An agricultural inspection robot autonomous obstacle avoidance system

Abstract

The application belongs to the technical field of autonomous obstacle avoidance, and particularly relates to an agricultural inspection robot autonomous obstacle avoidance system, which comprises a monitoring module, which collects visual area images and point cloud information in real time, generates an obstacle distribution map and a feasible probability, and predicts a probability of a moving direction based on a Bayesian optimization intention path model; meanwhile, an obstacle feasible probability shared by non-target robots in an occluded area is obtained by using a selection sharing unit; an active selection module generates an initial risk difference vector of a current position point by using a risk calculation unit, iteratively filters a low-risk direction by using a discrimination unit, optimizes direction selection by using an updating unit in combination with a risk fluctuation threshold, and finally determines a main direction and generates an expanded map layer based on an autonomous selection unit; an autonomous control module outputs and adjusts an obstacle avoidance instruction according to the expanded map layer and the main direction feasible probability; and the application realizes dynamic risk assessment in an occluded area, and significantly improves obstacle avoidance reliability in a complex farmland environment.

Description

Technical Field

[0001] This invention belongs to the field of autonomous obstacle avoidance technology, and in particular relates to an autonomous obstacle avoidance system for agricultural inspection robots. Background Technology

[0002] Current agricultural inspection robots primarily rely on LiDAR or monocular vision for environmental perception, which has limitations in highly occluded environments such as orchards and greenhouses: dense crop foliage leads to sparse LiDAR point clouds, visual systems are prone to texture misjudgments under varying lighting conditions, and single-sensor solutions have high obstacle miss rates in both continuous and randomly occluded areas. Existing obstacle avoidance systems are based on static environment modeling, making it difficult to integrate shared occlusion area information from multiple robots in real time, resulting in dynamic obstacle response delays. Furthermore, traditional path planning lacks a precise risk probability quantification assessment mechanism, failing to effectively coordinate the conflict between local obstacle avoidance and global path tracking, easily leading to robot jamming or collisions. There is an urgent need to overcome the technical bottlenecks of multimodal collaborative perception and dynamic risk decision-making to address the challenges of perception blind spots and real-time obstacle avoidance in complex agricultural environments. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes an autonomous obstacle avoidance system for agricultural inspection robots, comprising a monitoring module, an active selection module, and an autonomous control module. The monitoring module constructs a map layer of obstacle distribution in the visible area through multimodal data fusion, predicts the probability of travel direction using a Bayesian-optimized intention path generation model, and shares obstacle information in occluded areas. The active selection module uses a Markov algorithm to construct a multi-path risk obstacle avoidance selection matrix, and determines the optimal obstacle avoidance main direction through iterative calculation of risk difference vectors and dynamic threshold filtering. The autonomous control module generates real-time obstacle avoidance commands based on a fuzzy control algorithm, and achieves dynamic path optimization by combining extended map layer information and a direction deviation feedback mechanism. This application solves the perception blind spot problem of traditional methods in scenarios with dense foliage occlusion and dynamic obstacles by using multi-sensor information fusion and collaborative decision-making in occluded areas. It has advantages such as high obstacle avoidance success rate and strong real-time path planning, and can adapt to the inspection needs of complex agricultural scenarios such as orchards and greenhouses.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] An autonomous obstacle avoidance system for agricultural inspection robots includes:

[0006] The monitoring module is used to collect image and point cloud information of each robot's visible target area and occluded areas in the planned path;

[0007] The active selection module constructs a multi-path risk obstacle avoidance selection matrix based on the target robot's visible target area and occluded area image and point cloud information, robot distribution information, and Markov algorithm. The multi-path risk obstacle avoidance selection matrix is ​​constructed by combining the obstacle information of the target robot's visible area and the obstacle information of the occluded areas in different directions of the planned path with Markov algorithm.

[0008] The autonomous control module, based on the initial planned path and the multi-path risk avoidance selection matrix combined with the fuzzy control algorithm, obtains risk avoidance instructions and adjusts the planned path in real time.

[0009] Specifically, the monitoring module includes a data acquisition unit, a travel intention unit, and a selection and sharing unit;

[0010] The acquisition unit is used to acquire image and point cloud information in the visible area of ​​the planned path in real time, and combine it with the preset multimodal fusion model and map layer algorithm to obtain the obstacle distribution map layer of the visible area and mark the feasible probability distribution state of all obstacles on the planned path in the map layer.

[0011] The travel intention unit obtains the travel direction and travel probability corresponding to each position point of the target robot by combining the user's demand intention with the distribution state of target objects in the target area, the distribution state of the robot, and the obstacle distribution function in the visible area with the intention path generation model optimized by Bayesian function.

[0012] The selection and sharing unit obtains information on occlusion obstacles not shared by the target robot and their corresponding feasibility probability distribution status within the occlusion area in the corresponding direction, based on the ratio distribution of the travel probability of each position point of the target robot in the corresponding travel direction.

[0013] Specifically, the active selection module includes a risk calculation unit and a discrimination unit;

[0014] The risk calculation unit obtains the initial risk difference vector of the current position point based on the difference between the feasible probability distribution of the current position point's travel direction and the feasible probability distribution of the corresponding occluded area in each direction of the current position point.

[0015] The discrimination unit, based on the initial risk difference vector of the current location point, retains the initial risk difference corresponding to the occlusion area in the direction where the difference is less than 0, and re-acquires the occlusion direction sharing information in the corresponding direction based on the ratio distribution of the initial risk difference. At the same time, based on the retained secondary sharing information of the occlusion area, it obtains the secondary risk difference vector, and retains the secondary risk difference corresponding to the occlusion area in the direction where the difference is less than 0 in the secondary risk difference vector.

[0016] Specifically, the active selection module also includes an update unit and an autonomous selection unit;

[0017] The updating unit obtains the initial risk fluctuation vector based on the secondary risk difference corresponding to the retained occluded areas, and updates the risk fluctuation vector based on the initial risk fluctuation vector and a preset risk fluctuation threshold until the risk fluctuation vector of the occluded areas corresponding to all retained directions is less than the risk fluctuation threshold.

[0018] The autonomous selection unit selects the direction with the largest risk difference as the main direction of the next position point based on all risk difference vectors that satisfy the risk fluctuation vector retention direction. It then uses the regional shared information of the main direction combined with the map layer of the visible area corresponding to the current position point, and obtains the extended map layer corresponding to the current position of the target robot through map layer algorithm and coordinate transformation.

[0019] Specifically, the autonomous control module includes a control adjustment unit and an obstacle avoidance path update unit;

[0020] The control adjustment unit obtains the obstacle avoidance control command for the current position point based on the extended map layer information corresponding to the current position of the target robot, combined with the feasible probability distribution of the main direction of the next position point, and the fuzzy control algorithm.

[0021] The obstacle avoidance path update unit is used to feed back the obstacle avoidance control command of the current position point and the feasible probability distribution corresponding to the main direction of the next position point to the travel intention unit, adjust the travel direction of the next position point, and at the same time obtain the direction deviation angle and deviation distance of the next position point before and after the adjustment, and feed the direction deviation angle and deviation distance back to the intention path generation model for training until the direction deviation angle and deviation distance of the next position point before and after the adjustment both meet the corresponding preset deviation threshold.

[0022] Specifically, the process of obtaining the initial risk difference vector at the current location point includes:

[0023] Set the current position of the target robot, and ensure that the probability of feasible risks corresponding to the visible area of ​​the current position is greater than the threshold of the probability of abnormal driving of the target robot.

[0024] Based on the image information and lidar information collected at the current position of the target robot, the first image feature space and the first radar feature space are obtained through the parallel image feature extraction layer and radar feature extraction layer within the preset multimodal map layer model;

[0025] Based on the first image feature space and the first radar feature space, a multimodal fusion obstacle feature space is obtained by fusing them through a convolutional attention layer.

[0026] The multimodal fusion obstacle feature space is input in parallel into the map construction layer and the anchor frame layer to obtain the obstacle distribution status and density corresponding to the map background of the current location point, the anchor frames of the visible and occluded areas within the map background, and the sub-regions at the boundary between the visible and occluded areas.

[0027] Specifically, the process of obtaining the initial risk difference vector at the current location point also includes:

[0028] Based on the obstacle distribution and density of the sub-region at the boundary between the visible and occluded regions and the robot's orientation at the current location, the visible and occluded regions are segmented using a grid algorithm combined with a preset initial fan-shaped segmentation angle, resulting in an initial fan-shaped segmentation sub-region sequence corresponding to different fan-shaped segmentation angles at the boundary.

[0029] Based on the obstacle distribution and density within each initial sector segment, and combined with the target robot's abnormal limit control parameters, the first feasible risk probability corresponding to each initial sector segment is obtained using a Markov algorithm.

[0030] Specifically, the process of obtaining the initial risk difference vector at the current location point also includes:

[0031] Based on the occlusion area corresponding to each initial sector segmentation sub-region, the feasible risk probability at the same boundary between the non-target robot and the target robot within the corresponding occlusion area and the feasible probability value of the non-boundary area within the visible area are obtained.

[0032] Based on the first feasible risk probability of the target robot for each initial sector segmented sub-region, the feasible risk probability of the non-target robot at the same boundary within the target robot's occlusion area, and the feasible probability value of the non-target robot within the visible area, the shared probability distribution of the occlusion area corresponding to the initial sector segmented sub-region is obtained through a Bayesian algorithm.

[0033] The initial risk difference vector of the current location point is obtained by subtracting the first feasible risk probability of each initial sector segmented sub-region from the shared probability of the corresponding occluded region.

[0034] Specifically, the process of obtaining the extended map layer corresponding to the current position of the target robot includes:

[0035] Based on the initial risk difference vector of the current location point, obtain the ratio of the initial risk difference of the occluded area corresponding to the negative difference, sort them in descending order, and construct a data sharing factor set;

[0036] Based on the data sharing factor set, the corresponding non-target robot shared data is obtained, and the data shared by each non-target robot is input into the multimodal map layer model to obtain the current position of the extended map layer sequence corresponding to the target robot's visible area and each occluded area.

[0037] Based on the extended map layer sequence corresponding to the current location and the corresponding obstacle distribution information in the enhanced extended map layer corresponding to each occluded area, the average feasible risk probability and the corresponding second feasible risk probability of the target robot reaching all locations in the enhanced extended map layer corresponding to each occluded area from the current location point are obtained through simulation algorithm combined with Hidden Markov Model.

[0038] Specifically, the process of obtaining the extended map layer corresponding to the current position of the target robot also includes:

[0039] The second risk difference vector is obtained by subtracting the first feasible risk probability corresponding to each initial sector segmentation sub-region from the average feasible risk probability within the corresponding occluded region.

[0040] Based on the second risk difference vector, the second risk difference with a difference less than zero is retained, and the first risk fluctuation value is obtained based on the retained second risk difference and the initial risk difference corresponding to the same occlusion area.

[0041] A preset risk fluctuation threshold is set. If the first risk fluctuation value of any current position corresponding to the extended map layer meets the risk fluctuation threshold, then the enhanced extended map layer that meets the condition and corresponds to the smallest first risk fluctuation value and the largest second feasible risk probability is taken as the feasible enhanced extended map layer of the target robot.

[0042] If the first risk fluctuation value of the extended map layer corresponding to any current position does not meet the risk fluctuation threshold, then the second risk difference retained in the second risk difference vector is used as a new data sharing factor set, and the process of obtaining the feasible augmented extended map layer of the target robot is repeated until the feasible augmented extended map layer of the target robot that meets the conditions is obtained.

[0043] If the process of repeating N rounds of data sharing and simulated map construction fails to obtain a feasible augmented map layer for the target robot that meets the conditions, then the average feasible risk probability sequence corresponding to the round with the smallest risk fluctuation value in the N rounds of data sharing is selected, and the average feasible risk probability is fed back to the grid algorithm. The average feasible risk probability is multiplied by the initial sector segmentation angle to obtain the second sector segmentation angle ratio sequence.

[0044] Based on the second sector segmentation angle ratio sequence, the occluded area corresponding to the current position point is re-divided to obtain the second sector segmentation sub-region;

[0045] Based on the second sector segmentation sub-region combined with reinforcement learning algorithm, repeat the above process of obtaining the feasible augmentation and extension map layer of the target robot until a feasible augmentation and extension map layer that meets the conditions is obtained.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] This invention addresses the shortcomings of existing technologies by employing multimodal sensor fusion and multi-machine collaborative mechanisms to effectively solve obstacle avoidance challenges in agricultural high-occlusion scenarios. Specifically, it leverages the complementary heterogeneous data of LiDAR and binocular vision, combined with sector segmentation of boundary areas and shared probability fusion, to achieve joint perception modeling of occluded areas such as gaps in branches and temporary obstacles, eliminating blind spots of single sensors. Based on iterative filtering of risk difference vectors and Hidden Markov Risk Prediction, it dynamically constructs extended map layer sequences, ensuring path decisions consider both local occlusion changes and global risk distribution. Fuzzy control generates obstacle avoidance commands in real time, and directional deviation feedback continuously optimizes the intention path model, forming a closed-loop adaptive control of perception-decision-execution, ensuring trajectory smoothness and target orientation. When extreme occlusion causes initial planning failure, sector re-segmentation and reinforcement learning are used to re-explore feasible areas, significantly improving the system's robustness and fault tolerance in complex environments. Attached Figure Description

[0048] Figure 1 This is a flowchart of an autonomous obstacle avoidance system for an agricultural inspection robot according to Embodiment 1 of the present invention;

[0049] Figure 2 This is a diagram of the multimodal map layer model architecture of Embodiment 1 of the present invention. Detailed Implementation

[0050] Example 1

[0051] Please see Figure 1 The present invention provides an embodiment of an autonomous obstacle avoidance system for an agricultural inspection robot, comprising: a monitoring module, an active selection module, and an autonomous control module;

[0052] The monitoring module is used to collect image and point cloud information of each robot's visible target area and occluded areas in the planned path;

[0053] The active selection module constructs a multi-path risk obstacle avoidance selection matrix based on the target robot's visible target area and occluded area image and point cloud information, robot distribution information, and Markov algorithm. The multi-path risk obstacle avoidance selection matrix is ​​constructed by combining the obstacle information of the target robot's visible area and the obstacle information of the occluded areas in different directions of the planned path with Markov algorithm.

[0054] The autonomous control module, based on the initial planned path and the multi-path risk avoidance selection matrix combined with the fuzzy control algorithm, obtains risk avoidance instructions and adjusts the planned path in real time.

[0055] Furthermore, in this embodiment, the monitoring module includes a data acquisition unit, a travel intention unit, and a selection and sharing unit;

[0056] The acquisition unit is used to acquire image and point cloud information in the visible area of ​​the planned path in real time, and combine it with the preset multimodal fusion model and map layer algorithm to obtain the obstacle distribution map layer of the visible area and mark the feasible probability distribution state of all obstacles on the planned path in the map layer.

[0057] It should be noted that the image information in this embodiment includes RGB three-channel pixel data, object texture features, and edge contour information, which are used to identify the surface attributes and geometry of obstacles; the point cloud information includes three-dimensional spatial coordinates, reflection intensity values, and point density distribution, which are used to construct the precise spatial location and surface physical characteristics of obstacles; the two are fused to generate a rasterized map layer, in which each grid is labeled with the probability of obstacle existence, the probability distribution of passable areas, and the dynamic obstacle movement trend vector, which together constitute the core input of the feasible probability distribution state.

[0058] The travel intention unit obtains the travel direction and travel probability corresponding to each position point of the target robot by combining the user's demand intention with the distribution state of target objects in the target area, the distribution state of the robot, and the obstacle distribution function in the visible area with the intention path generation model optimized by Bayesian function.

[0059] Furthermore, in this embodiment, obtaining the direction of travel and the probability of travel corresponding to each position of the target robot includes:

[0060] First, the user's intent, target object distribution status, robot distribution status, and visible area obstacle distribution function are uniformly mapped to the robot's body coordinate system through a coordinate transformation matrix, generating a multi-dimensional environmental state tensor containing a global path gravitational field, an obstacle joint probability field, and a robot anti-collision repulsion force field. The obstacle probability field needs to be Gaussian smoothed to eliminate sensor noise. It should be noted that in this embodiment, the user's intent includes target coordinates and target task priority; the target object distribution status is constructed using a rasterized probability density function. For example, in this embodiment, the target object distribution status refers to the probability density function of the position and density distribution of fruit trees in an orchard and the number of fruits corresponding to each fruit tree; the robot distribution status includes, but is not limited to, position, direction of travel, and velocity vector.

[0061] Furthermore, in this process, the Lie group SE(3) transform is used to unify the lidar point cloud and the binocular image into the robot coordinate system, and the temporal deviation is compensated by quaternion spherical linear interpolation; in this embodiment, the obstacle distribution function is converted from discrete grid to continuous probability field by bicubic B-spline interpolation to eliminate map stitching gaps; it should be noted that in this embodiment, the non-target robot state is retrieved in real time by distributed KD-Tree index to find the positions of the 5 most recent devices, and the motion vector prediction uses an adaptive filter;

[0062] Second, a Bayesian network is constructed based on the environmental state tensor, with the intensity of user demand and the density of target objects as parent nodes and the direction of travel as child nodes. When calculating the prior probability of direction, the 360° space is discretized into 36 10° sectors. The probability distribution is generated by using the softmax function in combination with the path tracking intensity coefficient λ. The value of λ is dynamically adjusted according to the type of task. For example, for the picking task, λ = 2.0. The discretized prior probability of direction is output.

[0063] Furthermore, in this process, the dependency relationship between parent and child nodes is automatically constructed based on the strength of data association. When the statistical correlation between the distribution of the target object and the user's demand intention exceeds the threshold, connection edges are dynamically added. The prior probability calculation incorporates an obstacle penetration cost mechanism, and the penetration penalty coefficient is optimized online through gradient backpropagation to ensure that path planning takes into account both target orientation and obstacle avoidance requirements.

[0064] Third, the robot distribution state and obstacle distribution function are integrated, and the collision risk with non-target robots in each direction is calculated through the conflict probability model. After superimposing the obstacle probability field, the directional likelihood function is generated. It should be noted that the weight α in this process is estimated in real time by Kalman filtering to ensure the accuracy of dynamic obstacle response. The conflict probability model is preferably a Bayesian function.

[0065] Furthermore, in this embodiment, the uncertainty of the non-target robot's position is quantified as an elliptical probability distribution region, whose major axis is aligned with the velocity vector, and the region area expands linearly with the motion speed; the conflict probability is calculated based on the Mahalanobis distance exponential decay model of the spatial position; the obstacle likelihood fusion adopts a logarithmic scale hybrid algorithm to avoid the underflow problem of low probability value calculation.

[0066] Fourth, the prior probability and likelihood function are updated under the Bayesian framework. Log space calculation is used to avoid floating-point overflow. After pruning low-probability directions, the maximum posterior probability direction and the probability distribution vector of all directions are output. The main direction decision takes into account both target orientation and real-time obstacle avoidance requirements.

[0067] Furthermore, in this embodiment, the direction probability calculation adopts a hierarchical rounding strategy, specifically: first, the entire circumference space is divided into large-angle intervals, and local normalization is performed within each group; then, the probability balance between groups is achieved through S-shaped function weighting; pruning optimization is performed on low-probability directions, and adjacent sectors with excessively small probability differences are merged for calculation, with the probability quality lost during pruning being evenly distributed to the retained directions; furthermore, the initial division of the entire circumference space into large-angle intervals is based on the distribution of obstacles within the visible area, and is initially determined by those skilled in the art.

[0068] Fifth, Markov transition probabilities are introduced to smooth the posterior probability over time. The transition probability decays exponentially with the deviation of the historical direction, and the smoothing factor β is adaptively adjusted according to the robot's acceleration. The final output is a travel probability distribution that satisfies the continuity of motion. It should be noted that in this embodiment, β is gradually reduced at high acceleration to respond quickly to sudden obstacles. The gradual reduction of β is obtained through simulation algorithm.

[0069] Furthermore, in this embodiment, the direction transfer probability is incorporated into the path curvature constraint mechanism. When the rate of change of the robot's motion direction angle exceeds the dynamic limit, an exponential penalty is applied. The smoothing factor is linked with the acceleration in real time. Under high acceleration, the weight of the historical path is reduced to improve the response capability to sudden obstacles. Closed-loop adjustment is achieved through feedback from the inertial measurement unit.

[0070] Furthermore, the path curvature constraint mechanism in this embodiment is as follows: Based on the robot's kinematic model, a mapping relationship between the rate of change of the direction angle and mechanical stress is established; real-time steering curvature is calculated using wheel speed encoders and IMU data; when the rate of change of curvature exceeds the maximum allowable threshold, an exponential penalty function is triggered. This function causes the probability of transitioning to high-risk directions to decrease by an order of magnitude, ensuring that the path conforms to the physical limits of the mechanical structure. The robot kinematic model is preferably a differential-driven Ackerman steering algorithm; it should be noted that the maximum allowable threshold in this embodiment is determined by the tire-ground friction coefficient and the center of gravity height.

[0071] The selection and sharing unit obtains information on occlusion obstacles not shared by the target robot and their corresponding feasibility probability distribution status within the occlusion area in the corresponding direction, based on the ratio distribution of the travel probability of each position point of the target robot in the corresponding travel direction.

[0072] It should be noted that the process of obtaining occlusion obstacle information shared by non-target robots in this embodiment includes:

[0073] First, based on the probability distribution of the target robot's current position in each direction, select several directions with the highest probability values ​​and calculate the ratio of their probability to the total probability of the selected directions; set a ratio threshold to filter low-weight directions and generate a sequence of probability ratios for key directions to ensure that the sharing request focuses on high-risk occlusion areas.

[0074] Second, the selected directions are mapped to physical space sectors, and the corresponding occluded area ID is quickly located using polar coordinate spatial indexing; based on the robot cluster topology, the adjacent robot with the highest spatial overlap with the target direction and whose communication latency meets the real-time requirements is selected as the data source; a structured request protocol carrying the direction sector ID, timestamp, and data timeliness requirements is sent.

[0075] Third, coordinate transformation and time delay compensation are performed on the received shared data; local and shared data are fused through a confidence weighting strategy, wherein the upper limit of the weight of shared data is limited by its confidence threshold to suppress interference from low-quality data; it should be noted that the time delay compensation in this embodiment is a linear interpolation based on speed.

[0076] Fourth, the fused feasible probability is injected into the corresponding sector of the local map to cover the original perception results; for dynamic obstacles in the shared data, the current position is predicted by combining velocity vector backtracking and the probability field is updated; anisotropic diffusion filtering is used to smooth the probability distribution, eliminating random noise while preserving the sharpness of obstacle edges.

[0077] Fifth, dynamically adjust data accuracy based on network round-trip latency. Specifically, enable dual-channel redundant transmission of WiFi and LoRa, and prioritize the processing of the earliest arriving valid data packets at the receiving end; set data failure judgment rules to trigger local sensor rescanning to fill data gaps; it should be noted that the data failure judgment rules in this embodiment are as follows: automatically discard when latency exceeds the limit or coordinate transformation residual is too large.

[0078] Furthermore, in this embodiment, the active selection module includes a risk calculation unit and a discrimination unit;

[0079] The risk calculation unit obtains the initial risk difference vector of the current position point based on the difference between the feasible probability distribution of the current position point's travel direction and the feasible probability distribution of the corresponding occluded area in each direction of the current position point.

[0080] The discrimination unit, based on the initial risk difference vector of the current location point, retains the initial risk difference corresponding to the occlusion area in the direction where the difference is less than 0, and re-acquires the occlusion direction sharing information in the corresponding direction based on the ratio distribution of the initial risk difference. At the same time, based on the retained secondary sharing information of the occlusion area, it obtains the secondary risk difference vector, and retains the secondary risk difference corresponding to the occlusion area in the direction where the difference is less than 0 in the secondary risk difference vector.

[0081] Furthermore, in this embodiment, the active selection module also includes an update unit and an autonomous selection unit;

[0082] The updating unit obtains the initial risk fluctuation vector based on the secondary risk difference corresponding to the retained occluded areas, and updates the risk fluctuation vector based on the initial risk fluctuation vector and a preset risk fluctuation threshold until the risk fluctuation vector of the occluded areas corresponding to all retained directions is less than the risk fluctuation threshold.

[0083] The autonomous selection unit selects the direction with the largest risk difference as the main direction of the next position point based on all risk difference vectors that satisfy the risk fluctuation vector retention direction. It then uses the regional shared information of the main direction combined with the map layer of the visible area corresponding to the current position point, and obtains the extended map layer corresponding to the current position of the target robot through map layer algorithm and coordinate transformation.

[0084] Furthermore, in this embodiment, the autonomous control module includes a control adjustment unit and an obstacle avoidance path update unit;

[0085] The control adjustment unit obtains the obstacle avoidance control command for the current position point based on the extended map layer information corresponding to the current position of the target robot, combined with the feasible probability distribution of the main direction of the next position point, and the fuzzy control algorithm.

[0086] The obstacle avoidance path update unit is used to feed back the obstacle avoidance control command of the current position point and the feasible probability distribution corresponding to the main direction of the next position point to the travel intention unit, adjust the travel direction of the next position point, and at the same time obtain the direction deviation angle and deviation distance of the next position point before and after the adjustment, and feed the direction deviation angle and deviation distance back to the intention path generation model for training until the direction deviation angle and deviation distance of the next position point before and after the adjustment both meet the corresponding preset deviation threshold.

[0087] In summary, the monitoring module in this embodiment integrates RGB images and point cloud data to construct a rasterized probabilistic map layer. It then uses a Bayesian network to map user intent, obstacle distribution, and robot state into a multi-dimensional environmental tensor. Discretized direction probability calculation and Markov time-series smoothing generate a travel intent that balances target orientation and obstacle avoidance requirements. Secondly, the active selection module innovatively establishes a direction probability ratio-driven occlusion area information sharing mechanism. It integrates local and neighboring robot data through confidence-weighted fusion and employs an iterative optimization strategy based on risk difference vectors to dynamically filter high-risk occlusion areas for secondary perception, constructing a multi-path risk obstacle avoidance selection matrix covering both visible and occluded areas. Finally, the autonomous control module inputs the extended map layer and main direction probability distribution into the fuzzy control system to generate real-time obstacle avoidance commands and provide closed-loop feedback to the path. The generative model achieves dynamic and smooth path adjustment through continuous correction of directional deviations. This application addresses the limitations of traditional methods in perceiving blind spots in occluded areas and in multi-machine collaborative obstacle avoidance decision-making, achieving the following technical effects: First, it significantly enhances the robustness of environmental perception through probability field fusion and spatiotemporal compensation mechanisms, effectively addressing sensor noise and dynamic obstacle interference; second, it greatly improves the accuracy of risk assessment in occluded areas by utilizing collaborative sharing of directional probabilities and iterative risk optimization; third, it ensures that path adjustment conforms to kinematic constraints while maintaining target tracking stability through the linkage of fuzzy control and closed-loop feedback; and fourth, it maintains the continuity and safety of robot motion in complex scenarios based on an adaptive smoothing mechanism based on Markov transfer and curvature constraints, ultimately achieving efficient and reliable collaborative operation of multi-machine systems in unknown occluded environments.

[0088] Further, please refer to Figure 2 In this embodiment, the process of obtaining the initial risk difference vector of the current location point includes:

[0089] Set the current position of the target robot, and ensure that the probability of feasible risks corresponding to the visible area of ​​the current position is greater than the threshold of the probability of abnormal driving of the target robot.

[0090] Based on the image information and lidar information collected at the current position of the target robot, a first image feature space and a first radar feature space are obtained through parallel image feature extraction layers and radar feature extraction layers within a preset multimodal map layer model; furthermore, the image feature extraction layer in this embodiment is preferably the YOLOv5-Transformer algorithm; the radar feature extraction layer in this embodiment is preferably PointPillars-DGCNN;

[0091] It is important to note that in this embodiment, the image feature extraction layer uses a YOLOv5 backbone network for target detection and bounding box regression, and a cascaded Transformer encoder is used to perform global self-attention computation. Motion blur is compensated through feature propagation guided by optical flow. The radar feature extraction layer uses PointPillars to convert the point cloud into pseudo-image raster columns, and applies DGCNN to construct a dynamic K-nearest neighbor graph for edge convolution aggregation, combined with adaptive weighted features based on reflection intensity values. Cross-modal processing achieves spatiotemporal synchronous calibration through quaternion spherical interpolation to obtain the first image feature space and the first radar feature space.

[0092] Based on the first image feature space and the first radar feature space, a multimodal fusion obstacle feature space is obtained by fusing them through a convolutional attention layer.

[0093] It should be noted that the multimodal fusion obstacle feature space acquisition process in this embodiment includes:

[0094] First, the radar coordinates are mapped to the image pixel system through the extrinsic parameter calibration matrix, and the resolution is unified by bilinear interpolation to obtain the first image feature space and the first radar feature space. It should be noted that in this embodiment, 1×1 convolution is used to compress the image and radar features to the same number of channels. It should be further explained that in this embodiment, the channel attention branch performs global average pooling, fully connected, and sigmoid function to generate weights for the image features, and generates weights for the radar features through max pooling, fully connected, and sigmoid function.

[0095] Based on the obtained image weights and radar feature weights, the spatial attention branch concatenates the bimodal features and generates a spatial map through a 3×3 convolution, while combining Sigmoid activation to enhance cross-modal consistency regions.

[0096] Second, based on the features extracted by channel attention and spatial attention, multi-receptive field features are extracted through three-level dilated convolution, and then multimodal fusion obstacle feature space is output after 3×3 convolution channel splicing and layer normalization. It should be noted that in this embodiment, the dilation rates of the three-level dilated convolution are 1, 3, and 5.

[0097] The multimodal fusion obstacle feature space is input in parallel into the map construction layer and the anchor box layer to obtain the map background of the current location point, the anchor boxes of the visible and occluded areas within the map background, and the obstacle distribution status and density corresponding to the sub-regions at the boundary between the visible and occluded areas. It should be noted that in this embodiment, the map construction layer preferably uses the TSDF algorithm, and the anchor box layer preferably uses the PointRCNN + graph cut algorithm.

[0098] It should be further explained that, in this embodiment, the map building layer first divides the fused feature space into a fixed-size voxel grid of size 'a', with each voxel storing coordinates and feature vectors; then, it calculates the signed distance 'd' from each voxel to the nearest obstacle surface, and obtains the TSDF value through a truncation function, where the truncation distance in this embodiment is 3 times the voxel size; secondly, it uses Bayesian methods to update the TSDF values ​​of the fused historical frames, assigning low prior probabilities to dynamic obstacle regions to suppress interference; thirdly, it extracts voxels with TSDF values ​​in the range [0.2, 1] as passable areas, and marks voxels with TSDF values ​​of 0 and located at the edge of the sensor's field of view as... The occlusion area is defined, and a three-level voxel pyramid (b, c, e) is constructed for different levels of planning. It should be further noted that in this embodiment, the bottom layer voxel b is used for real-time obstacle avoidance, the middle layer voxel c supports local path planning, and the top layer voxel e serves global navigation. Max pooling downsampling ensures hierarchical consistency. Furthermore, in this embodiment, it should be noted that the voxel size is balanced according to the robot size, the cutoff distance must be greater than the sensor noise amplitude and less than the typical obstacle size, an exponentially decaying weight is applied to the voxel containing the moving object to handle dynamic obstacles, and the occlusion area is only marked within the sensor's maximum range where adjacent voxels have obstacles to prevent mislabeling.

[0099] It should be further explained that when the anchor box layer in this embodiment uses PointRCNN combined with the graph cut algorithm, firstly, 3D candidate boxes are generated by uniformly sampling in a spherical space based on point cloud features to ensure the coverage of occluded areas; secondly, RoI pooling is performed on the point cloud within the candidate boxes to extract local features, and the fine-tuning of the box pose, including position coordinates and travel direction angle, is regressed with the help of a fully connected layer, while obtaining the classification score of the obstacle to complete the refinement and classification; thirdly, in the obstacle seam detection stage, the graph cut algorithm is used to construct a graph structure, with voxel points on the anchor box boundary as nodes, and the edge weights are calculated based on feature differences and position distances, using a combination of exponential and reciprocal functions, combined with the feature sensitivity control parameter σ, to determine the edge weight size; fourthly, energy minimization is performed, and the data term estimates the probability that a voxel belongs to a seam by jointly estimating the RGB gradient and point cloud density, while the smoothing term is used to penalize the label abrupt changes of adjacent voxels; fifthly, density quantification is performed, counting the number of connected components per unit area, such as 1 square meter, and marking areas exceeding a set threshold as high-density areas.

[0100] Based on the obstacle distribution and density of the sub-region at the boundary between the visible and occluded regions and the robot's orientation at the current location, the visible and occluded regions are segmented using a grid algorithm combined with a preset initial fan-shaped segmentation angle, resulting in an initial fan-shaped segmentation sub-region sequence corresponding to different fan-shaped segmentation angles at the boundary.

[0101] In this step, it is necessary to further explain the following: First, based on the robot's current location data and orientation angle, the boundary area between the visible and occluded areas is discretized into grid cells using a rasterized map. Each grid cell stores the obstacle distribution density value and voxel state (visible, occluded). Second, with the robot's position as the center and the orientation as the central axis, radial dividing lines are generated according to a preset initial fan-shaped segmentation angle, dividing the boundary area into multiple fan-shaped sub-regions. Third, grid traversal is performed on each fan-shaped sub-region, and the proportion of obstacle voxels and the distribution of anchor frame boundary intersections within the cell are statistically analyzed. The coordinates of the anchor frame vertices in the boundary area are quickly retrieved using a spatial hash table. Fourth, the sub-region boundaries are dynamically adjusted according to the obstacle density. When the obstacle density within a unit fan-shaped area exceeds a threshold, an adaptive subdivision mechanism is triggered. It should be noted that, while maintaining the initial segmentation angle base, auxiliary dividing lines are inserted to generate an initial fan-shaped segmented sub-region sequence containing the angle range, obstacle density characteristics, and anchor frame intersection coordinates.

[0102] Based on the obstacle distribution and density within each initial sector segmentation region, and combined with the target robot's abnormal limit control parameters, the first feasible risk probability corresponding to each initial sector segmentation region is obtained through a Markov algorithm.

[0103] Based on the occlusion area corresponding to each initial sector segmentation sub-region, the feasible risk probability at the same boundary between the non-target robot and the target robot within the corresponding occlusion area and the feasible probability value of the non-boundary area within the visible area are obtained.

[0104] Based on the first feasible risk probability of the target robot for each initial sector segmented sub-region, the feasible risk probability of the non-target robot at the same boundary within the target robot's occlusion area, and the feasible probability value of the non-target robot within the visible area, the shared probability distribution of the occlusion area corresponding to the initial sector segmented sub-region is obtained through a Bayesian algorithm.

[0105] The initial risk difference vector of the current location point is obtained by subtracting the first feasible risk probability of each initial sector segmented sub-region from the shared probability of the corresponding occluded region.

[0106] It should be further explained that, in this embodiment, based on the robot's current position coordinates and orientation angle, obstacle feature extraction is first performed on each initial sector segmented sub-region: obstacle density, height distribution, and dynamic obstacle motion vectors are statistically analyzed by traversing a voxel grid, and a state transition matrix is ​​constructed in combination with the robot's abnormal limit control parameters. When calculating the state transition probability using a Markov algorithm, a time decay factor is introduced to process the historical trajectory of dynamic obstacles. Secondly, for occluded areas, Bayesian inference is used to learn the behavior patterns of non-target robots at the boundary from historical sensor data, and a convolutional neural network is used to predict the feasible risk probability within the occluded area. For the visible area, the feasible probability value of the non-boundary area is directly calculated based on the LiDAR point cloud density. Then, the first feasible risk probability of the target robot in the sub-region is multiplied element-wise with the risk probability of the corresponding area of ​​the non-target robot, and a shared probability distribution matrix is ​​generated by smoothing with a Gaussian kernel function. Finally, a matrix point difference operation is performed on each sub-region, and the results are arranged in angular order to form a risk difference vector, where a positive difference value indicates a potential collision risk in that direction, and a negative difference value indicates a feasible safe area. The vector dimension is consistent with the number of sector sub-regions. Abnormal limit control parameters include, but are not limited to, maximum permissible acceleration and steering angular velocity threshold.

[0107] Furthermore, the process of obtaining the extended map layer corresponding to the current position of the target robot includes:

[0108] Based on the initial risk difference vector of the current location point, obtain the ratio of the initial risk difference of the occluded area corresponding to the negative difference, sort them in descending order, and construct a data sharing factor set;

[0109] It should be further explained that for the data sharing factor set, this embodiment traverses the initial risk difference vector of the current location point, extracts the elements with negative differences, calculates the ratio of each negative value to the sum of the absolute values ​​of all negative values ​​in the vector, forms a probability distribution sequence, and constructs the data sharing factor set by sorting the ratios from high to low. This process is implemented through vector normalization and quick sorting algorithms to ensure that the factor set reflects the feasibility priority of the occluded area.

[0110] Based on the data sharing factor set, the corresponding non-target robot shared data is obtained, and the data shared by each non-target robot is input into the multimodal map layer model to obtain the current position of the extended map layer sequence corresponding to the target robot's visible area and each occluded area.

[0111] Further explanation is needed regarding the input and processing of multimodal map layer model data. In this embodiment, based on the priority of the data sharing factor set, multimodal information such as LiDAR point clouds, visual images, and inertial measurement unit data for the corresponding region are obtained from the communication cache of the non-target robot. The data is then unified to the coordinate system of the target robot through a spatiotemporal alignment algorithm. This data is input into a multimodal map layer model composed of convolutional neural networks and graph neural networks, and the output is an enhanced extended map layer sequence containing the visible region and each occluded region. Transfer learning is used to optimize the parameters during model training.

[0112] Based on the extended map layer sequence corresponding to the current location and the corresponding obstacle distribution information in the enhanced extended map layer corresponding to each occluded area, the average feasible risk probability and the corresponding second feasible risk probability of the target robot reaching all locations in the enhanced extended map layer corresponding to each occluded area from the current location point are obtained through simulation algorithm combined with Hidden Markov Model.

[0113] Further explanation is needed regarding the calculation of the average feasible risk probability. In this embodiment, for each enhanced extended map layer, a large number of virtual trajectories are generated using the Monte Carlo simulation algorithm, covering all location points within the map layer. For each trajectory, the state transition probability is calculated using a Hidden Markov Model. The state is defined as {feasible, obstacle, unknown}, and the observation value is sensor simulation data. The average feasible risk probability from the current location point to each location point is obtained through forward-backward algorithm iteration. At the same time, the historical motion pattern of dynamic obstacles is considered as a correction factor for the transition probability.

[0114] The second risk difference vector is obtained by subtracting the first feasible risk probability corresponding to each initial sector segmentation sub-region from the average feasible risk probability within the corresponding occluded region.

[0115] Further explanation is needed regarding the generation of the second risk difference vector. In this embodiment, the first feasible risk probability of each initial sector segmented sub-region is calculated element-wise with the average feasible risk probability of the corresponding enhanced extended map layer to generate the second risk difference vector. Elements with a difference less than zero in the vector are retained through matrix indexing operations. These elements correspond to areas where the risk has decreased, providing basic data for subsequent risk fluctuation analysis.

[0116] Based on the second risk difference vector, the second risk difference with a difference less than zero is retained, and the first risk fluctuation value is obtained based on the retained second risk difference and the initial risk difference corresponding to the same occlusion area.

[0117] It should be further explained that for the calculation of the first risk fluctuation value, this embodiment compares the retained second risk difference value with the initial risk difference value of the same occluded area, and calculates the absolute difference between the two as the first risk fluctuation value. This value reflects the change in risk assessment after the introduction of the extended map layer. The calculation uses a sliding window smoothing process to reduce the impact of noise and ensure that the fluctuation value can accurately reflect the risk change trend.

[0118] A preset risk fluctuation threshold is set. If the first risk fluctuation value of any current position corresponding to the extended map layer meets the risk fluctuation threshold, then the enhanced extended map layer that meets the condition and corresponds to the smallest first risk fluctuation value and the largest second feasible risk probability is taken as the feasible enhanced extended map layer of the target robot.

[0119] Further explanation is needed regarding the selection of feasible enhanced extended map layers. In this embodiment, a risk fluctuation threshold is preset. The first risk fluctuation value of all extended map layers is traversed, and the map layer that meets the threshold condition, has the smallest risk fluctuation value, and the highest second feasible risk probability is selected as the target. If none of them are met, the retained second risk difference is used as a new factor set. The data sharing and map generation process is repeated, and the optimization is continuously carried out through a loop iteration mechanism until a map layer that meets the conditions is found or the maximum number of iterations is triggered.

[0120] If the first risk fluctuation value of the extended map layer corresponding to any current position does not meet the risk fluctuation threshold, then the second risk difference retained in the second risk difference vector is used as a new data sharing factor set, and the process of obtaining the feasible augmented extended map layer of the target robot is repeated until the feasible augmented extended map layer of the target robot that meets the conditions is obtained.

[0121] Further explanation is needed regarding the adjustment of the sector segmentation angle and the re-division of regions. In this embodiment, when the risk difference of all occluded regions is removed, the average feasible risk probability sequence of the round with the smallest risk fluctuation value in the historical process is extracted. Each probability value is multiplied by the initial sector segmentation angle to obtain the second sector segmentation angle ratio sequence. The current occluded region is re-divided through the angle interpolation algorithm to generate a second sector segmentation sub-region that better fits the risk distribution, providing a more refined state space for subsequent reinforcement learning.

[0122] If the process of repeating N rounds of data sharing and simulated map construction fails to obtain a feasible augmented map layer for the target robot that meets the conditions, then the average feasible risk probability sequence corresponding to the round with the smallest risk fluctuation value in the N rounds of data sharing is selected, and the average feasible risk probability is fed back to the grid algorithm. The average feasible risk probability is multiplied by the initial sector segmentation angle to obtain the second sector segmentation angle ratio sequence.

[0123] Based on the second sector segmentation angle ratio sequence, the occluded area corresponding to the current position point is re-divided to obtain the second sector segmentation sub-region;

[0124] Based on the second sector segmentation sub-region combined with reinforcement learning algorithm, repeat the above process of obtaining the feasible augmentation and extension map layer of the target robot until a feasible augmentation and extension map layer that meets the conditions is obtained.

[0125] Further explanation is needed regarding the reinforcement learning iterative optimization. In this embodiment, the second sector segmented sub-region is used as the state space. The process of generating a feasible augmented map layer is defined as a reinforcement learning task. The agent's actions are to select a data sharing strategy and a map expansion method. The reward function is designed by combining the risk fluctuation value and the feasibility probability. The policy network parameters are iteratively updated through the PPO algorithm until the generated map layer meets the risk fluctuation threshold requirement, thereby realizing closed-loop adaptive control from data sharing to map optimization.

[0126] In summary, this embodiment constructs an enhanced extended map layer with environmental adaptability through multimodal fusion and dynamic risk assessment techniques, enabling the robot to achieve accurate environmental perception and risk prediction in complex scenarios. Technically, YOLOv5-Transformer and PointPillars-DGCNN are used in parallel to extract image and radar features. Quaternion interpolation is used to calibrate spatiotemporal biases, and after fusion via a convolutional attention mechanism, a multimodal obstacle feature space is generated, effectively improving the feature representation accuracy of dynamic obstacles and occluded areas. A three-level voxel pyramid is constructed using the TSDF algorithm, and anchor boxes are generated and obstacles are detected using PointRCNN and graph cut algorithms. This achieves multi-scale environmental modeling from global path planning to real-time obstacle avoidance at the lower level. The spherical sampling and dynamic obstacle attenuation mechanisms significantly improve the modeling accuracy of occluded areas and moving objects. In the risk assessment phase, the feasibility probability of a sector sub-region is calculated by combining the Markov algorithm with the robot's motion limit parameters. Shared data from non-target robots is introduced to construct an enhanced extended map layer. The Hidden Markov Model is used to simulate trajectory risks, and the segmentation angle is iteratively optimized using the PPO algorithm, forming a closed-loop optimization mechanism of "feature fusion, map construction, risk assessment, and map expansion". This application improves the completeness of environmental representation through multimodal data complementarity. The dynamic calibration mechanism based on Bayesian update and reinforcement learning effectively solves the problems of sensor field of view limitations and risk assessment lag, enabling the robot to adaptively expand the map coverage in unknown environments and greatly reducing the risk assessment error of occluded areas. At the same time, through sector segmentation and risk fluctuation threshold control, a dynamic balance between computing resources and environmental modeling accuracy is achieved, providing a reliable environmental cognition foundation for autonomous navigation in complex scenarios.

[0127] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

Claims

1. An autonomous obstacle avoidance system for agricultural inspection robots, characterized in that, include: The monitoring module is used to collect image and point cloud information of each robot's visible target area and occluded areas in the planned path; The active selection module constructs a multi-path risk obstacle avoidance selection matrix based on the target robot's visible target area and occluded area image and point cloud information, robot distribution information, and Markov algorithm. The multi-path risk obstacle avoidance selection matrix is ​​constructed by combining the obstacle information of the target robot's visible area and the obstacle information of the occluded areas in different directions of the planned path with Markov algorithm. The autonomous control module obtains risk avoidance instructions based on the initial planned path and the multi-path risk avoidance selection matrix, combined with a fuzzy control algorithm, and adjusts the planned path in real time. The active selection module includes a risk calculation unit and a discrimination unit; The risk calculation unit obtains the initial risk difference vector of the current position point based on the difference between the feasible probability distribution of the current position point's travel direction and the feasible probability distribution of the corresponding occluded area in each direction of the current position point. The discrimination unit, based on the initial risk difference vector of the current location point, retains the initial risk difference corresponding to the occlusion area in the direction where the difference is less than 0, and re-acquires the corresponding occlusion direction sharing information based on the ratio distribution of the initial risk difference. At the same time, based on the retained secondary sharing information of the occlusion area, it obtains the secondary risk difference vector, and retains the secondary risk difference corresponding to the occlusion area in the direction where the difference is less than 0 in the secondary risk difference vector.

2. The autonomous obstacle avoidance system for agricultural inspection robots as described in claim 1, characterized in that, The monitoring module includes a data acquisition unit, a travel intention unit, and a selection and sharing unit; The acquisition unit is used to acquire image and point cloud information in the visible area of ​​the planned path in real time, and combine it with the preset multimodal fusion model and map layer algorithm to obtain the obstacle distribution map layer of the visible area and mark the feasible probability distribution state of all obstacles on the planned path in the map layer. The travel intention unit obtains the travel direction and travel probability corresponding to each position point of the target robot based on the user's demand intention, the distribution state of target objects in the target area, the robot's distribution state, and the obstacle distribution function in the visible area combined with the intention path generation model optimized by Bayesian function; the user's demand intention includes target coordinates and target task priority. The selection and sharing unit obtains information on occlusion obstacles not shared by the target robot in the corresponding occlusion area and the corresponding feasible probability distribution state based on the ratio distribution of the travel probability of each position point of the target robot in the corresponding travel direction.

3. The autonomous obstacle avoidance system for agricultural inspection robots as described in claim 2, characterized in that, The active selection module also includes an update unit and an autonomous selection unit; The updating unit obtains the initial risk fluctuation vector based on the secondary risk difference corresponding to the retained occluded areas, and updates the risk fluctuation vector based on the initial risk fluctuation vector and a preset risk fluctuation threshold until the risk fluctuation vector of the occluded areas corresponding to all retained directions is less than the risk fluctuation threshold. The autonomous selection unit selects the direction with the largest risk difference as the main direction of the next position point based on all risk difference vectors that satisfy the risk fluctuation vector retention direction. It then uses the regional shared information of the main direction combined with the map layer of the visible area corresponding to the current position point, and obtains the extended map layer corresponding to the current position of the target robot through map layer algorithm and coordinate transformation.

4. The autonomous obstacle avoidance system for an agricultural inspection robot as described in claim 3, characterized in that, The autonomous control module includes a control adjustment unit and an obstacle avoidance path update unit; The control adjustment unit obtains the obstacle avoidance control command for the current position point based on the extended map layer information corresponding to the current position of the target robot, combined with the feasible probability distribution of the main direction of the next position point, and the fuzzy control algorithm. The obstacle avoidance path update unit is used to feed back the obstacle avoidance control command of the current position point and the feasible probability distribution corresponding to the main direction of the next position point to the travel intention unit, adjust the travel direction of the next position point, and at the same time obtain the direction deviation angle and deviation distance of the next position point before and after the adjustment, and feed the direction deviation angle and deviation distance back to the intention path generation model for training until the direction deviation angle and deviation distance of the next position point before and after the adjustment both meet the corresponding preset deviation threshold.

5. The autonomous obstacle avoidance system for agricultural inspection robots as described in claim 4, characterized in that, The process of obtaining the initial risk difference vector at the current location point includes: Set the current position of the target robot, and ensure that the probability of feasible risks corresponding to the visible area of ​​the current position is greater than the threshold of the probability of abnormal driving of the target robot. Based on the image information and lidar information collected at the current position of the target robot, the first image feature space and the first radar feature space are obtained through the parallel image feature extraction layer and radar feature extraction layer within the preset multimodal map layer model; Based on the first image feature space and the first radar feature space, a multimodal fusion obstacle feature space is obtained by fusing them through a convolutional attention layer. The multimodal fusion obstacle feature space is input in parallel into the map construction layer and the anchor frame layer to obtain the obstacle distribution status and density corresponding to the map background of the current location point, the anchor frames of the visible and occluded areas within the map background, and the sub-regions at the boundary between the visible and occluded areas.

6. The autonomous obstacle avoidance system for an agricultural inspection robot as described in claim 5, characterized in that, The process of obtaining the initial risk difference vector at the current location point also includes: Based on the obstacle distribution and density of the sub-region at the boundary between the visible and occluded regions and the robot's orientation at the current location, the visible and occluded regions are segmented using a grid algorithm combined with a preset initial fan-shaped segmentation angle, resulting in an initial fan-shaped segmentation sub-region sequence corresponding to different fan-shaped segmentation angles at the boundary. Based on the obstacle distribution and density within each initial sector segment, and combined with the target robot's abnormal limit control parameters, the first feasible risk probability corresponding to each initial sector segment is obtained using a Markov algorithm.

7. The autonomous obstacle avoidance system for an agricultural inspection robot as described in claim 6, characterized in that, The process of obtaining the initial risk difference vector at the current location point also includes: Based on the occlusion area corresponding to each initial sector segmentation sub-region, the feasible risk probability at the same boundary between the non-target robot and the target robot within the corresponding occlusion area and the feasible probability value of the non-boundary area within the visible area are obtained. Based on the first feasible risk probability of the target robot for each initial sector segmented sub-region, the feasible risk probability of the non-target robot at the same boundary within the target robot's occlusion area, and the feasible probability value of the non-target robot within the visible area, the shared probability distribution of the occlusion area corresponding to the initial sector segmented sub-region is obtained through a Bayesian algorithm. The initial risk difference vector of the current location point is obtained by subtracting the first feasible risk probability of each initial sector segmented sub-region from the shared probability of the corresponding occluded region.

8. The autonomous obstacle avoidance system for an agricultural inspection robot as described in claim 7, characterized in that, The process of obtaining the extended map layer corresponding to the current position of the target robot includes: Based on the initial risk difference vector of the current location point, obtain the ratio of the initial risk difference of the occluded area corresponding to the negative difference, sort them in descending order, and construct a data sharing factor set; Based on the data sharing factor set, the corresponding non-target robot shared data is obtained, and the data shared by each non-target robot is input into the multimodal map layer model to obtain the current position of the extended map layer sequence corresponding to the target robot's visible area and each occluded area. Based on the extended map layer sequence corresponding to the current location and the corresponding obstacle distribution information in the enhanced extended map layer corresponding to each occluded area, the average feasible risk probability and the corresponding second feasible risk probability of the target robot reaching all locations in the enhanced extended map layer corresponding to each occluded area from the current location point are obtained through simulation algorithm combined with Hidden Markov Model.

9. The autonomous obstacle avoidance system for an agricultural inspection robot as described in claim 8, characterized in that, The process of obtaining the extended map layer corresponding to the current position of the target robot also includes: The second risk difference vector is obtained by subtracting the first feasible risk probability corresponding to each initial sector segmentation sub-region from the average feasible risk probability within the corresponding occluded region. Based on the second risk difference vector, the second risk difference with a difference less than zero is retained, and the first risk fluctuation value is obtained based on the retained second risk difference and the initial risk difference corresponding to the same occlusion area. A preset risk fluctuation threshold is set. If the first risk fluctuation value of any current position corresponding to the extended map layer meets the risk fluctuation threshold, then the enhanced extended map layer that meets the condition and corresponds to the smallest first risk fluctuation value and the largest second feasible risk probability is taken as the feasible enhanced extended map layer of the target robot. If the first risk fluctuation value of the extended map layer corresponding to any current position does not meet the risk fluctuation threshold, then the second risk difference retained in the second risk difference vector is used as a new data sharing factor set, and the process of obtaining the feasible augmented extended map layer of the target robot is repeated until the feasible augmented extended map layer of the target robot that meets the conditions is obtained. If the process of repeating N rounds of data sharing and simulated map construction fails to obtain a feasible augmented map layer for the target robot that meets the conditions, then the average feasible risk probability sequence corresponding to the round with the smallest risk fluctuation value in the N rounds of data sharing is selected, and the average feasible risk probability is fed back to the grid algorithm. The average feasible risk probability is multiplied by the initial sector segmentation angle to obtain the second sector segmentation angle ratio sequence. Based on the second sector segmentation angle ratio sequence, the occluded area corresponding to the current position point is re-divided to obtain the second sector segmentation sub-region; Based on the second sector segmentation sub-region combined with reinforcement learning algorithm, repeat the above process of obtaining the feasible augmentation and extension map layer of the target robot until a feasible augmentation and extension map layer that meets the conditions is obtained.

Images