Ground mobile unmanned equipment autonomous obstacle avoidance control system optimized by artificial intelligence

Abstract

The invention relates to the technical field of ground mobile unmanned equipment control, and discloses a ground mobile unmanned equipment autonomous obstacle avoidance control system optimized by artificial intelligence. The system comprises an environment perception layer, a bimodal risk assessment layer, a dynamic decision-making layer, a trajectory optimization layer and a feedback optimization layer. The environment sensing layer adopts a retina fovea centralis imitating mechanism to perform non-uniform sampling on laser radar point cloud data to generate dynamic point cloud partitions; the bimodal risk assessment layer fuses two types of radar data to generate static and dynamic obstacle risk assessment diagrams; the dynamic decision-making layer establishes space-time mapping and generates an obstacle confidence coefficient matrix through a graph neural network; the trajectory optimization layer converts the matrix into a control parameter based on a multi-objective evolutionary algorithm, and issues the control parameter through a time-sensitive network protocol; and the feedback optimization layer monitors environment change, calculates deviation, generates an effectiveness index, and dynamically adjusts a point cloud acquisition strategy until the index is optimal. According to the system, the autonomous obstacle avoidance capability and adaptability of the ground mobile unmanned equipment in a complex environment are enhanced.

Description

Technical Field

[0001] This invention relates to the field of ground mobile unmanned equipment control technology, specifically to an autonomous obstacle avoidance control system for ground mobile unmanned equipment optimized using artificial intelligence. Background Technology

[0002] With the widespread application of ground-based mobile unmanned equipment in logistics, agriculture, security, and other fields, autonomous obstacle avoidance capability has become a core indicator for measuring its operational safety and reliability. Currently, most mainstream obstacle avoidance systems for ground-based mobile unmanned equipment rely on a single sensor to collect environmental data, such as lidar or millimeter-wave radar, which are susceptible to interference in complex and dynamic environments. While lidar can provide high-precision point cloud data, the data volume is enormous, making real-time processing difficult; millimeter-wave radar has strong penetration capabilities but lacks accuracy in static obstacle recognition.

[0003] In existing obstacle avoidance control methods, the environmental perception layer often uses uniform sampling to process point cloud data, resulting in excessive redundant information, consuming significant computational resources, and impacting system response speed. Risk assessment often relies solely on data from a single sensor, making it difficult to assess both dynamic and static obstacles, and prone to missed or incorrect assessments. The decision-making layer lacks an effective spatiotemporal correlation mechanism when fusing multi-source data, resulting in obstacle avoidance strategies that do not accurately match the actual environment. Furthermore, trajectory optimization does not adequately consider multi-objective constraints, leading to less smooth obstacle avoidance paths and affecting the operational stability of ground-based mobile unmanned equipment. In addition, most systems lack closed-loop feedback mechanisms, failing to dynamically adjust data acquisition and decision-making strategies according to environmental changes, resulting in poor adaptability in complex and variable environments. Summary of the Invention

[0004] The purpose of this invention is to provide an autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized by artificial intelligence, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence, the system comprising:

[0006] The environmental perception layer performs non-uniform sampling of lidar point cloud data based on a retinal fovea-like mechanism to generate dynamic point cloud partitions.

[0007] The dual-modal risk assessment layer deploys a millimeter-wave radar array based on the dynamic point cloud partitioning, and fuses lidar point cloud data and millimeter-wave radar echo data to generate a static obstacle risk assessment map and a dynamic obstacle risk assessment map.

[0008] In the dynamic decision-making layer, a spatiotemporal mapping between the static obstacle risk assessment map and the dynamic obstacle risk assessment map is established. A graph neural network is used to fuse dual-modal risk assessment data to generate an obstacle confidence matrix.

[0009] The trajectory optimization layer transforms the obstacle confidence matrix into obstacle avoidance trajectory control parameters based on a multi-objective evolutionary algorithm, and sends them to the ground mobile unmanned equipment drive unit via a time-sensitive network protocol.

[0010] The feedback optimization layer monitors the changes in the environmental state after executing the obstacle avoidance trajectory control parameters in real time, calculates the deviation between the environmental state changes and the preset safety threshold, generates an obstacle avoidance strategy effectiveness index, and dynamically adjusts the point cloud acquisition strategy until the obstacle avoidance strategy effectiveness index reaches its optimum.

[0011] Preferably, the method for non-uniform sampling of lidar point cloud data based on a foveal-like mechanism includes:

[0012] The space in front of the ground-mobile unmanned equipment is divided into a high-resolution core area and a low-resolution edge area, and a dynamic point cloud partitioning module is defined based on a preset historical distribution heat map of obstacles.

[0013] Dense voxel sampling is used in the high-resolution core region, and sparse voxel sampling is used in the low-resolution edge region.

[0014] The system uses a three-dimensional convolutional neural network to identify abnormal geometric features in point cloud data in real time, and dynamically adjusts the spatial range of the high-resolution core area based on the identification results.

[0015] Preferably, the method for generating static obstacle risk assessment maps and dynamic obstacle risk assessment maps includes:

[0016] The dynamic point cloud partitions are spatiotemporally aligned to output a standardized point cloud frame sequence.

[0017] The millimeter-wave radar array is deployed at the corresponding spatial location of the high-resolution core area, transmitting frequency-modulated continuous waves and receiving Doppler echo signals.

[0018] The phase change of the Doppler echo signal is analyzed by a moving target detection algorithm to generate a dynamic obstacle motion vector field;

[0019] The standardized point cloud frame sequence is processed by a point cloud feature pyramid network to output a static obstacle risk assessment map with spatial coordinate annotations.

[0020] By integrating the static obstacle risk assessment map with the dynamic obstacle motion vector field, a dynamic obstacle risk assessment map with risk level labels is generated.

[0021] Preferably, the method for generating a static obstacle risk assessment map includes:

[0022] A multi-scale point cloud feature extraction network is constructed, which includes a bottom-up geometric feature encoding path and a top-down semantic feature decoding path.

[0023] The curvature features, normal vector distribution features, and surface roughness features of the point cloud data are extracted layer by layer in the geometric feature encoding path.

[0024] In the semantic feature decoding path, high-level semantic features and low-level geometric features are fused across layers;

[0025] The obstacle risk probability value of each point cloud cluster is calculated using a spatial probability model, and a static obstacle risk assessment map containing the location, size, and risk level of static obstacles is generated.

[0026] Preferably, the method for generating a dynamic obstacle risk assessment map with risk level markers includes:

[0027] Establish the spatial mapping relationship between millimeter-wave radar echo signals and lidar point clouds, and define a dynamic obstacle tracking coordinate system;

[0028] The trajectory state of a moving target is predicted using the Kalman filter algorithm, and the residual between the predicted trajectory value and the actual observed value is calculated.

[0029] Based on the residuals, a motion uncertainty model of the dynamic obstacle is constructed, and the dynamic obstacle risk assessment map containing velocity vectors, acceleration, and risk confidence is output.

[0030] Preferably, the method for establishing the spatiotemporal mapping between the static obstacle risk assessment map and the dynamic obstacle risk assessment map includes:

[0031] A global spatial coordinate system is established with the centroid of the ground-mobile unmanned equipment as the origin, and the static obstacle risk assessment map is transformed into the global spatial coordinate system through external calibration parameters.

[0032] A hardware timestamp synchronization mechanism is used to align the time references of the static obstacle risk assessment map and the dynamic obstacle risk assessment map.

[0033] A spatial position mapping table for obstacles is established in the global spatial coordinate system, linking the coordinates of static obstacles with the coordinates of the motion trajectory of dynamic obstacles.

[0034] Preferably, the method for fusing dual-modal risk assessment data using graph neural networks includes:

[0035] Each obstacle region in the static obstacle risk assessment map is defined as a static node, and geometric feature vectors are extracted as static node attributes.

[0036] Each moving target in the dynamic obstacle risk assessment map is defined as a dynamic node, and the motion feature vector is extracted as the attribute of the dynamic node.

[0037] Construct a heterogeneous graph structure containing all static and dynamic nodes, and define the spatial proximity relationship between nodes as connecting edges;

[0038] By aggregating the feature information of adjacent nodes through a multi-head graph attention mechanism, a comprehensive risk score matrix for each obstacle region is output.

[0039] Preferably, the method for converting the obstacle confidence matrix into obstacle avoidance trajectory control parameters based on a multi-objective evolutionary algorithm includes:

[0040] Three types of optimization objectives are set: safe passage objective, energy consumption optimization objective, and trajectory smoothing objective.

[0041] Initialize multiple sets of candidate trajectory control parameters, each set of parameters including steering angle sequence, velocity sequence and acceleration sequence;

[0042] Evaluate the avoidance effect of each set of candidate trajectory control parameters on the comprehensive risk score matrix, and calculate the achievement score of the three optimization objectives;

[0043] The Pareto optimal solution set is iteratively selected, and the set of candidate trajectory control parameters with the highest achievement score is output as the obstacle avoidance trajectory control parameters.

[0044] Preferably, the method for generating the obstacle avoidance strategy effectiveness index includes:

[0045] Collect environmental perception data when actually executing obstacle avoidance trajectory control parameters, and construct the minimum distance change curve of obstacles;

[0046] Calculate the root mean square error between the minimum distance variation curve of the obstacle and the preset safe distance curve;

[0047] The root mean square error is normalized, and the resulting output is an effectiveness index of the obstacle avoidance strategy, ranging from zero to one.

[0048] Preferably, the method for dynamically adjusting the point cloud acquisition strategy includes:

[0049] Preset a first threshold and a second threshold for the effectiveness index of the obstacle avoidance strategy;

[0050] When the obstacle avoidance strategy effectiveness index is lower than the first threshold of the obstacle avoidance strategy effectiveness index, the spatial range and sampling frequency of the high-resolution core area are increased.

[0051] When the obstacle avoidance strategy effectiveness index is between the first threshold and the second threshold of the obstacle avoidance strategy effectiveness index, the current point cloud acquisition strategy is maintained.

[0052] When the obstacle avoidance strategy effectiveness index is higher than the second threshold of the obstacle avoidance strategy effectiveness index, the spatial range of the high-resolution core area is reduced by a preset ratio.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] By employing a foveal-like mechanism in the environmental perception layer to perform non-uniform sampling of LiDAR point cloud data, redundant data can be reduced while preserving detailed information in key areas. This lowers the computational burden on subsequent processing stages and improves the system's response speed to environmental changes. The generation of dynamic point cloud partitions makes sensor data acquisition more targeted, providing accurate foundational data for subsequent risk assessment.

[0055] The dual-modal risk assessment layer fuses LiDAR and millimeter-wave radar data to generate static and dynamic obstacle risk assessment maps, taking into account the identification characteristics of different types of obstacles. The high precision of LiDAR improves the accuracy of static obstacle assessment, while the sensitivity of millimeter-wave radar to dynamic targets enhances the timeliness of dynamic obstacle monitoring. The complementary fusion of the two data sources reduces the limitations of a single sensor, making the risk assessment results more comprehensive and reliable.

[0056] The dynamic decision-making layer establishes a spatiotemporal mapping of a dual-modal risk assessment map. Leveraging the powerful correlation modeling capabilities of graph neural networks, it effectively uncovers the intrinsic connections between obstacle data from different times and regions. The generated obstacle confidence matrix more accurately reflects the actual distribution and changing trends of obstacles in the environment, providing a scientific basis for subsequent trajectory optimization. This fusion approach overcomes the problem of insufficient spatiotemporal correlation in traditional data fusion, improving the rationality of decision-making.

[0057] The trajectory optimization layer, based on a multi-objective evolutionary algorithm, transforms the obstacle confidence matrix into control parameters. This achieves a balance between multiple objectives, including obstacle avoidance safety, path smoothness, and operational efficiency. The generated trajectory better matches the motion characteristics of ground-based mobile unmanned equipment, reducing bumps and energy consumption during operation. Parameters are distributed via a time-sensitive networking protocol, ensuring the real-time performance and reliability of control commands and preventing obstacle avoidance errors due to command delays.

[0058] The feedback optimization layer forms a closed-loop control system, which monitors environmental changes in real time and calculates deviations. The generated obstacle avoidance strategy effectiveness index provides a quantitative reference for the system's self-adjustment. The point cloud acquisition strategy is dynamically adjusted based on the index, enabling the system to adaptively allocate acquisition resources according to environmental complexity. In simple environments, it reduces data volume to improve efficiency, while in complex environments, it increases sampling in key areas to ensure accuracy, thus maintaining good obstacle avoidance performance in different scenarios. Attached Figure Description

[0059] Figure 1 This is a timing diagram of the autonomous obstacle avoidance control system for ground mobile unmanned equipment optimized by artificial intelligence as described in this invention.

[0060] Figure 2 Flowcharts generated for static and dynamic risk assessment diagrams;

[0061] Figure 3 A flowchart generated for a static obstacle risk assessment diagram;

[0062] Figure 4 A flowchart for dual-modal risk assessment data fusion;

[0063] Figure 5 A flowchart for generating obstacle avoidance trajectory control parameters. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Please see Figure 1 This invention provides an autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence, the system comprising:

[0066] Intelligent obstacle avoidance decision-making in dynamic environments is achieved through biomimetic visual perception and multimodal data fusion. The environmental perception layer incorporates a lidar processing module mimicking the fovea mechanism of the retina, establishing a non-uniform sampling area in front of the device to convert raw point cloud data into a dynamically partitioned structure. The dual-modal risk assessment layer, configured with a millimeter-wave radar array working in conjunction with lidar, constructs both static and dynamic risk assessment models. The dynamic decision-making layer uses a graph neural network to process the spatiotemporally mapped risk assessment data, generating a confidence matrix that includes the comprehensive threat level of obstacles. The trajectory optimization layer employs a multi-objective evolutionary algorithm to transform the risk assessment results into a sequence of optimal control parameters, achieving microsecond-level command transmission through a time-sensitive network. The feedback optimization layer continuously monitors changes in environmental conditions, establishing a quantitative evaluation mechanism for obstacle avoidance effectiveness, forming a closed-loop control system.

[0067] Example 1: See Figure 2Non-uniform sampling processing of the LiDAR point cloud is performed by a spatial partitioning processor, which integrates a historical obstacle distribution heatmap database. The 120-degree conical field of view in front of the device is divided into a core high-resolution area with a diameter of 15 meters and an outer low-resolution edge area. The core area uses a 0.1-meter cubic pixel grid for dense sampling, while the edge area uses a 0.5-meter cubic pixel grid for sparse sampling. A 3D convolutional neural network processes the input data at a constant operating frequency of 10Hz. The network architecture includes a three-layer feature extraction structure: the first layer uses a 5×5×5 convolutional kernel for primary geometric feature capture, the second layer uses a 3×3×3 convolutional kernel to refine the feature expression, and the third layer uses a 1×1×1 convolutional kernel to complete feature compression. The anomaly detection module continuously monitors the point cloud curvature gradient and normal vector mutation features. When a vertical height change exceeding 2 meters is detected for three consecutive frames and the horizontal distribution dispersion is below a threshold, the partitioning adjustment mechanism is automatically triggered. At this time, the core area diameter expands to 20 meters, the voxel resolution of the edge area increases to 0.3 meters, and the LiDAR scan line distribution density increases by 30% in the expanded area.

[0068] The millimeter-wave radar array consists of eight 77GHz transceiver modules, arranged at 15-degree equidistant intervals along the leading edge arc of the device. The transmitting unit employs a sawtooth wave frequency-modulated continuous wave mode, with waveform parameters set as follows: starting frequency 76GHz, bandwidth 1GHz, and sweep period 2ms. The receiving channel is configured with a digital down-conversion processing chain, first shifting the echo signal to baseband via a mixer, and then generating a digital intermediate frequency signal through 24-bit ADC sampling. The Doppler processing unit performs a fast Fourier transform on 256 continuous pulses, achieving a velocity resolution accuracy of 0.2m / s. The moving target detection algorithm uses phase difference technology, calculating the radial velocity component by comparing the phase offset of adjacent echo signals, while simultaneously combining multi-frame correlation filtering to eliminate false targets.

[0069] The point cloud data processing workflow includes spatiotemporal alignment preprocessing, using nearest neighbor interpolation to map the original point cloud to a unified timestamp coordinate system. The feature pyramid network adopts a five-level hierarchical architecture: the first level processes the original point cloud to generate basic feature maps; the second level performs feature downsampling through sparse convolutions with a stride of 2; the third level fuses the features from the first two levels and extracts local geometric structures; the fourth level constructs global context features; and the fifth level gradually restores spatial resolution through transposed convolutions. The network output layer connects to three parallel branches: the position branch regresses the 3D bounding box coordinates of obstacles, the size branch predicts the physical scale of obstacles, and the risk branch calculates the static obstacle threat value. The static risk assessment map is presented in the form of a point cloud, with each point containing 3D coordinates and a risk value attribute. Fixed obstacles are marked in red, and temporary stationary objects are marked in yellow. The risk level is represented by a color transparency gradient (transparency from 0.1 to 0.9 corresponds to risk levels 1 to 9).

[0070] The dynamic risk assessment generation process first establishes a rigid transformation matrix between the millimeter-wave radar coordinate system and the lidar coordinate system, and determines the relative pose between the sensors through hand-eye calibration. The motion vector field construction employs data association technology, associating the trajectories of five consecutive frames of detected moving targets using a Hungarian algorithm. Each dynamic target's state vector includes position, velocity, acceleration, and trajectory curvature. Risk confidence is calculated based on the stability of the target's motion state: when the target's rate of acceleration exceeds a set threshold or the trajectory prediction residual increases, the confidence decays exponentially. The final generated dynamic risk assessment map uses a vector arrow visualization scheme. The arrow's starting point indicates the target's current position, its direction indicates the direction of motion, its length is proportional to the velocity magnitude, and its color coding transitions from green (low risk) to purple (high risk). A three-decimal-digit confidence value is displayed next to the arrow.

[0071] Sensor data fusion is performed after spatiotemporal alignment. LiDAR point clouds and millimeter-wave radar detected targets are associated using a nearest-neighbor distance matching algorithm. A correspondence is established when the spatial distance between the two detection results is less than 0.5 meters. Unmatched LiDAR point clouds are considered static obstacles, and unmatched millimeter-wave targets initiate temporary dynamic trajectory tracking. Static and dynamic risk assessment maps are overlaid in a unified coordinate system to form a complete risk distribution model of the surrounding environment. This model is updated every 100ms and transmitted to the decision-making processing unit via gigabit Ethernet.

[0072] Example 2: See Figure 3 The multi-scale point cloud feature extraction network adopts a bidirectional information transmission architecture, including a bottom-up geometric feature encoding path and a top-down semantic feature decoding path. The geometric feature encoding path is configured with four processing stages: The first stage performs voxelization on the input point cloud, divides the space using a 0.2-meter cubic grid, and extracts the curvature tensor features within each voxel through sparse convolution operations, calculating the rate of curvature change of the point cloud surface in the three principal directions; The second stage downsamples the feature map to half the original resolution, extracts the normal vector distribution features, and statistically analyzes the cluster center and dispersion of the point cloud normal vectors within each voxel unit; The third stage further downsamples to one-quarter resolution, analyzes the surface roughness features, and calculates the correlation between the point cloud reflection intensity and the local plane fitting residual; The fourth stage generates a global feature map at one-eighth resolution to capture large-scale geometric structure information. The semantic feature decoding path adopts a mirror-symmetric structure: the first stage receives the global features from the fourth stage of the encoding path, and restores them to one-quarter resolution through transposed convolution upsampling; the second stage concatenates the upsampled features with the features from the third stage of the encoding path, and then upsamples them again to half resolution; the third stage similarly fuses the corresponding layer features from the encoding path, finally outputting a semantic feature map with the same resolution as the input point cloud. The cross-layer fusion module uses a gating mechanism, controlling the fusion ratio of low-level geometric features and high-level semantic features through learnable weights.

[0073] The spatial probability model is constructed based on a Gaussian mixture distribution. The obstacle risk probability calculation for each point cloud cluster includes three dynamically adjusted weighting factors. The first weighting factor, distance, is inversely proportional to the current position of the ground-based mobile unmanned equipment. It is calculated using a hyperbolic decay function, maintaining a constant weight value within 5 meters in front of the equipment, and decreasing with increasing distance beyond 5 meters. The second weighting factor, historical frequency, is obtained by querying a historical database that records the number of times obstacles have appeared in each spatial grid within the past 300 seconds, using a sliding time window update mechanism. The third weighting factor, geometric stability, assesses the morphological changes of the point cloud cluster, calculating the standard deviation of the bounding box dimensions and the variance of the center point displacement over five consecutive frames. The final risk probability value is calculated by the weighted product of the three weighting factors. The output static obstacle risk assessment map contains three types of information: obstacle location represented by the center coordinates of the 3D bounding box, dimension information recording the length, width, and height data of the bounding box, and a nine-level quantitative scale for the risk level.

[0074] The dynamic obstacle tracking coordinate system is established in polar coordinates, with the centroid of the ground-based mobile unmanned equipment as the origin. A 0.5-meter resolution scale ring is used for the radial coordinates, and a 360-degree azimuth grid is established with a minimum division of 1 degree for the angular coordinates. The millimeter-wave radar echo signal undergoes digital down-conversion processing, and the Doppler frequency shift component is extracted using a phase demodulator, achieving a velocity calculation accuracy of 0.1 m / s. A Kalman filter is configured with a six-dimensional state vector, including radial distance, azimuth angle, radial velocity, tangential velocity, radial acceleration, and tangential acceleration components. The state transition matrix adopts a constant acceleration model, and the observation matrix selects distance and radial velocity as observations. Trajectory prediction residual calculation is performed in each processing cycle, comparing the Euclidean distance between the predicted state value and the actual observed value.

[0075] The motion uncertainty model uses Mahalanobis distance to measure the statistical properties of the predicted residuals. A sliding window statistical model of the residual covariance matrix is ​​established to store the residual vectors for the most recent 20 periods. When the Mahalanobis distance of a new residual vector exceeds three standard deviations, a risk confidence decay mechanism is triggered. The confidence decay function adopts an exponential descent model, with an initial confidence level set to 0.95. Each decay event multiplies the confidence level by a decay factor of 0.7, with a minimum limit of 0.3. Simultaneously, when the predicted residuals are below a threshold for five consecutive frames, a confidence recovery procedure is initiated, increasing the recovery value by 0.05 each frame until the initial level is reached.

[0076] The vector field visualization of the dynamic risk assessment map employs a composite coding scheme. Each dynamic target is represented by a conical arrow, with the arrow's origin positioned at the target's current radial location and its direction aligned with the direction of motion. The arrow length is proportional to the velocity modulus, with a baseline length of 10 cm corresponding to a velocity of 1 m / s. The color coding scheme defines the HSV color space: hue gradients from 120 degrees (green) to 300 degrees (purple), corresponding to risk levels 1 to 9; saturation is fixed at 100%; and brightness varies linearly with confidence level. A text label attached to the arrow tip displays the risk confidence value with three decimal places. The motion trajectory prediction line is represented by a semi-transparent strip graphic, with bandwidth proportional to the acceleration modulus. The color gradient within the strip reflects the risk level changes of each trajectory point within the next three seconds.

[0077] Example 3: See Figure 4 The global spatial coordinate system is established with the center of mass of the ground-based mobile unmanned equipment as the origin. The coordinate axes are determined according to the right-hand rule: the positive X-axis points in the direction of the equipment's movement, the positive Y-axis points to the left side of the equipment, and the positive Z-axis extends vertically upwards. External calibration parameters include the installation position deviation matrix of the lidar and millimeter-wave radar, represented by a 4×4 homogeneous transformation matrix, recording the three-dimensional translation and three-axis rotation of the sensors relative to the equipment's center of mass. The equipment's center of mass projection compensation value is calculated in real time by the dynamic load monitoring system, triggering an update of the compensation parameters when the load change exceeds 5%. The spatial coordinate system transformation process uses a quaternion rotation interpolation algorithm to transform the static obstacle risk assessment map from the sensor coordinate system to the global coordinate system, with transformation accuracy controlled at the millimeter level.

[0078] The time synchronization system employs Precision Time Protocol version 2 (PTP 2) for hardware-level clock alignment, with the master clock source being the pulse-per-second signal from the Global Navigation Satellite System (GNSS) receiver. Timestamp units are integrated into the data acquisition boards of each sensor. When the lidar completes a frame scan or the millimeter-wave radar completes a frequency sweep cycle, a hardware timestamp accurate to the nanosecond level is immediately written into the data packet header. The time alignment processor establishes a circular buffer to interpolate and resample the static and dynamic risk assessment maps, ensuring that the time deviation between the two sets of data does not exceed 100 microseconds. Synchronized data frames are appended with a unified timestamp, formatted as a microsecond count since system startup.

[0079] The spatial location mapping of obstacles is managed through a quadtree index structure, which establishes spatial partitions in the horizontal XY plane. The root node covers a 50m x 50m area around the device, and each child node is recursively divided into the smallest 1m x 1m unit. The location mapping table stores three types of relationships: the binding relationship between the unique identifier of a static obstacle and the center coordinates of its bounding box; the correspondence between the trajectory identifier of a dynamic obstacle and its motion state vector; and records of the spatial proximity relationships between static and dynamic obstacles. The spatial proximity relationship is calculated using the position difference formula:

[0080]

[0081] in: Indicates the degree of positional difference. , , These represent the distance differences between the two obstacles along the three coordinate axes. , , These are the weighting coefficients for each axis. When When the value is less than the set threshold, an associated entry for static obstacles and dynamic obstacles is created in the mapping table, and the historical position data of the last ten frames is recorded.

[0082] The heterogeneous graph builder defines two node types: static nodes correspond to each obstacle region in the static risk assessment graph, and dynamic nodes correspond to each moving target in the dynamic risk assessment graph. The static node feature vector extracts 128-dimensional attribute data, including geometric features (bounding box aspect ratio, mean surface curvature), material features (laser reflection intensity distribution, point cloud density), and spatiotemporal stability features (positional movement variance, rate of change of size). The dynamic node feature vector contains 128-dimensional attributes, including 64-dimensional motion features (velocity vector, acceleration vector, trajectory curvature radius) and 64-dimensional morphological features (radar cross section, Doppler spectral width).

[0083] The establishment of node connections employs a dual-criteria system: the spatial proximity criterion requires that the Euclidean distance between two nodes in the global coordinate system be less than 5 meters; the motion trajectory prediction criterion calculates the shortest intersection time between the predicted trajectory of a dynamic node and the position of a static node, establishing a connection when the intersection time is less than 3 seconds. Each connection edge stores three types of attributes: spatial distance value, relative velocity vector angle, and risk coupling factor. The graph structure is updated every 200 milliseconds; newly added obstacle nodes are automatically added, while disappeared obstacle nodes are retained for five frames before being removed.

[0084] The graph attention mechanism employs eight parallel computation heads, each with its feature transformation matrix trained independently. The query vector dimension is set to 32, extracted from node feature vectors through linear transformation; the key and value vector dimensions are set to 64, also obtained through feature transformation. Adjacent node features are aggregated using a weighted summation method, with attention weights calculated using three factors: cosine similarity of node features, inverse spatial distance, and risk level product. After multi-head output features are fused in the concatenation layer, they are reduced to a 128-dimensional feature vector through a fully connected layer.

[0085] The comprehensive risk score matrix generation module receives the updated node feature vectors and calculates the risk interaction strength between each static node and all dynamic nodes. The score calculation includes three components: static risk base value, dynamic risk propagation value, and coupled risk additional value. The output matrix row index corresponds to the static obstacle number, and the column index corresponds to the dynamic obstacle number. The matrix element values ​​are compressed to the range of 0 to 1 using the Sigmoid function. An additional dimension of the matrix stores the comprehensive risk score of each static region affected by the dynamic target, which is extracted from the corresponding row vector through a max pooling operation.

[0086] The spatiotemporal mapping consistency verification module continuously monitors data validity and triggers a data reset when the following anomalies are detected: the deviation between the position mapping of static and dynamic nodes exceeds 2 meters for three consecutive frames; the abrupt change in the entropy value of the graph attention weight distribution exceeds twice the standard deviation of the historical mean; and the outlier of the risk scoring matrix increases abnormally. When the verification fails, the system automatically reverts to the previous valid state for reprocessing and marks the abnormal data frames for offline analysis. All output data is encapsulated using a time-sensitive network protocol and transmitted to the trajectory optimization layer within a specified time window.

[0087] Example 4: See Figure 5 The multi-objective optimization module receives an obstacle confidence matrix from the decision layer. The row vectors of this matrix correspond to static obstacle region numbers, and the column vectors correspond to dynamic obstacle target numbers. The matrix element values ​​represent the comprehensive risk score for a specific obstacle combination. The optimization objective function is defined in three dimensions: the safe passage objective requires the trajectory to maintain the minimum safe distance from all obstacles; the energy consumption optimization objective limits the magnitude of acceleration changes; and the trajectory smoothing objective constrains the rate of change of the steering angle. The weights of the three objectives are allocated as follows: safety objective 0.6, energy consumption objective 0.2, and smoothing objective 0.2.

[0088] The initial candidate parameter set generator creates 50 trajectory control schemes, each containing a parameter sequence of 20 consecutive control cycle points. Each control cycle point includes three control variables: steering angle (range -30° to +30°), velocity (range 0-10 m / s), and acceleration (range -3 to +3 m / s²). The parameter set generation uses Latin hypercube sampling to ensure uniform coverage of the parameter space. The table below shows example data for the first five control cycle points of candidate parameter set #7 (see Table 1).

[0089] Table 1: Example Table of Candidate Trajectory Control Parameter Sets (Partial)

[0090] The safe passage score calculator iterates through the sequence of trajectory points for each candidate parameter set. For each trajectory point, it calculates the minimum Euclidean distance to the surfaces of all static obstacle bounding boxes and the shortest distance to the predicted trajectory line of the dynamic obstacle. The minimum of these distance values, d_min, is taken, and the safety score for that point is calculated as 1 / (d_min+ε), where ε is a small constant to prevent division by zero. The overall safety target score for the entire trajectory is the maximum score of all trajectory points, reflecting the risk level of the most dangerous contact point.

[0091] The energy consumption score is calculated using the integral method of acceleration change. The absolute value of acceleration at each control cycle point is accumulated, while considering a penalty term for abrupt changes in acceleration direction. When the sign of acceleration changes between adjacent cycles, an additional penalty coefficient is added. The final energy consumption score is normalized to a value between 0 and 1, with a smaller value indicating better energy efficiency.

[0092] The trajectory smoothing scorer focuses on the characteristics of steering angle changes. First, it calculates the first-order difference sequence of the steering angle, reflecting the rate of change; then, it calculates the second-order difference sequence, characterizing the acceleration of the steering angle change. The smoothing score consists of two parts: the average of the absolute values ​​of the first-order differences reflects the overall magnitude of the change, and the sum of the absolute values ​​of the second-order differences characterizes the degree of drastic change. The weighted sum of the two parts is then converted into a score value using an inverse proportional function.

[0093] The evolutionary optimization engine employs an improved non-dominated sorting genetic algorithm. The initial population contains 50 individuals, each encoding a complete sequence of trajectory parameters. The selection operator uses a binary tournament mechanism, randomly selecting two individuals each time and comparing the vectors formed by their three-dimensional objective function values. The dominance criterion is: if individual A's safety score is less than B's, its energy consumption score is greater than B's, and its smoothness score is greater than B's, then individual B dominates A. Non-dominated sorting divides the population into multiple frontier levels, with the first frontier level representing the current optimal solution set.

[0094] The crossover operation is performed between two selected parent individuals. Five consecutive control points in the trajectory sequence are randomly selected as the crossover segment, and the parameter values ​​of the two parents are exchanged in this segment. The mutation operation occurs probabilistically. When mutation is triggered, three control points are randomly selected, and Gaussian noise with a standard deviation of 2 is added to their steering angle parameters. The velocity parameters are randomly perturbed within a range of ±10%, and the acceleration parameters vary within ±0.3 m / s².

[0095] Each generation of evolution maintains a population size of 100. After crossover and mutation to generate new individuals, the parent and offspring generations are merged to form a set of 200 individuals. After non-dominated sorting, the top 100 optimal individuals are selected to enter the next generation. The evolutionary process continues for 50 generations and then terminates, outputting all non-dominated solutions of the first frontier level to form a Pareto optimal solution set.

[0096] The optimal solution selector employs a fuzzy decision-making mechanism. For each individual in the Pareto solution set, the satisfaction membership degree of the three objective functions is calculated. The safety objective membership function is defined as a trapezoidal distribution: satisfaction is 1 when the safety score is less than 0.3; it linearly decreases to 0.5 in the interval between 0.3 and 0.6; and satisfaction is 0 when it is greater than 0.6. The membership functions for the energy consumption objective and the smoothing objective adopt a similar design, but the threshold intervals are adjusted accordingly. Finally, the individual with the largest product of the satisfaction scores of the three objectives is selected as the optimal trajectory scheme.

[0097] The parameter sequence post-processor performs kinematic smoothing on the selected trajectory. Cubic B-spline curves are used to fit the original parameter points, generating continuously differentiable steering angle functions, velocity curves, and acceleration curves. The smoothed control parameter sequence is packaged and transmitted via a time-sensitive network protocol. The data packet contains a timestamp sequence and the corresponding three-axis control parameters, with the transmission cycle strictly controlled within 2 milliseconds. The drive interface module converts the control parameters into motor steering angle commands and power unit torque commands, achieving precise trajectory tracking.

[0098] Example 5: The environmental condition monitoring module collects multi-source sensor data at a fixed frequency of 100 Hz to construct a three-dimensional observation vector. The first dimension of this vector records the real-time distance between the ground-based mobile unmanned equipment and the nearest obstacle, obtained through a nearest neighbor search algorithm using lidar point clouds. The second dimension measures the relative velocity, calculated by fusing millimeter-wave radar Doppler frequency shift and visual optical flow features. The third dimension collects obstacle azimuth information, using polar coordinate transformation to process the global coordinate system position data. A timestamp is added to the data acquisition process, achieving microsecond-level accuracy and maintaining strict time synchronization with the control system clock.

[0099] The safety distance curve generator dynamically adjusts the baseline value based on the current motion state of the equipment. The basic curve shape adopts a piecewise linear design, dividing the area in front of the equipment into three zones: the zero to two-meter range is the emergency braking zone, with a negative constant curve slope; the two to five-meter range is the buffer adjustment zone, with the absolute value of the slope increasing with speed; and the area beyond five meters is the early warning monitoring zone, maintaining a constant safety threshold. The real-time curve parameter update mechanism is linked to the speed sensor input. When the equipment acceleration exceeds the set threshold, the emergency braking zone range is automatically compressed and the buffer zone slope is increased.

[0100] The obstacle minimum distance change curve plotter continuously records the actual minimum distance value when executing the obstacle avoidance trajectory. The data buffer stores distance observations from the last two hundred control cycles, sampled at 20-millisecond intervals. The curve smoothing processor employs a moving average filtering technique with a window width of five sampling points to eliminate jitter caused by sensor noise. The curve coordinate system uses time series on the horizontal axis and distance scale on the vertical axis, with the time resolution consistent with the control system cycle.

[0101] The root mean square error (RMSE) calculation unit receives input data from the safe distance curve and the measured distance curve. The processing first aligns the two curves along their time axes, using linear interpolation to unify the sampling point positions. The error sequence generator calculates the difference in the ordinate at each corresponding time point, squares it, and stores it in the error sequence array. The RMSE value is calculated using a sliding window method, with the window width covering twenty consecutive control cycles. Each window outputs an error statistic. The window sliding step size is set to one control cycle to ensure the real-time continuity of error assessment.

[0102] The obstacle avoidance strategy effectiveness index generator comprises a normalization processor and a dynamic threshold adjuster. The normalization processor receives the root mean square error sequence and obtains the error statistical characteristics through a historical data logger. The historical data logger stores the final error values ​​of the last one hundred effective obstacle avoidance processes, maintaining two dynamic variables: the minimum error value and the maximum error value. The normalization formula uses a linear mapping method; when the new error value is less than the historical minimum, the maximum exponent is output; when it is greater than the historical maximum, the minimum exponent is output. The dynamic threshold adjuster automatically adjusts the mapping range according to the complexity of the device's operating environment, expanding the dynamic range of the effective index in complex scenarios.

[0103] The point cloud acquisition strategy adjuster employs a three-level response mechanism corresponding to different effectiveness index ranges. When the index is below 0.6, an enhanced perception mode is triggered: the core high-resolution area diameter expands from the baseline 15 meters to 22.5 meters, the edge area voxel resolution increases from 0.5 meters to 0.3 meters, and the lidar scanning frequency increases from 10 Hz to 20 Hz. When the index is between 0.6 and 0.8, the current acquisition parameters remain unchanged. When the index is above 0.8, an energy efficiency optimization mode is activated: the core area diameter shrinks to 10.5 meters, the edge area voxel resolution decreases to 0.7 meters, and the scanning frequency decreases to 5 Hz.

[0104] The parameter transition controller ensures data continuity during acquisition strategy switching. A proportional-integral-derivative (PID) control algorithm is used to smoothly adjust parameter changes, with the proportional gain set at 0.5, the integral time constant at 0.2 seconds, and the derivative time constant at 0.05 seconds. When parameter adjustment is triggered, the system completes the transition within three control cycles, avoiding decision oscillations caused by sudden changes in perceived data. The adjustment process monitors point cloud quality indicators in real time, automatically pausing parameter reduction when the point cloud density decrease rate exceeds a threshold.

[0105] The feedback data closed loop establishes a connection via the control bus. The output value of the effectiveness index calculation unit is written to the shared memory area in real time, and the point cloud acquisition controller reads this value at a period of fifty milliseconds. Acquisition parameter adjustment commands are sent to the LiDAR driver board via dedicated data packets, and the commands include spatial partitioning parameters, scanning frequency values, and voxel resolution settings. The LiDAR execution status feedback signal is returned to the adjuster, forming a complete closed-loop control system. All data exchange processes follow a time-sensitive network protocol, and the transmission delay is controlled within two hundred microseconds.

[0106] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence, characterized in that, include: The environmental perception layer performs non-uniform sampling of lidar point cloud data based on a retinal fovea-like mechanism to generate dynamic point cloud partitions. The dual-modal risk assessment layer deploys a millimeter-wave radar array based on the dynamic point cloud partitioning, and fuses lidar point cloud data and millimeter-wave radar echo data to generate a static obstacle risk assessment map and a dynamic obstacle risk assessment map. In the dynamic decision-making layer, a spatiotemporal mapping between the static obstacle risk assessment map and the dynamic obstacle risk assessment map is established. A graph neural network is used to fuse dual-modal risk assessment data to generate an obstacle confidence matrix. The trajectory optimization layer transforms the obstacle confidence matrix into obstacle avoidance trajectory control parameters based on a multi-objective evolutionary algorithm, and sends them to the ground mobile unmanned equipment drive unit via a time-sensitive network protocol. The feedback optimization layer monitors the changes in the environmental state after executing the obstacle avoidance trajectory control parameters in real time, calculates the deviation between the environmental state changes and the preset safety threshold, generates an obstacle avoidance strategy effectiveness index, and dynamically adjusts the point cloud acquisition strategy until the obstacle avoidance strategy effectiveness index reaches its optimum.

2. The autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence as described in claim 1, characterized in that, The method for non-uniform sampling of lidar point cloud data based on a foveal-like mechanism includes: The space in front of the ground-mobile unmanned equipment is divided into a high-resolution core area and a low-resolution edge area, and a dynamic point cloud partitioning module is defined based on a preset historical distribution heat map of obstacles. Dense voxel sampling is used in the high-resolution core region, and sparse voxel sampling is used in the low-resolution edge region. The system uses a three-dimensional convolutional neural network to identify abnormal geometric features in point cloud data in real time, and dynamically adjusts the spatial range of the high-resolution core area based on the identification results.

3. The autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence as described in claim 2, characterized in that, The methods for generating static and dynamic obstacle risk assessment maps include: The dynamic point cloud partitions are spatiotemporally aligned to output a standardized point cloud frame sequence. The millimeter-wave radar array is deployed at the corresponding spatial location of the high-resolution core area, transmitting frequency-modulated continuous waves and receiving Doppler echo signals. The phase change of the Doppler echo signal is analyzed by a moving target detection algorithm to generate a dynamic obstacle motion vector field; The standardized point cloud frame sequence is processed by a point cloud feature pyramid network to output a static obstacle risk assessment map with spatial coordinate annotations. By integrating the static obstacle risk assessment map with the dynamic obstacle motion vector field, a dynamic obstacle risk assessment map with risk level labels is generated.

4. The autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence as described in claim 3, characterized in that, The method for generating a static obstacle risk assessment map includes: A multi-scale point cloud feature extraction network is constructed, which includes a bottom-up geometric feature encoding path and a top-down semantic feature decoding path. The curvature features, normal vector distribution features, and surface roughness features of the point cloud data are extracted layer by layer in the geometric feature encoding path. In the semantic feature decoding path, high-level semantic features and low-level geometric features are fused across layers; The obstacle risk probability value of each point cloud cluster is calculated using a spatial probability model, and a static obstacle risk assessment map containing the location, size, and risk level of static obstacles is generated.

5. The autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence according to claim 3, characterized in that, The method for generating a dynamic obstacle risk assessment map with risk level labels includes: Establish the spatial mapping relationship between millimeter-wave radar echo signals and lidar point clouds, and define a dynamic obstacle tracking coordinate system; The trajectory state of a moving target is predicted using the Kalman filter algorithm, and the residual between the predicted trajectory value and the actual observed value is calculated. Based on the residuals, a motion uncertainty model of the dynamic obstacle is constructed, and the dynamic obstacle risk assessment map containing velocity vectors, acceleration, and risk confidence is output.

6. The autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence as described in claim 5, characterized in that, The method for establishing the spatiotemporal mapping between the static obstacle risk assessment map and the dynamic obstacle risk assessment map includes: A global spatial coordinate system is established with the centroid of the ground-mobile unmanned equipment as the origin, and the static obstacle risk assessment map is transformed into the global spatial coordinate system through external calibration parameters. A hardware timestamp synchronization mechanism is used to align the time references of the static obstacle risk assessment map and the dynamic obstacle risk assessment map. A spatial position mapping table for obstacles is established in the global spatial coordinate system, linking the coordinates of static obstacles with the coordinates of the motion trajectory of dynamic obstacles.

7. The autonomous obstacle avoidance control system for ground mobile unmanned equipment optimized using artificial intelligence according to claim 6, characterized in that, The method for fusing dual-modal risk assessment data using graph neural networks includes: Each obstacle region in the static obstacle risk assessment map is defined as a static node, and geometric feature vectors are extracted as static node attributes. Each moving target in the dynamic obstacle risk assessment map is defined as a dynamic node, and the motion feature vector is extracted as the attribute of the dynamic node. Construct a heterogeneous graph structure containing all static and dynamic nodes, and define the spatial proximity relationship between nodes as connecting edges; By aggregating the feature information of adjacent nodes through a multi-head graph attention mechanism, a comprehensive risk score matrix for each obstacle region is output.

8. The autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence according to claim 7, characterized in that, The method for converting the obstacle confidence matrix into obstacle avoidance trajectory control parameters based on a multi-objective evolutionary algorithm includes: Three types of optimization objectives are set: safe passage objective, energy consumption optimization objective, and trajectory smoothing objective. Initialize multiple sets of candidate trajectory control parameters, each set of parameters including steering angle sequence, velocity sequence and acceleration sequence; Evaluate the avoidance effect of each set of candidate trajectory control parameters on the comprehensive risk score matrix, and calculate the achievement score of the three optimization objectives; The Pareto optimal solution set is iteratively selected, and the set of candidate trajectory control parameters with the highest achievement score is output as the obstacle avoidance trajectory control parameters.

9. The autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence as described in claim 8, characterized in that, The method for generating the obstacle avoidance strategy effectiveness index includes: Collect environmental perception data when actually executing obstacle avoidance trajectory control parameters, and construct the minimum distance change curve of obstacles; Calculate the root mean square error between the minimum distance variation curve of the obstacle and the preset safe distance curve; The root mean square error is normalized, and the resulting output is an effectiveness index of the obstacle avoidance strategy, ranging from zero to one.

10. The autonomous obstacle avoidance control system for ground-based mobile unmanned equipment optimized using artificial intelligence according to claim 9, characterized in that, The method for dynamically adjusting the point cloud acquisition strategy includes: Preset a first threshold and a second threshold for the effectiveness index of the obstacle avoidance strategy; When the obstacle avoidance strategy effectiveness index is lower than the first threshold of the obstacle avoidance strategy effectiveness index, the spatial range and sampling frequency of the high-resolution core area are increased. When the obstacle avoidance strategy effectiveness index is between the first threshold and the second threshold of the obstacle avoidance strategy effectiveness index, the current point cloud acquisition strategy is maintained. When the obstacle avoidance strategy effectiveness index is higher than the second threshold of the obstacle avoidance strategy effectiveness index, the spatial range of the high-resolution core area is reduced by a preset ratio.

Images