificial intelligence-based laser radar environment modeling and navigation system constructed by the present invention achieves the goal of improving the system's environmental modeling accuracy, improving path planning efficiency and real-time performance, improving navigation control adaptability, and improving the system's engineering and generalization effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below with reference to the accompanying drawings and examples.

[0019] In the picture: Figure 1 This is a flow chart of an artificial intelligence-based lidar environment modeling and navigation system in the present invention.

[0020] Figure 2 This is a flowchart of the point cloud denoising and completion process in the present invention.

[0021] Figure 3 This is a flowchart of the algorithm for point cloud feature extraction and semantic recognition of the present invention.

[0022] Figure 4 Schematic diagram of point cloud processing based on deep learning of the present invention.

[0023] Figure 5 This is a schematic diagram of the laser radar data of the present invention.

[0024] Figure 6 It is a structural diagram of a network-side server provided according to a third embodiment of the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0026] The embodiment of the present invention provides an artificial intelligence-based laser radar environment modeling and navigation system, which obtains three-dimensional point cloud data in a complex environment to obtain an initial data set; pre-processes the initial data set by data enhancement to obtain a target data set; S2, uses a deep learning model based on a generative adversarial network to denoise and complete the initial data set to obtain a target data set; divides the target data set into a training set and a test set; uses an attention enhancement algorithm to distinguish between static obstacles and dynamic objects in the training set; uses a path planning algorithm to identify the state space in a complex environment; based on a convolutional neural network combined with a recurrent neural network, evaluates the complexity of the environment and dynamically adjusts the control parameters to form an adaptive navigation strategy update mechanism; sends the test set to the constructed network model to complete the modeling and perform navigation. The artificial intelligence-based laser radar environment modeling and navigation system constructed by the present invention achieves the goal of improving the system's environmental modeling accuracy, improving path planning efficiency and real-time performance, improving navigation control adaptability, and improving the system's engineering and generalization effects. The following is a detailed description of the implementation details of an artificial intelligence-based lidar environment modeling and navigation system in this embodiment. The following content is only for the convenience of understanding the implementation details and is not necessary for the implementation of this solution. See Figure 5 and Figure 6 This invention uses Hesai Technology's AT128 high-performance lidar, featuring a 128-beam sensor, a 10Hz scanning frequency, ±2cm ranging accuracy, and a 200-meter effective range. This module efficiently captures 3D point cloud data in complex environments, providing high-quality raw data input for the system.

[0027] Data processing: The NVIDIA Orin NX edge computing platform, equipped with a multi-core CPU, an Ampere(8) architecture GPU, and a dedicated AI acceleration module, achieves a computing power of 275 TOPS, supporting the parallel processing of large-scale point cloud data and efficient inference of deep learning models. This platform can ensure real-time response speed under high load and complex scenarios, ensuring the system's processing power and the effectiveness of real-time decision-making.

[0028] NVIDIA Orin NX is a high-performance edge computing platform launched by NVIDIA, designed to support complex artificial intelligence (AI) applications in areas such as robots, drones, smart cameras and portable medical devices.

[0029] Sensor Fusion: Integrating the IMX586 vision sensor and the BMI088 inertial measurement unit (IMU), the module fuses multi-source data using the extended Kalman filter (EKF) algorithm. This module significantly improves positioning accuracy, perception robustness, and system redundancy, enhancing the system's adaptability to complex and dynamic environments and ensuring stable operation in these conditions.

[0030] See Figures 1 to 3 The first embodiment of the present invention provides an artificial intelligence-based laser radar environment modeling and navigation system, comprising: S1, obtain the three-dimensional point cloud data in the complex environment and obtain the initial data set.

[0031] During the training process, a large amount of real-scene point cloud data from the IQmulus & TerraMobilita Contest dataset and the WHU-TLS / MLS dataset is used as the initial dataset.

[0032] The TerraMobilita Contest was initiated by the French National Geoinformation Agency (IGN) and the Center for Mathematical Morphology (CMM). It focuses on the dense urban street scenes of Paris and contains 300 million points of high-density mobile laser scanning data. It covers complex environments such as dynamic pedestrians / vehicles, narrow roads and building occlusions, and provides object-level unique ID annotation and 20+ fine-grained semantic categories.

[0033] WHU-TLS was collected by Wuhan University and covers a variety of typical scenes such as urban buildings, campus roads, and vegetation areas. It includes two modal data types: terrestrial laser scanning (TLS) and mobile laser scanning (MLS). It has the characteristics of high-density point cloud, multi-echo information and precise geographic positioning, and can effectively support point cloud denoising, completion and semantic modeling tasks in complex environments.

[0034] TLS, a terrestrial 3D laser scanning system and related technologies developed by Wuhan University in China, covers 11 different environments, including subway stations, high-speed rail stations, mountains, forests, parks, campuses, residential areas, riverbanks, cultural heritage buildings, underground mines, and tunnels. It includes 115 survey stations, 1.74 billion 3D points, and a true transformation matrix between point clouds. It is primarily used in high-precision topographic surveying, building modeling, geological disaster monitoring, and cultural heritage protection. The system combines laser ranging technology, the Global Navigation Satellite System (GNSS), and an Inertial Measurement Unit (IMU) to rapidly acquire 3D point cloud data of surfaces or objects, providing foundational data support for Geographic Information Systems (GIS) and Computer-Aided Design (CAD).

[0035] S2, uses a deep learning model based on generative adversarial networks to denoise and complete the initial dataset to obtain the target dataset; The point cloud data collected by the drone in real time through the lidar is transmitted to the data processing unit and processed by the GAN-based denoising and completion model.

[0036] See 2. A GAN is a deep learning model trained through an adversarial game mechanism. Its core concept is derived from the two-person zero-sum game in game theory. It consists of two neural networks: a generator and a discriminator. These two networks compete with each other and evolve together to ultimately generate high-quality data.

[0037] S21 uses the initial dataset as multi-source data input, inputs multi-frame time series point clouds (including timestamps) + IMU attitude data input (acceleration / angular velocity), and performs spatiotemporal alignment (extended Kalman filter) on the multi-frame time series. Dynamic occlusion holes (such as pedestrians blocking vehicles) are repaired in the multi-frame time series, and IMU attitude assists in spatial positioning to reduce completion errors.

[0038] An IMU is a sensor module that measures an object's acceleration, angular velocity, and (sometimes) magnetic field. It is widely used in robotics, autonomous driving, drones, virtual reality (VR), aerospace, and other fields. It fuses data from multiple sensors to estimate an object's motion state and posture in real time.

[0039] S22, perform point cloud feature processing on the initial dataset, extract geometric features (curvature / density) through the PointConv layer, and the PointConv layer supports unordered point clouds.

[0040] The PointConv layer is used to measure the deep hierarchical feature learning network of point sets in space. It is mainly used to process 3D point cloud data and is widely used in autonomous driving, robot perception, 3D reconstruction and other fields.

[0041] S23 downsamples the initial dataset to reduce the computational effort (e.g., from 100,000 points to 10,000). A point cloud is a massive collection of points representing the surface characteristics of a target, typically obtained through laser measurement or photogrammetry. Point cloud data accurately reflects the true surface conditions, such as ground conditions and reflective features.

[0042] FPS is a technology that reduces the amount of video data or reduces the processing load by lowering the frame rate (Frames Per Second, FPS). It is commonly used in scenarios such as video compression, video processing, and game rendering optimization.

[0043] S24, generates point clouds through the generator (G), predicts dynamic occlusion missing point clouds (such as vehicle outline completion), filters noise points (such as leaf misdetection points, the noise rate is less than 5%), and learns the laws of dynamic scenes to repair large-area missing points that traditional filtering cannot handle. S25, evaluates authenticity and geometric consistency through the discriminator (D), and processes unstructured point cloud data through Chamfer Distance. Chamfer Distance is a metric method for measuring the similarity between two point sets. It is widely used in computer vision, geometric processing, and generative model evaluation, especially when processing unstructured point cloud data. Given two point sets A and B, Chamfer Distance calculates the sum of the squares of the nearest distances from each point in point set A to point set B, plus the sum of the squares of the nearest distances from each point in point set B to point set A. Semantic consistency: pre-trained semantic labels (such as the "vehicle" category) ensure the authenticity and semantic accuracy of the generated point cloud through dual loss constraints.

[0044] Determine whether the loss function converges. When it converges, output the target data set. If the function does not converge, improve the model accuracy through continuous iterative optimization, update the G / D parameters through back propagation, and make the function converge through continuous training to obtain the target data set.

[0045] As an example, after 1000 training iterations, the loss converged, the completion error dropped from more than 40% to more than 15%, and the point cloud completeness (+25%) and semantic accuracy (+12%) were improved, and the target dataset was obtained.

[0046] S3: Divide the target data set into a training set and a test set.

[0047] Specifically, the target dataset images are randomly divided into a training set and a test set in a ratio of approximately 8:2.

[0048] S4, uses an attention enhancement algorithm to distinguish static obstacles and dynamic objects in the training set; S41, through the multi-scale grouping method, the variable radius in the training set is grouped into groups (0.2m / 0.8m / 2.0m). Among them, the small scale (0.2m) is used to capture small targets such as pedestrians; the large scale (2.0m) is used to extract global features such as road boundaries. By covering different target sizes at multiple scales, the small object missed detection rate is improved from 30% to 12%.

[0049] S42 adjusts the training set through the dynamic attention module, inputs the temporal difference data (including the data of the current frame and the previous frame), and calculates the dynamic probability weight based on the temporal difference data. When the weight of the dynamic point (such as a pedestrian) increases, the weight of the static point decreases, and when the weight of the dynamic point decreases, the static point increases. The characteristics of the radio target are enhanced first, and the recall rate is increased from 75% to 84%. The recall rate refers to an indicator used to measure the ability to recognize dynamic objects.

[0050] S43, by adjusting the cross-layer feature fusion of the training set, skip connections fuse low-level details (point cloud coordinates) with high-level semantics (such as the "vehicle" category) to avoid loss of deep network details. The feature integrity of complex occlusion scenes is improved by 22%, and the semantic segmentation index (mIoU) is improved from 60% to 75%.

[0051] The S44 processes the training set through semantic style and dynamic recognition, extracts features in layers through PointNet++ (sampling → grouping → classification), and outputs semantic labels (such as "pedestrian") and dynamic states (static / dynamic). This improves the accuracy of recognizing details such as rods by 20%, and is compatible with the Jetson Nano platform, an embedded development board developed by NVIDIA, reducing inference time from 50ms to 32ms.

[0052] This attention-enhanced PointNet++ is based on PointNet++, an upgraded version of PointNet, a point cloud deep learning framework proposed by Charles R. Qi et al. in 2017. Its hierarchical structure and local feature extraction significantly improve the accuracy and efficiency of point cloud analysis, making it particularly adept at processing complex scenes and large-scale point cloud data.

[0053] S5, uses path planning algorithms to identify state spaces in complex environments; The paths in the training set model are optimized using a bidirectional search strategy. The search tree is expanded simultaneously at the start and end points, terminating when they meet. High-threat areas (such as nodes near pedestrians) are prioritized for expansion. This approach reduces node traversal by 42% and planning time from 120ms to 60ms (a 50% improvement).

[0054] Path planning achieves efficient obstacle avoidance and path optimization in complex environments by integrating dynamic learning and heuristic search.

[0055] The dynamic cost function was optimized, and the agent was trained through reinforcement learning (PPO algorithm). The dynamic weights were trained: pedestrian threat coefficient = 2, vehicle = 1.5, obstacle avoidance detour distance was reduced by 15%, and path length was shortened from 1.2 km to 1.08 km (10%).

[0056] The Proximal Policy Optimization (PPO) algorithm is a policy optimization algorithm for reinforcement learning. It aims to solve some key problems in the policy gradient method, such as unstable training and low sample utilization efficiency.

[0057] Through incremental replanning, only the locally affected area (such as 5 meters around a sudden obstacle) is updated, reducing the computational workload by 68% and the replanning delay from 110ms to 40ms, bringing real-time performance close to millisecond requirements and avoiding 70% of potential collisions.

[0058] Incremental replanning is an efficient path planning method that significantly reduces computational effort and improves response speed by reusing the results of previous planning instead of replanning from scratch each time.

[0059] Model training process, state space: contains obstacle speed / semantic labels. By determining the speed of the obstacle, the agent can predict the possible future location of the obstacle, thereby better performing path planning and obstacle avoidance operations. Semantic labels can help the agent understand the different elements and their characteristics in the environment, so that the different influences of these elements can be considered when making decisions. Reward function: A collision penalty of 1000 points means that when the agent collides with an obstacle, it will receive a large negative reward (-1000 points). This high penalty is intended to encourage the agent to avoid collisions as much as possible, because collisions are usually dangerous or failure situations that need to be avoided in practical applications. The shortest path is +1 point / meter, that is, the agent will receive a reward of +1 point for every meter it moves along the shortest path, thereby achieving the effect of identifying the complex state space in the training set model.

[0060] S6, based on convolutional neural networks combined with recurrent neural networks, evaluates environmental complexity and dynamically adjusts control parameters to form an adaptive navigation strategy update mechanism; Refer to 1. The familiarity threshold is adjusted through environmental perception (CNN). A training set is input, including lidar projection images and visual images. The environmental complexity score (0-100) is output and the speed threshold is dynamically adjusted. By scoring the environmental complexity, the speed adjustment accuracy reaches 82%, and the navigation success rate in heavy rain scenarios is increased from 60% to 78%.

[0061] Convolutional neural network (CNN) is a deep learning model designed specifically for processing grid-like data (such as images, point cloud projections, and time series). It significantly reduces parameter size and improves feature expression capabilities through local perception, weight sharing, and multi-layer feature extraction.

[0062] By adjusting PID parameters through time series prediction (RNN), inputting historical IMU data and environmental score sequences, and predicting environmental changes (such as crosswind / pedestrian intrusion probability) in the next 500ms, the sudden crosswind response delay was reduced from 200ms to 60ms, and the steering delay was reduced from 0.8s to 0.45s (-44%).

[0063] A RNN (Recurrent Neural Network) is a deep learning model designed specifically for processing sequential data (such as time series, text, and sensor signals). It captures temporal dependencies in data through recurrent connections between hidden layer neurons. Its core advantage lies in its ability to model the impact of past information on current decisions. It is widely used in fields such as speech recognition, natural language processing, and sensor signal prediction.

[0064] After adjusting the speed threshold and PID (feedback control algorithm) parameters, multi-mode adaptive switching is used for control, including safety mode and emergency mode. Safety mode (activated when the score is greater than 70 points): full sensor fusion, emergency mode: triggering maximum braking / steering, the highest priority, the multi-mode switching success rate is 92%, and the stability in extreme scenarios is improved (such as strong wind attitude error of ±2.5° to ±5°).

[0065] Real-time feedback optimization is adopted, and the sensor provides real-time feedback to correct the model prediction error, forming a "perception-decision-execution" closed-loop control. After switching modes, control instructions are generated for control, and the control results are transmitted to the actuator and fed back to the sensor in real time.

[0066] As an example, the drone's altitude fluctuation error has been reduced from ±2 meters to ±1 meter, cargo bumps have been reduced by 65%, and control accuracy has been significantly improved.

[0067] For example, a neural network-based adaptive navigation control algorithm monitors environmental changes in real time and automatically adjusts vehicle speed and steering angle to ensure safe driving when encountering unexpected situations, such as a preceding vehicle changing lanes or a pedestrian crossing the road. For example, if the system detects a sudden lane change, it can adjust within 0.3 seconds to maintain a safe following distance. And if a pedestrian crosses the road, the system can initiate emergency braking within 0.5 seconds, effectively avoiding a collision.

[0068] S7, the test set is sent into the constructed network model to complete the modeling and perform navigation.

[0069] Specifically, the test set is fed into the constructed system network model to complete the modeling, and navigation is performed based on the modeling. For specific improvements, please refer to Table 1 and Table 2.

[0070] Table 1 Comparison of progress between the present invention and the prior art

[0071] Table 2 Comparative data performance comparison table between the prior art and the present invention

[0072] The embodiment of the present invention provides an artificial intelligence-based laser radar environment modeling and navigation system, which obtains three-dimensional point cloud data in a complex environment to obtain an initial data set; pre-processes the initial data set by data enhancement to obtain a target data set; uses a deep learning model based on a generative adversarial network to denoise and complete the initial data set to obtain a target data set; divides the target data set into a training set and a test set; uses an attention enhancement algorithm to distinguish static obstacles and dynamic objects in the training set; uses a path planning algorithm to identify the state space in the complex environment; evaluates the complexity of the environment based on a convolutional neural network combined with a recurrent neural network and dynamically adjusts control parameters to form an adaptive navigation strategy update mechanism; sends the test set to the constructed network model to complete modeling and perform navigation. The artificial intelligence-based laser radar environment modeling and navigation system constructed by the present invention achieves the effect of improving the system's environmental modeling accuracy through dynamic occlusion repair and dynamic obstacle recognition, improves path planning efficiency and real-time performance through dynamic adaptive adjustment and multi-scene adaptive recognition, and improves the navigation control adaptability. By adjusting the system algorithm and changing the hardware adaptation conditions, the system engineering and generalization effects are improved.

[0073] The second embodiment of the present invention relates to a network side server, such as Figure 6 As shown, it includes at least one processor 302; and a memory 301 that is communicatively connected to the at least one processor 302; wherein the memory 301 stores instructions that can be executed by the at least one processor 302, and the instructions are executed by the at least one processor 302 to enable the at least one processor 302 to execute the above-mentioned data processing method. Memory 301 and processor 302 are connected using a bus. The bus can include any number of interconnected buses and bridges, connecting various circuits of one or more processors 302 and memory 301. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits. These are all well known in the art and are therefore not described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 302 is transmitted over a wireless medium via an antenna. Furthermore, the antenna receives data and transmits it to processor 302.

[0074] The processor 302 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. The memory 301 can be used to store data used by the processor 302 when performing operations.

[0075] A third embodiment of the present invention relates to a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent rice field weed detection method based on multimodal data fusion in the first embodiment.

[0076] The steps of the various methods above are divided only for the purpose of clear description. During implementation, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process without changing the core design of the algorithm and process are all within the scope of protection of this patent.

[0077] The above is only an embodiment of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A laser radar environment modeling and navigation system based on artificial intelligence, characterized in that: include: S1, obtain 3D point cloud data in a complex environment to obtain the initial data set; S2, uses a deep learning model based on generative adversarial networks to denoise and complete the initial dataset to obtain the target dataset; S3, dividing the target data set into a training set and a test set; S4, uses an attention enhancement algorithm to distinguish static obstacles and dynamic objects in the training set; S5, uses path planning algorithms to identify state spaces in complex environments; S6, based on convolutional neural networks combined with recurrent neural networks, evaluates environmental complexity and dynamically adjusts control parameters to form an adaptive navigation strategy update mechanism; S7, the test set is sent into the constructed network model to complete the modeling and perform navigation.

2. The artificial intelligence-based laser radar environment modeling and navigation system according to claim 1, characterized in that: In S2, a deep learning model based on a generative adversarial network is used to denoise and complete the initial dataset, resulting in a target dataset including: S21, which uses the initial dataset as multi-source data input, inputs multi-frame time series point clouds (including timestamps) + IMU attitude data input (acceleration / angular velocity), performs spatiotemporal alignment (extended Kalman filter), repairs dynamic occlusion holes (such as pedestrians blocking vehicles) in the multi-frame time series, and uses IMU attitude to assist in spatial positioning to reduce completion errors; An IMU is a sensor module that can measure an object's acceleration, angular velocity, and (sometimes) magnetic field.

3. The artificial intelligence-based laser radar environment modeling and navigation system according to claim 2, characterized in that: S22, perform point cloud feature processing on the initial dataset and extract geometric features (curvature / density) through the PointConv layer. The PointConv layer supports unordered point clouds. The PointConv layer is used for deep hierarchical feature learning networks of point sets in metric space; FPS is a technology that reduces the amount of video data or reduces the processing load by lowering the frame rate (Frames Per Second, FPS).

4. The artificial intelligence-based laser radar environment modeling and navigation system according to claim 3, characterized in that: In S4, an attention enhancement algorithm is used to distinguish static obstacles from dynamic objects in the training set, including: An attention-enhanced PointNet++ based on PointNet++, a point cloud deep learning framework that improves the accuracy and efficiency of point cloud analysis through hierarchical structure and local feature extraction; The variable radius is grouped by multi-scale grouping. The small scale is used to capture small targets such as pedestrians; the large scale is used to extract global features such as road boundaries.

5. The artificial intelligence-based laser radar environment modeling and navigation system according to claim 4, characterized in that: The dynamic attention module is adjusted, and the temporal difference data (including data of the current frame and the previous frame) is input. The dynamic probability weight is calculated based on the temporal difference data. When the weight of the dynamic point (such as a pedestrian) increases, the static point decreases. When the weight of the dynamic point decreases, the static point increases, and the characteristics of the radio target are enhanced first.

6. The artificial intelligence-based laser radar environment modeling and navigation system according to claim 5, characterized in that: The generator generates point clouds, predicts dynamic occlusion missing point clouds, filters noise points, and allows the generator to learn the laws of dynamic scenes and repair large areas of missing images.

7. The artificial intelligence-based laser radar environment modeling and navigation system according to claim 1, characterized in that: In S5, a path planning algorithm is used to identify the state space in a complex environment, including: optimizing the path in the training set model through a bidirectional search strategy, expanding the search tree at the starting point and the end point at the same time, terminating when they meet, and prioritizing the expansion of high-threat areas (such as nodes near pedestrians).

8. The artificial intelligence-based laser radar environment modeling and navigation system according to claim 1, characterized in that: In S6, based on a convolutional neural network combined with a recurrent neural network, the environment complexity is evaluated and the control parameters are dynamically adjusted to form an adaptive navigation strategy update mechanism, including: inputting a training set, which includes lidar projection images and visual images, outputting an environment complexity score, and dynamically adjusting the speed threshold.

9. A network side server, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the artificial intelligence-based lidar environment modeling and navigation system as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the artificial intelligence-based lidar environment modeling and navigation system according to any one of claims 1 to 8 is implemented.

Images