Autonomous obstacle avoidance system of agricultural inspection robot

Abstract

The invention 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 used for acquiring a visual area image and point cloud information in real time, generating an obstacle distribution map and a feasible probability, and predicting an advancing direction probability based on a Bayesian optimization intention path model; meanwhile, obtaining an obstacle feasible probability shared by a non-target robot in the shielding area by utilizing a selection sharing unit; the active selection module is used for generating an initial risk difference vector of a current position point through a risk calculation unit, iteratively screening a low-risk direction through a judgment unit, optimizing direction selection by utilizing an updating unit in combination with a risk fluctuation threshold value, finally determining a main direction based on an autonomous selection unit and generating an extended map layer; the autonomous control module outputs and adjusts an obstacle avoidance instruction according to the extended map layer and the main direction feasible probability; according to the invention, the dynamic risk assessment of the sheltered area is realized, and the obstacle avoidance reliability in a complex farmland environment is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of autonomous obstacle avoidance, and particularly relates to an autonomous obstacle avoidance system for an agricultural inspection robot. BACKGROUND

[0002] Current agricultural inspection robots mainly rely on laser radar or monocular vision for environment perception, and have limitations in high-shading scenes such as orchards and greenhouses: dense crop branches and leaves cause laser radar point cloud to be sparse, the vision system is prone to texture misjudgment under light changes, and a single sensor solution has a high obstacle miss detection rate in continuous and random shading areas. Existing obstacle avoidance systems are based on static environment modeling and cannot real-time fuse shading area information shared by multiple robots, resulting in dynamic obstacle response delay; at the same time, traditional path planning lacks precise risk probability quantitative evaluation mechanism and cannot effectively coordinate the conflict between local obstacle avoidance and global path tracking, which easily causes robot to be stuck or collide. It is urgent to break through the technical bottleneck of multi-modal collaborative perception and dynamic risk decision-making to cope with the challenges of perception blind area and real-time obstacle avoidance in complex agricultural environments. SUMMARY

[0003] In view of the deficiencies of the prior art, the application provides an autonomous obstacle avoidance system for an agricultural inspection robot, which comprises a monitoring module, an active selection module and an autonomous control module. The monitoring module constructs a visible area obstacle distribution map layer through multi-modal data fusion, predicts the probability of the advancing direction in combination with a Bayesian optimized intention path generation model, and shares the shading area obstacle information; the active selection module constructs a multi-path risk obstacle selection matrix using a Markov algorithm, determines the optimal obstacle avoidance main direction through iterative calculation of a risk difference vector and dynamic threshold screening; and the autonomous control module generates real-time obstacle avoidance instructions based on a fuzzy control algorithm, and realizes path dynamic optimization in combination with extended map layer information and direction deviation feedback mechanism. The application solves the problem of perception blind area in the dense branch and leaf shading and dynamic obstacle scene through multi-sensor information fusion and shading area collaborative decision-making, has the advantages of high obstacle avoidance success rate and strong real-time path planning, and can adapt to the inspection needs of complex agricultural scenes such as orchards and greenhouses.

[0004] To achieve the above object, the application provides the following technical scheme.

[0005] An autonomous obstacle avoidance system for an agricultural inspection robot comprises:

[0006] A monitoring module is used for collecting image and point cloud information in a visible target area and a shading area in a planned path of each robot.

[0007] The active selection module combines the target robot visible target area and the occlusion area image and the point cloud information, the robot distribution information, and a Markov algorithm to construct a multi-path risk obstacle avoidance selection matrix; the multi-path risk obstacle avoidance selection matrix is constructed by combining the target robot visible area obstacle information and the obstacle information of the occlusion area in different directions of the planned path with the Markov algorithm;

[0008] The autonomous control module combines the initial planned path and the multi-path risk obstacle avoidance selection matrix with a fuzzy control algorithm to obtain a risk obstacle avoidance instruction to adjust the planned path in real time.

[0009] Specifically, the monitoring module includes a collection unit, a travel intention unit, and a selection sharing unit.

[0010] The collection unit is configured to collect image and point cloud information in the visible area on the planned path in real time, and combine a preset multi-modal fusion model and a map layer algorithm to obtain a visible area obstacle distribution map layer and mark the feasible probability distribution state of all obstacles on the planned path in the map layer.

[0011] The travel intention unit combines the target area target object distribution state, the robot distribution state, and the visible area obstacle distribution function with a Bayesian function optimized intention path generation model according to the user demand intention to obtain the travel direction and travel probability corresponding to each position point of the target robot.

[0012] The selection sharing unit obtains the shared obstacle information in the occlusion area and the corresponding feasible probability distribution state of non-target robots in the corresponding direction according to the ratio distribution of the travel probability of the travel direction corresponding to each position point of the target robot.

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

[0014] The risk calculation unit obtains an initial risk difference vector of the current position point according to the difference between the feasible probability distribution state of the travel direction of the target robot at the current position point and the feasible probability distribution of the occlusion area corresponding to each direction of the current position point.

[0015] The discrimination unit retains the initial risk difference corresponding to the occlusion area in the corresponding direction whose difference value is less than 0 according to the initial risk difference vector of the current position point, reacquires the shared information of the corresponding direction occlusion according to the ratio distribution of the initial risk difference, and obtains a secondary risk difference vector according to the retained secondary shared information of the occlusion area, and retains the secondary risk difference corresponding to the occlusion area in the corresponding direction whose difference value is less than 0 in the secondary risk difference vector.

[0016] Specifically, the active selection module further includes an updating unit and an autonomous selection unit.

[0017] The updating unit obtains a primary risk fluctuation vector according to the secondary risk difference value corresponding to the retained occlusion area, and updates according to the primary risk fluctuation vector combined with a preset risk fluctuation threshold until the risk fluctuation vector of the occlusion area corresponding to all retained directions is less than the risk fluctuation threshold;

[0018] The autonomous selection unit selects a direction corresponding to a maximum risk difference value as a main direction of the next position point based on all risk difference vectors corresponding to risk fluctuation vector retained directions, and obtains an extended map layer corresponding to the current position of the target robot by using the region sharing information of the main direction combined with the visible region map layer corresponding to the current position point, through a map layer algorithm and coordinate conversion.

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

[0020] The control adjustment unit obtains an obstacle avoidance regulation instruction of the current position point by combining the extended map layer information corresponding to the current position of the target robot with the feasible probability distribution of the main direction of the next position point, and combining a fuzzy control algorithm.

[0021] The obstacle avoidance path updating unit is configured to feed back the obstacle avoidance regulation instruction of the current position point and the feasible probability distribution corresponding to the main direction of the next position point to the marching intention unit, adjust the marching direction of the next position point, obtain the direction deviation angle and deviation distance of the next position point before and after adjustment, and feed back the direction deviation angle and deviation distance to the intention path generation model for training until the direction deviation angle and deviation distance of the next position point before and after adjustment both satisfy the corresponding preset deviation threshold.

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

[0023] A current position point of the target robot is set, and the feasible risk probability of the visible region corresponding to the current position point is greater than the abnormal risk probability threshold of the target robot;

[0024] According to the image information and the laser radar information collected at the current position of the target robot, a first image feature space and a first radar feature space are obtained through the image feature extraction layer and the radar feature extraction layer in the preset multi-modal map layer model in parallel.

[0025] Based on the first image feature space and the first radar feature space, a multi-modal fusion obstacle feature space is obtained through a convolution attention layer.

[0026] The multi-modal fusion obstacle feature space is input into the map construction layer and the anchor box layer in parallel to obtain the obstacle distribution state and the density corresponding to the map background, the anchor box of the visible region and the occlusion region in the map background, and the intersection sub-region of the visible region and the occlusion region.

[0027] Specifically, the obtaining process of the initial risk difference value vector of the current position point further includes:

[0028] Based on the obstacle distribution state and the density of the sub-region corresponding to the intersection of the visible region and the occlusion region and the orientation of the robot at the current position point, the grid algorithm is combined with the preset initial sector segmentation angle to segment the visible region and the occlusion region anchor frame and the intersection sub-region, and an initial sector segmentation sub-region sequence corresponding to different sector segmentation angles at the intersection is obtained.

[0029] Based on the obstacle distribution state and the density in each initial sector segmentation sub-region, the target robot abnormal limit control parameter is combined, and the Markov algorithm is used to obtain the first feasible risk probability corresponding to each initial sector segmentation sub-region.

[0030] Specifically, the obtaining process of the initial risk difference value vector of the current position point further includes:

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

[0032] Based on the first feasible risk probability of the target robot corresponding to each initial sector segmentation sub-region, the feasible risk probability of the non-target robot corresponding to the same intersection in the occlusion region of the target robot, and the feasible probability value of the non-target robot in the visible region, the Bayesian algorithm is used to obtain the shared probability distribution of the corresponding occlusion region of the initial sector segmentation sub-region.

[0033] The first feasible risk probability of each initial sector segmentation sub-region is subtracted from the shared probability of the corresponding occlusion region to obtain the initial risk difference value vector of the current position point.

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

[0035] Based on the initial risk difference value vector of the current position point, the ratio of the initial risk difference value of the occlusion region corresponding to the negative difference value is obtained and arranged in descending order to construct a data sharing amount factor set;

[0036] Based on the data sharing amount factor set, the shared data of the non-target robot is obtained, and the shared data of each non-target robot is input into the multi-modal map layer model to obtain the extended map layer sequence corresponding to the current position of the target robot corresponding to the visible region and each occlusion region corresponding to the enhanced extended map layer construction;

[0037] Based on the sequence of the extended map layers corresponding to the current position and the corresponding obstacle distribution information in the enhanced extended map layer corresponding to each occlusion region, the average feasible risk probability and the corresponding second feasible risk probability of the target robot from the current position to all position points in the enhanced extended map layer corresponding to each occlusion region are obtained through a simulation algorithm combined with a hidden Markov model;

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

[0039] The first feasible risk probability corresponding to each initial sector segmentation sub-region is subtracted from the average feasible risk probability in the corresponding occlusion region to obtain a second risk difference value vector;

[0040] Based on the second risk difference value vector, the second risk difference value smaller than zero is retained, and based on the retained second risk difference value and the initial risk difference value corresponding to the same occlusion region, a first risk fluctuation value is obtained;

[0041] A preset risk fluctuation threshold is set, if the first risk fluctuation value of any extended map layer corresponding to the current position meets the risk fluctuation threshold, the enhanced extended map layer corresponding to the first risk fluctuation value meeting the condition and having the minimum value and the maximum 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 any extended map layer corresponding to the current position does not meet the risk fluctuation threshold, the retained second risk difference value in the second risk difference value vector is taken as a new data sharing amount factor set, and the process of obtaining the feasible enhanced extended map layer of the target robot is repeated until the feasible enhanced extended map layer of the target robot meeting the condition is obtained;

[0043] If the process of repeating N rounds of data sharing and simulation map construction does not obtain the feasible enhanced extended map layer of the target robot meeting the condition, the average feasible risk probability sequence corresponding to the minimum risk fluctuation value in the round corresponding to the minimum 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, and the second sector segmentation angle ratio sequence is obtained by multiplying the average feasible risk probability and the initial sector segmentation angle;

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

[0045] Based on the second sector segmentation sub-region and the reinforcement learning algorithm, the above process of obtaining the feasible enhanced extended map layer of the target robot is repeated until the feasible enhanced extended map layer meeting the condition is obtained.

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

[0047] The present application aims at the deficiencies of the prior art, effectively solves the obstacle avoidance problem in the agricultural high-shading scene through multi-modal sensor fusion and multi-machine cooperation mechanism, especially utilizes the heterogeneous data complementarity of laser radar and binocular vision, combines the sector division and shared probability fusion of the boundary area, realizes the joint perception modeling of the shading area such as branch leaf gap and temporary obstacle, and eliminates the perception blind area of a single sensor; based on the iterative screening and hidden Markov risk prediction of the risk difference vector, an extended map layer sequence is dynamically constructed, the path decision considers the local shading change and global risk distribution; the obstacle avoidance instruction is generated in real time through fuzzy control, and the intended path model is continuously optimized by means of directional deviation feedback, so as to form a closed-loop adaptive regulation and control of perception-decision-execution, guarantee the trajectory smoothness and target directivity; when the initial planning is invalid due to extreme shading, the feasible area is re-explored through sector re-division and reinforcement learning, which significantly improves the robustness and fault tolerance of the system in complex environment. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 An agricultural inspection robot autonomous obstacle avoidance system flowchart of embodiment 1 of the present application;

[0049] Figure 2 A multi-modal map layer model architecture diagram of embodiment 1 of the present application. DETAILED DESCRIPTION

[0050] Embodiment 1

[0051] Please refer to Figure 1 The present application provides an embodiment: an agricultural inspection robot autonomous obstacle avoidance system, comprising: a monitoring module, an active selection module and an autonomous control module.

[0052] The monitoring module is used for collecting image and point cloud information in the visible target area and the shading area of each robot in the planned path;

[0053] The active selection module constructs a multi-path risk obstacle avoidance selection matrix according to the image and point cloud information of the visible target area and the shading area of the target robot, the robot distribution information and the Markov algorithm; the multi-path risk obstacle avoidance selection matrix is obtained by combining the obstacle information of the visible area of the target robot and the obstacle information of the shading area in different directions of the planned path with the Markov algorithm;

[0054] The autonomous control module obtains a risk obstacle avoidance instruction to adjust the planned path in real time according to the initial planned path, the multi-path risk obstacle avoidance selection matrix and the fuzzy control algorithm.

[0055] Further, the monitoring module in the embodiment comprises an acquisition unit, a travel intention unit and a selection sharing unit.

[0056] The acquisition unit is configured to acquire image and point cloud information in a visible area on the planning path in real time, and obtain a visible area obstacle distribution map layer and mark a feasible probability distribution state of all obstacles on the planning path in the map layer by combining a preset multi-modal fusion model and a map layer algorithm.

[0057] It should be noted that the image information in the embodiment includes RGB three-channel pixel data, object texture features and edge contour information, which are used to identify the surface attributes and geometric shapes of the obstacles; the point cloud information includes three-dimensional space coordinates, reflection intensity values and point density distribution, which are used to construct the accurate spatial position and surface physical characteristics of the obstacles; after the fusion of the two, a rasterized map layer is generated, in which each grid is marked with obstacle existence probability, passable area probability distribution and dynamic obstacle motion 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 according to the user demand intention, the target object distribution state, the robot distribution state and the visible area obstacle distribution function combined with the intention path generation model optimized by the Bayesian function.

[0059] Further, the acquisition of the travel direction and travel probability corresponding to each position point of the target robot in the embodiment includes:

[0060] First, the user demand intention, the target object distribution state, the robot distribution state and the visible area obstacle distribution function are uniformly mapped to the robot body coordinate system through a coordinate transformation matrix to generate a multi-dimensional environment state tensor including a global path attractive force field, an obstacle joint probability field and a robot anti-collision repulsive force field, wherein the obstacle probability field needs to be smoothed by Gauss to eliminate sensor noise; it should be noted that the user demand intention in the embodiment includes target coordinates and target task priority; the target object distribution state is constructed by a grid probability density function; for example, the target object distribution state in the embodiment refers to the position and density distribution state of the fruit trees in the orchard and the probability density function of the fruit number corresponding to each fruit tree; the robot distribution state includes but is not limited to position, travel direction and speed vector.

[0061] It should be further explained that in this process, the laser radar point cloud and binocular image are unified to the robot coordinate system by Lie group SE(3) transformation, and the time sequence deviation is compensated by spherical linear interpolation of quaternions; in the embodiment, the obstacle distribution function converts discrete grids into a continuous probability field by bicubic B-spline interpolation to eliminate map stitching gaps; it should be noted that the non-target robot state in the embodiment is retrieved by distributed KD-Tree index to obtain the positions of the nearest 5 devices in real time, and the motion vector prediction uses an adaptive filter.

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

[0063] Further, in the process, the dependency relationship between the parent nodes and the child nodes is automatically constructed based on the data association strength, and a connection edge is dynamically added when the statistical correlation between the target object distribution and the user demand intention exceeds a threshold value; the prior probability calculation integrates the obstacle penetration cost mechanism, and the penetration penalty coefficient is optimized online through gradient back propagation, so as to ensure that the path planning takes into account the target directionality and obstacle avoidance demand;

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

[0065] Further, in the embodiment, the position uncertainty of the non-target robot is quantified as an elliptical probability distribution area, the long axis direction of which is aligned with the velocity vector, and the area of the area is linearly expanded 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 mixing algorithm to avoid the underflow problem in the operation of low probability values;

[0066] Fourth, the prior probability and the likelihood function are updated under the Bayesian framework, the logarithmic space is calculated to avoid floating point overflow, the low probability direction is pruned, and the maximum a posteriori probability direction and the full direction probability distribution vector are output, and the main direction decision takes into account the target directionality and real-time obstacle avoidance demand;

[0067] Further, in the embodiment, the direction probability calculation adopts a hierarchical normalization strategy, specifically: first, the full circle space is divided into large angle intervals, the local normalization in the group is implemented, and then the probability balance between groups is realized through the S-shaped function weighting; the low probability direction is pruned and optimized, the adjacent sector probability difference is too small, the calculation is combined, and the probability mass lost by pruning is evenly distributed to the retained direction; further, the first division of the full circle space into large angle intervals is based on the obstacle distribution state in the visible area, which is initially divided by a person skilled in the art;

[0068] Fifth, the Markov transition probability is introduced to time sequence smooth the posterior probability, the transition probability decays exponentially with the historical directional deviation index, and the smoothing factor β is adaptively adjusted according to the robot acceleration, and finally the motion probability distribution satisfying the motion continuity is output; It should be noted that in this embodiment, β is gradually reduced at high acceleration to quickly respond to sudden obstacles, and the gradual reduction process of β is obtained by simulation algorithm;

[0069] Further, in this embodiment, the direction transition probability incorporates the path curvature constraint mechanism, and when the rate of change of the robot motion direction angle exceeds the kinetic limit, an exponential penalty is applied; The smoothing factor is linked in real time with the acceleration, and in the high acceleration state, the historical path weight is reduced to improve the response ability to sudden obstacles, and closed-loop adjustment is realized through inertial measurement unit feedback.

[0070] Further, in this embodiment, the path curvature constraint mechanism is specifically: based on the robot kinematics model, a mapping relationship between the rate of change of the direction angle and the mechanical stress is established: the real-time steering curvature is calculated through the wheel speed encoder and IMU data, and when the rate of change of the curvature exceeds the maximum allowable threshold, an exponential penalty function is triggered. The function makes the transition probability of high-risk directions decay by orders of magnitude, ensuring that the path meets the physical limit of the mechanical structure. The robot kinematics model is preferably a differential drive 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 height of the center of gravity;

[0071] The selection sharing unit obtains the occlusion obstacle information shared by non-target robots in the corresponding direction within the occlusion area and the corresponding feasible probability distribution state according to the ratio distribution of the motion probability of each position point of the target robot corresponding to the motion direction.

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

[0073] First, based on the direction motion probability distribution of the target robot at the current position, a plurality of directions with the highest probability value are selected, and the ratio of the probability to the total sum of the selected direction is calculated; Set the ratio threshold to filter low-weight directions, generate a key direction probability ratio sequence, and ensure that the sharing request focuses on high-risk occlusion areas;

[0074] Second, map the selected directions to the physical space sector, quickly locate the corresponding occlusion area ID through polar coordinate space indexing; According to the topology relationship of the robot cluster, select the adjacent robot with the highest spatial overlap degree with the target direction and the communication delay meeting the real-time requirement as the data source; Send a structured request protocol carrying the direction sector ID, timestamp and data timeliness requirement;

[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 the shared data is limited by the confidence threshold to suppress low-quality data interference; it should be noted that the time delay compensation in the embodiment is a linear interpolation based on speed;

[0076] Fourth, the feasible probability after fusion 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 the speed vector and the probability field is updated; an anisotropic diffusion filter is used to smooth the probability distribution, which retains the sharpness of the obstacle edge while eliminating random noise;

[0077] Fifth, the data accuracy is dynamically adjusted according to the network round-trip delay, specifically: enable dual-channel redundant transmission of WiFi and LoRa, and the receiving end processes the earliest arriving valid data packet first; set data invalidation judgment rules to trigger local sensor rescan to fill data gaps; it should be noted that the data invalidation judgment rules in the embodiment are specifically: automatically discard when the time delay is out of limit or the coordinate transformation residual is too large;

[0078] Further, the active selection module in the embodiment includes a risk calculation unit and a discrimination unit.

[0079] The risk calculation unit obtains an initial risk difference vector of the current position point according to the difference between the feasible probability distribution state of the current position point of the target robot in the travel direction and the feasible probability distribution of each direction corresponding to the blocked area of the current position point.

[0080] The discrimination unit retains the initial risk difference value corresponding to the blocked area of the direction corresponding to the difference value less than 0 according to the initial risk difference vector of the current position point, and reacquires the shared information of the blocked direction corresponding to the direction according to the ratio distribution of the initial risk difference value, and obtains a secondary risk difference vector according to the retained secondary shared information of the blocked area, and retains the secondary risk difference value corresponding to the blocked area of the direction corresponding to the difference value less than 0 in the secondary risk difference vector.

[0081] Further, the active selection module in the embodiment further includes an update unit and an autonomous selection unit.

[0082] The update unit obtains a primary risk fluctuation vector according to the retained secondary risk difference value corresponding to the blocked area, and updates according to the primary risk fluctuation vector combined with the preset risk fluctuation threshold until the risk fluctuation vector of the blocked area corresponding to all the retained directions is less than the risk fluctuation threshold.

[0083] The autonomous selection unit selects a direction corresponding to the maximum risk difference value as the main direction of the next position point based on all risk difference vectors corresponding to the risk fluctuation vector reservation direction, and obtains an extended map layer corresponding to the current position of the target robot by combining the region sharing information of the main direction with the visible region map layer corresponding to the current position point, through a map layer algorithm and coordinate conversion.

[0084] Further, the autonomous control module in the embodiment includes a control adjustment unit and an obstacle avoidance path updating unit.

[0085] The control adjustment unit obtains an obstacle avoidance regulation instruction for the current position point by combining the extended map layer information corresponding to the current position of the target robot with the feasible probability distribution of the main direction of the next position point, and combining a fuzzy control algorithm.

[0086] The obstacle avoidance path updating unit is configured to feed back the obstacle avoidance regulation instruction for 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, obtain the direction deviation angle and the deviation distance of the next position point before and after the adjustment, and feed back the direction deviation angle and the deviation distance to the intention path generation model for training until the direction deviation angle and the deviation distance of the next position point before and after the adjustment both satisfy the corresponding preset deviation threshold.

[0087] In summary, the monitoring module in the embodiment fuses RGB images and point cloud data to construct a rasterized probability map layer, maps user intent, obstacle distribution and robot state into a multi-dimensional environmental tensor through a Bayesian network, and generates a travel intent that takes into account target orientation and obstacle avoidance requirements through discretized directional probability calculation and Markov temporal smoothing; secondly, the active selection module innovatively establishes a sheltered area information sharing mechanism driven by a directional probability ratio, fuses local and neighborhood robot data through confidence weighting, and uses an iterative optimization strategy of a risk difference vector to dynamically select high-risk sheltered areas for secondary perception, thereby constructing a multi-path risk avoidance selection matrix covering visible and sheltered areas; finally, the autonomous control module inputs the extended map layer and the main directional probability distribution into a fuzzy control system to generate real-time obstacle avoidance instructions and feedback to the path generation model, and realizes dynamic smoothing adjustment of the path through continuous correction of the directional deviation; the application solves the limitations of traditional methods in the perception blind area of the sheltered area and multi-machine cooperative obstacle avoidance decision-making, and achieves the following technical effects: first, the robustness of environmental perception is significantly enhanced through a probability field fusion and time-space compensation mechanism, effectively dealing with sensor noise and dynamic obstacle interference; second, the accuracy of risk assessment in the sheltered area is greatly improved through the cooperative sharing and risk iterative optimization of the directional probability; third, the fuzzy control and closed-loop feedback linkage ensures that the path adjustment meets the kinematic constraints and maintains the stability of target tracking; fourth, the adaptive smoothing mechanism based on Markov transition and curvature constraints maintains the continuity and safety of robot motion in complex scenarios, ultimately achieving efficient and reliable cooperative operation of the multi-machine system in unknown sheltered environments.

[0088] Further, referring to Figure 2 The process of obtaining the initial risk difference vector of the current position point in the embodiment includes:

[0089] The current position point of the target robot is set, and the feasible risk probability corresponding to the visible area of the current position point is greater than the abnormal risk probability threshold of the target robot driving;

[0090] According to the image information and laser radar information collected at the current position of the target robot, a first image feature space and a first radar feature space are obtained through the image feature extraction layer and the radar feature extraction layer in the preset multi-modal map layer model; further, the image feature extraction layer in the embodiment is preferably a YOLOv5-Transformer algorithm; the radar feature extraction layer in the embodiment is preferably a PointPillars-DGCNN;

[0091] It should be noted that in the embodiment, the image feature extraction layer adopts the YOLOv5 backbone network for target detection and bounding box regression, and cascades the Transformer encoder to implement global self-attention calculation, and compensates motion blur through optical flow guided feature propagation; the radar feature extraction layer uses PointPillars to convert the point cloud into a pseudo image grid column, and applies DGCNN to construct a dynamic K-neighbor graph for edge convolution aggregation, and combines the adaptive weighted features of the reflection intensity value. The cross-modal processing realizes the space-time synchronous calibration through the 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, the multi-modal fusion obstacle feature space is obtained by convolution attention layer fusion.

[0093] It should be noted that the multi-modal fusion obstacle feature space acquisition process in the embodiment includes:

[0094] First, the radar coordinates are mapped to the image pixel system through the external parameter calibration matrix, and the resolution is unified through the bilinear interpolation to obtain the first image feature space and the first radar feature space; it should be noted that in the embodiment, 1x1 convolution is used to compress the image and radar features to the same number of channels; it should be further explained that the channel attention branch of the embodiment performs global average pooling, full connection and Sigmoid function on the image features to generate weights, and the radar features pass through max pooling, full connection and Sigmoid function to generate weights;

[0095] Based on the obtained image weight and radar feature weight, the spatial attention branch generates a spatial map through 3x3 convolution after splicing the double-modal features, and simultaneously strengthens the cross-modal consistent region through Sigmoid activation.

[0096] Second, based on the features extracted by the above channel attention and the features extracted by the spatial attention, a multi-receptive field feature is extracted through three-level dilated convolution, and a multi-modal fusion obstacle feature space is output through 3x3 convolution channel splicing and layer normalization. It should be noted that in the embodiment, the dilation rates of the three-level dilated convolution are 1, 3 and 5.

[0097] The multi-modal fusion obstacle feature space is input into the map construction layer and the anchor box layer in parallel to obtain the current position point map background, the visible area and the occluded area anchor box in the map background, and the obstacle distribution state and the density of the sub-region corresponding to the visible area and the occluded area. It should be noted that in the embodiment, the map construction layer is preferably TSDF algorithm, and the anchor box layer is preferably PointRCNN+graph cut algorithm.

[0098] It needs to be further explained that in the embodiment, the map construction layer first divides the fusion feature space into a fixed size voxel grid with size a, each voxel stores coordinates and feature vectors; then the signed distance d of each voxel to the nearest obstacle surface is calculated, and the TSDF value is obtained through the truncation function, wherein the truncation distance in the embodiment is 3 times the voxel size; secondly, the TSDF value of the fusion historical frame is updated by Bayes, and a low prior probability is given to the dynamic obstacle area to suppress interference; thirdly, the voxels with TSDF value in [0.2, 1] are extracted as the passable area, the voxels with TSDF value of 0 and located at the edge of the sensor field of view are marked as the occlusion area, and a three-level voxel pyramid of b, c and e is constructed for different levels of planning; it needs to be further explained that in the embodiment, the bottom layer b voxel is used for real-time obstacle avoidance, the middle layer c voxel supports local path planning, and the top layer e voxel serves global navigation, and the hierarchical consistency is ensured through maximum pooling downsampling; further, it needs to be noted in the embodiment that the voxel size is balanced according to the size of the robot, the truncation distance needs to be greater than the amplitude of the sensor noise and less than the typical obstacle size, the voxels where the moving object is located are processed with exponential decay weight to deal with dynamic obstacles, and the occlusion area is only marked in the area within the maximum range of the sensor and adjacent to the obstacle to prevent false marking.

[0099] It needs to be further explained that when the anchor frame layer in the embodiment adopts PointRCNN combined with the graph cut algorithm, first, 3D candidate boxes are generated by uniformly sampling the spherical space based on the point cloud features, so as to ensure the coverage of the occlusion area; secondly, the RoI pooling operation is performed on the point cloud in the candidate box, so as to extract local features, and the fine adjustment amount of the box pose is regressed through the full connection layer, including the position coordinates and the travel direction angle, and the classification score of the obstacle is obtained at the same time, so as to complete the finishing and classification; thirdly, in the obstacle joint detection link, the graph cut algorithm is used to construct a graph structure, the voxel points on the boundary of the anchor box are taken as nodes, and the edge weight is calculated according to the feature difference and the position distance; through the combination of the exponential function and the reciprocal function, the edge weight size is determined in combination with the feature sensitivity control parameter σ; fourthly, energy minimization processing is performed, the data item estimates the probability that the voxel belongs to the joint through the joint estimation of the RGB gradient and the point cloud density, and the smoothing term is used to punish the label mutation of adjacent voxels; fifthly, dense quantization is performed, the number of connected domains in a unit area such as 1 square meter is counted, and the area exceeding the set threshold is marked as a high-density area.

[0100] Based on the obstacle distribution state and the density of the sub-region corresponding to the intersection of the visible region and the occlusion region and the current position and direction of the robot, the visible region and the occlusion region anchor frame and the intersection sub-region are segmented through the grid algorithm combined with the preset initial sector segmentation angle, and the initial sector segmentation sub-region sequence corresponding to different sector segmentation angles at the intersection is obtained.

[0101] In this step, it needs to be further explained that first, based on the current position point of the robot, the positioning data and the orientation angle, the anchor frame intersection area of the visible area and the occlusion area is discretized into grid cells through the raster map, each grid cell stores the obstacle distribution density value and the voxel state (visible, occlusion); second, taking the robot position as the center and the orientation as the central axis, a radial division line is generated according to the preset initial sector division angle, and the intersection area is divided into multiple sector sub-areas; third, grid traversal is performed on each sector sub-area, and the obstacle voxel proportion and anchor frame boundary intersection distribution in the unit are counted, and the anchor frame vertex coordinates of the intersection area are quickly retrieved through the space hash table; fourth, the sub-area boundary is dynamically adjusted according to the obstacle density, when the obstacle density in the unit sector area exceeds the threshold, the adaptive subdivision mechanism is triggered, it needs to be noted that the auxiliary division line is inserted under the premise of maintaining the initial division angle base, and the initial sector division sub-area sequence containing the angle range, obstacle density characteristics and anchor frame intersection coordinates is generated.

[0102] Based on the obstacle distribution state and density in each initial sector division sub-area, combined with the target robot abnormal limit control parameter, the first feasible risk probability corresponding to each initial sector division sub-area is obtained through the Markov algorithm;

[0103] Based on the corresponding occlusion area of each initial sector division sub-area, the feasible risk probability of the same intersection of non-target robots and target robots in the corresponding occlusion area and the feasible probability value in the non-intersection of the visible area are obtained;

[0104] Based on the first feasible risk probability of each initial sector division sub-area corresponding to the target robot, the feasible risk probability of the same intersection of non-target robots in the target robot occlusion area and the feasible probability value of the non-target robot in the visible area, the shared probability distribution of the initial sector division sub-area corresponding to the occlusion area is obtained through the Bayesian algorithm;

[0105] The first feasible risk probability of each initial sector division sub-area is subtracted from the shared probability of the corresponding occlusion area, and the initial risk difference value vector of the current position point is obtained.

[0106] It needs to be further explained that in this embodiment, based on the current position coordinates and the orientation angle of the robot, obstacle feature extraction is first performed on each initial sector sub-region: the obstacle density, height distribution and dynamic obstacle motion vector are counted by voxel grid traversal, and the state transition matrix is constructed combined with the robot abnormal limit control parameters. When calculating the state transition probability using Markov algorithm, a time decay factor is introduced to process the historical trajectory of dynamic obstacles; secondly, for the occluded area, the behavior pattern of non-target robots at the boundary is learned from the historical sensor data by Bayesian inference, and the feasible risk probability in the occluded area is predicted by convolutional neural network, and the visible area directly calculates the feasible probability value of the non-boundary area based on the laser radar point cloud density; then, the first feasible risk probability of the target robot in the sub-region is multiplied with the risk probability of the corresponding area of the non-target robot point by point, and the shared probability distribution matrix is generated by smoothing processing through Gaussian kernel function; finally, the matrix point difference operation is performed on each sub-region, and the result is arranged in angle order to form the risk difference vector, where the positive difference value represents the potential collision risk in that direction, and the negative difference value represents the feasible safe area, and the vector dimension is consistent with the number of sector sub-regions. The abnormal limit control parameters include but are not limited to the maximum allowed acceleration and the steering angular velocity threshold;

[0107] Further, the acquisition process of 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 position point, the ratio of the initial risk difference value of the occluded area corresponding to the negative difference value is obtained and arranged in descending order, and a data sharing amount factor set is constructed;

[0109] It needs to be further explained that the initial risk difference vector of the current position point is traversed in this embodiment, the elements with negative difference value are extracted, the ratio of each negative value to the sum of the absolute values of all negative values in the vector is calculated, a probability distribution sequence is formed, and the data sharing amount factor set is constructed after sorting from high to low according to the ratio. This process is realized by vector normalization and quicksort algorithm, which ensures that the factor set reflects the priority of the feasibility of the occluded area.

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

[0111] Further explanation is needed for the multi-modal map layer model data input and processing. According to the priority of the data sharing amount factor set, the multi-modal information such as laser radar point cloud, visual image, inertial measurement unit data of the corresponding area is obtained from the communication buffer of the non-target robot, the data is unified to the coordinate system of the target robot through the space-time alignment algorithm, and the multi-modal map layer model composed of the convolutional neural network and the graph neural network is input, and the enhanced expansion map layer sequence containing the visible area and each occluded area is output. During model training, transfer learning is used to optimize parameters.

[0112] Based on the expansion map layer sequence corresponding to the current position and the corresponding obstacle distribution information in the enhanced expansion map layer corresponding to each occluded area, the average feasible risk probability and the corresponding second feasible risk probability of the target robot from the current position point to all position points in each enhanced expansion map layer corresponding to each occluded area are obtained through the simulation algorithm combined with the hidden Markov model;

[0113] Further explanation is needed for the average feasible risk probability calculation. For each enhanced expansion map layer, a large number of virtual trajectories are generated using the Monte Carlo simulation algorithm to cover all position points in the map layer. The state transition probability is calculated for each trajectory combined with the hidden Markov model, the state is defined as {feasible, obstacle, unknown}, and the observation value is the sensor simulation data. The average feasible risk probability from the current position point to each position point is obtained by iterative calculation through the forward-backward algorithm, and the historical motion pattern of the dynamic obstacle is considered as a correction factor of the transition probability.

[0114] The first feasible risk probability corresponding to each initial sector segmentation sub-area is subtracted from the average feasible risk probability in the corresponding occluded area to obtain a second risk difference value vector.

[0115] Further explanation is needed for the second risk difference value vector generation. The first feasible risk probability of each initial sector segmentation sub-area is subtracted from the average feasible risk probability of the corresponding enhanced expansion map layer to generate a second risk difference value vector. The elements with a difference value less than zero in the vector are retained through matrix index operation. These elements correspond to the risk reduction area, which provides basic data for subsequent risk fluctuation analysis.

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

[0117] For further explanation of the first risk fluctuation value calculation, the embodiment calculates the absolute difference between the reserved second risk difference value and the initial risk difference value of the same occlusion area as the first risk fluctuation value, which reflects the change range of the risk assessment after introducing the extended map layer. A sliding window smoothing process is used to reduce noise and ensure that the fluctuation value accurately reflects the risk change trend.

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

[0119] For further explanation of the feasible enhanced extended map layer screening, the embodiment sets a risk fluctuation threshold and iterates all first risk fluctuation values of the extended map layers to select the map layer that meets the threshold condition and has the smallest risk fluctuation value and the largest second feasible risk probability as the target. If none of them meets the condition, the reserved second risk difference value is used as a new factor set, and the data sharing and map generation process is repeated to continuously optimize until a map layer that meets the condition is found or the maximum iteration number is reached.

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

[0121] For further explanation of the fan segmentation angle adjustment and region redivision, when all risk difference values of the occlusion regions are excluded, the average feasible risk probability sequence of the minimum risk fluctuation value round in the historical process is extracted, each probability value is multiplied by the initial fan segmentation angle to obtain a second fan segmentation angle ratio sequence, and the current occlusion region is redivided by an angle interpolation algorithm to generate a second fan segmentation sub-region that is more consistent with the risk distribution, providing a more detailed state space for subsequent reinforcement learning.

[0122] If the process of repeating N rounds of data sharing and simulated map construction does not obtain a target robot feasible enhanced extended map layer that meets the condition, 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 fan segmentation angle to obtain a second fan segmentation angle ratio sequence.

[0123] re-divide the occlusion area corresponding to the current position point based on the second fan-shaped segmentation angle ratio sequence to obtain a second fan-shaped segmentation sub-region;

[0124] Based on the second fan-shaped segmentation sub-region, the above-mentioned process of obtaining the feasible enhanced expansion map layer of the target robot is repeated by combining the reinforcement learning algorithm until a feasible enhanced expansion map layer that meets the conditions is obtained.

[0125] For further explanation of the reinforcement learning iterative optimization, the second fan-shaped segmentation sub-region is taken as the state space, the generation process of the feasible enhanced expansion map layer is defined as a reinforcement learning task, the action of the agent is to select the data sharing strategy and the map expansion mode, the reward function is designed in combination with the risk fluctuation value and the feasibility probability, the policy network parameters are iteratively updated by the PPO algorithm until the generated map layer meets the risk fluctuation threshold requirement, and the closed-loop adaptive control from data sharing to map optimization is realized.

[0126] In summary, the embodiment constructs an enhanced expansion map layer with environmental adaptability by using multi-modal fusion and dynamic risk assessment technology, and realizes precise environmental perception and risk prediction of the robot in a complex scene. Technically, YOLOv5-Transformer and PointPillars-DGCNN are used to extract image and radar features in parallel, the spatiotemporal deviation is calibrated by using quaternion interpolation, and a multi-modal obstacle feature space is generated after convolution attention mechanism fusion, which effectively improves the feature expression accuracy of dynamic obstacles and occlusion areas. A three-level voxel pyramid is constructed by using the TSDF algorithm, and anchor boxes are generated and obstacles are detected by combining PointRCNN and graph cut algorithm, realizing multi-scale environmental modeling from global path planning to real-time obstacle avoidance at the bottom layer, and the spherical sampling and dynamic obstacle decay mechanism significantly improves the modeling accuracy of occlusion areas and moving objects. In the risk assessment link, the Markov algorithm is combined with the robot motion limit parameters to calculate the feasibility probability of the fan-shaped sub-region, the non-target robot shared data is introduced to construct the enhanced expansion map layer, the hidden Markov model is used to simulate the trajectory risk, and the PPO algorithm is used to iteratively optimize the segmentation angle, forming a closed-loop optimization mechanism of "feature fusion, map construction, risk assessment, and map expansion". The present application improves the completeness of environmental representation by complementing multi-modal data, and effectively solves the problems of limited sensor field of view and lagging risk assessment by using the dynamic calibration mechanism based on Bayesian update and reinforcement learning. The robot can adaptively expand the map coverage range in an unknown environment, greatly reducing the occlusion area risk assessment error, and through fan-shaped segmentation and risk fluctuation threshold control, the dynamic balance between computing resources and environmental modeling accuracy is realized, providing a reliable environmental awareness basis for autonomous navigation in complex scenes.

[0127] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative, but not restrictive, and a person of ordinary skill in the art can make changes, modifications, replacements and variations to the above-described embodiments without departing from the purpose of the present application and the scope protected by the claims under the inspiration of the present application, and these are all within the protection of the present application.

Claims

1. An autonomous obstacle avoidance system for an agricultural inspection robot, characterized in that, The application relates to a robot path planning method and device. The method comprises the following steps: a monitoring module is used to collect image and point cloud information in a visible target area and a blocked area in a planned path of each robot; an active selection module is used to combine image and point cloud information of the visible target area and the blocked area of a target robot, robot distribution information and a Markov algorithm to construct a multi-path risk obstacle avoidance selection matrix; the multi-path risk obstacle avoidance selection matrix is constructed by combining obstacle information of the visible area of the target robot and obstacle information of the blocked area in different directions of the planned path and the Markov algorithm; and an autonomous control module is used to combine an initial planned path and the multi-path risk obstacle avoidance selection matrix and a fuzzy control algorithm to obtain a risk obstacle avoidance instruction and to adjust the planned path in real time. The monitoring module comprises a collection unit, a travel intention unit and a selection sharing unit. The collection unit is used to collect image and point cloud information in a visible area of a planned path in real time, to combine a preset multi-modal fusion model and a map layer algorithm, to obtain a visible area obstacle distribution map layer and to mark a feasible probability distribution state of all obstacles on the planned path in the map layer.

2. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 1, wherein, The travel intention unit is used to combine a user demand intention, a target area target object distribution state, a robot distribution state and a visible area obstacle distribution function, to combine a Bayesian function optimized intention path generation model, to obtain a travel direction and a travel probability of each position point of a target robot. The selection sharing unit is used to obtain shared obstacle information and a corresponding feasible probability distribution state of a non-target robot in a blocked area in a corresponding direction according to a ratio distribution of the travel probability of each position point of the target robot in the corresponding travel direction. The active selection module comprises a risk calculation unit and a discrimination unit. The risk calculation unit is used to obtain an initial risk difference vector of a current position point according to a feasible probability distribution state of a travel direction of the target robot and a difference value of a feasible probability distribution of a blocked area corresponding to each direction of the current position point.

3. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 2, wherein, The discrimination unit is used to retain an initial risk difference value of a blocked area corresponding to a direction with a difference value less than 0 according to an initial risk difference vector of the current position point, to reacquire shared information of a corresponding blocked direction according to a ratio distribution of the initial risk difference value, to obtain a secondary risk difference vector according to the retained secondary shared information of the blocked area and to retain a secondary risk difference value of a blocked area corresponding to a direction with a difference value less than 0 in the secondary risk difference vector. The active selection module further comprises an updating unit and an autonomous selection unit. The updating unit is used to obtain a primary risk fluctuation vector according to the retained secondary risk difference value of the blocked area, to update the primary risk fluctuation vector in combination with a preset risk fluctuation threshold until a risk fluctuation vector of a blocked area corresponding to a direction retained is less than the risk fluctuation threshold.

4. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 3, wherein, The autonomous selection unit is used to select a direction corresponding to a maximum risk difference value as a main direction of a next position point based on a risk difference vector of a direction satisfying a risk fluctuation vector retained and to obtain an extended map layer corresponding to a current position of the target robot by using region shared information of the main direction in combination with a visible area map layer corresponding to the current position point, through a map layer algorithm and coordinate conversion. ​ ​ 5. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 4, wherein, The autonomous control module comprises a control adjustment unit and an obstacle avoidance path updating unit; The control adjustment unit obtains an obstacle avoidance regulation instruction for the current position point according to the extended map layer information corresponding to the current position of the target robot in combination with the main direction feasible probability distribution of the next position point and in combination with a fuzzy control algorithm; The obstacle avoidance path updating unit is configured to feed back the obstacle avoidance regulation instruction for 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, obtain the direction deviation angle and the deviation distance of the next position point before and after the adjustment, and feed back the direction deviation angle and the deviation distance to the intention path generation model for training until the direction deviation angle and the deviation distance of the next position point before and after the adjustment both satisfy the corresponding preset deviation threshold.

6. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 5, wherein, The process of obtaining the initial risk difference vector of the current position point comprises: A current position point of the target robot is set, and the feasible risk probability corresponding to the visible area of the current position point is greater than the abnormal travel risk probability threshold of the target robot; According to the image information and the laser radar information collected at the current position of the target robot, a first image feature space and a first radar feature space are obtained through the image feature extraction layer and the radar feature extraction layer in the preset multi-modal map layer model in parallel; Based on the first image feature space and the first radar feature space, a multi-modal fusion obstacle feature space is obtained through a convolution attention layer. The multi-modal fusion obstacle feature space is input into the map construction layer and the anchor box layer in parallel to obtain the obstacle distribution state and the density corresponding to the anchor box of the visible area and the occluded area in the map background, the anchor box of the visible area and the occluded area in the map background, and the intersection sub-area of the visible area and the occluded area.

7. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 6, wherein, The process of obtaining the initial risk difference vector of the current position point further comprises: Based on the obstacle distribution state and the density corresponding to the intersection sub-area of the visible area and the occluded area and the orientation of the robot at the current position point, the visible area and the occluded area anchor box and the intersection sub-area are segmented through a grid algorithm in combination with a preset initial sector segmentation angle to obtain an initial sector segmentation sub-area sequence corresponding to different sector segmentation angles at the intersection. Based on the obstacle distribution state and the density in each initial sector segmentation sub-area, the first feasible risk probability corresponding to each initial sector segmentation sub-area is obtained through a Markov algorithm in combination with the abnormal limit control parameter of the target robot.

8. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 7, wherein, The process of obtaining the initial risk difference vector of the current position point further comprises: Based on the occluded area corresponding to each initial sector segmentation sub-area, the feasible risk probability at the same intersection of the non-target robot and the target robot in the corresponding occluded area and the feasible probability value of the non-intersection in the visible area are obtained; Based on the first feasible risk probability of the target robot corresponding to each initial sector segmentation sub-area, the feasible risk probability of the non-target robot corresponding to the same intersection in the occluded area of the target robot, and the feasible probability value of the non-target robot in the visible area, the shared probability distribution of the occluded area corresponding to the initial sector segmentation sub-area is obtained through a Bayesian algorithm. Differences between the first feasible risk probability of each initial sectorized sub-region and the shared probability of the corresponding occluded region are obtained to obtain an initial risk difference value vector of the current position point.

9. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 8, wherein, The obtaining process of the extended map layer corresponding to the current position of the target robot comprises: Based on the initial risk difference value vector of the current position point, the ratio of the initial risk difference value of the corresponding occluded region with a negative difference value is obtained and arranged in descending order to construct a data sharing amount factor set; Based on the data sharing amount factor set, the shared data of the corresponding non-target robot is obtained, and the data shared by each non-target robot is input into the multi-modal map layer model to obtain an extended map layer sequence corresponding to the current position of the target robot constructed by the corresponding enhanced extended map layer of each occluded region and the visible region of the target robot; Based on the extended map layer sequence corresponding to the current position and the corresponding obstacle distribution information in the enhanced extended map layer of each occluded region, the average feasible risk probability and the corresponding second feasible risk probability of the target robot from the current position point to all position points in the enhanced extended map layer of each occluded region are obtained by a simulation algorithm combined with a hidden Markov model.

10. The autonomous obstacle avoidance system for an agricultural inspection robot of claim 9, wherein, The obtaining process of the extended map layer corresponding to the current position of the target robot further comprises: Differences between the first feasible risk probability of each initial sectorized sub-region and the shared probability of the corresponding occluded region are obtained to obtain an initial risk difference value vector of the current position point. Based on the second risk difference value vector, the second risk difference value with a difference value less than zero is retained, and the first risk fluctuation value is obtained based on the retained second risk difference value and the initial risk difference value of the corresponding same occluded region. A preset risk fluctuation threshold is set, if the first risk fluctuation value of any extended map layer corresponding to the current position satisfies the risk fluctuation threshold, the enhanced extended map layer corresponding to the minimum first risk fluctuation value and the maximum second feasible risk probability that satisfy the condition is taken as the feasible enhanced extended map layer of the target robot. If the first risk fluctuation value of any extended map layer corresponding to the current position does not satisfy the risk fluctuation threshold, the retained second risk difference value in the second risk difference value vector is taken as a new data sharing amount factor set, and the process of obtaining the feasible enhanced extended map layer of the target robot is repeated until the feasible enhanced extended map layer of the target robot that satisfies the condition is obtained. If the process of repeated data sharing and simulation map construction for N rounds does not obtain the feasible enhanced extended map layer of the target robot that satisfies the condition, the average feasible risk probability sequence corresponding to the minimum risk fluctuation value in one round of the N rounds of data sharing process is selected, and the average feasible risk probability is fed back to the grid algorithm, and the second sectorized angle ratio sequence is obtained by multiplying the average feasible risk probability and the initial sectorized angle. Based on the second sectorized angle ratio sequence, the corresponding occluded region of the current position point is re-divided to obtain a second sectorized sub-region. Based on the second sectorized sub-region and the reinforcement learning algorithm, the above process of obtaining the feasible enhanced extended map layer of the target robot is repeated until the feasible enhanced extended map layer that satisfies the condition is obtained.

Images