Cross-view angle unmanned agricultural machine safe operation method and system

Abstract

The application discloses a kind of cross visual angle unmanned agricultural machine safe operation method and system, the method acquires multi-view image, locates inertial navigation data and agricultural machine operation state data, completes time synchronization, coordinate unification and generates predicted travel corridor;Scene structure analysis is carried out using basic vision model, and cross visual angle association and time sequence tracking are carried out on candidate instance, to form known target detection evidence and unknown target discovery evidence;According to the visual angle reliability factor, the semantic, geometric, motion and unknown target discovery evidence are discounted and fused, to obtain joint evidence, uncertainty value and unknown target score, and a spatiotemporal probability occupancy risk field is constructed accordingly, and the action feasible region is determined under the control barrier constraint, and the output is kept, speed limit, detour or brake instruction.The application can improve the stability of unknown obstacle identification and safe action decision in complex field environment.

Description

Technical Field

[0001] This invention relates to a method and system for safe operation of unmanned agricultural machinery from multiple perspectives, belonging to the fields of intelligent agricultural equipment, machine vision, and safety control technology for unmanned agricultural machinery. Background Technology

[0002] With the development of smart agriculture and unmanned operation technologies, unmanned agricultural machinery is gradually being applied to scenarios such as tillage, sowing, plant protection, harvesting, and field transportation. During operation, unmanned agricultural machinery needs to continuously observe the surrounding environment, identify targets such as people, vehicles, crop boundaries, ditches, rocks, farm implements, and temporary obstacles, and control the machinery to maintain movement, decelerate, detour, brake, or prompt manual intervention based on the identification results.

[0003] Field environments differ from road environments. Work areas often present challenges such as strong sunlight, backlighting, dust, mud, crop obstruction, varying ground textures, and temporary debris. Some obstacles are not fixed in shape and their locations are random, making it difficult to reliably determine the presence of safety risks ahead using only a single camera or a fixed-category recognition model.

[0004] Meanwhile, the safety control of unmanned agricultural machinery depends not only on whether obstacles are detected, but also on the current speed, braking distance, steering ability, implement width, positioning error, and operating mode. If simple judgments are made based solely on target detection confidence or obstacle distance, problems may easily arise such as failing to stop in time when necessary, or stopping frequently when not required.

[0005] Existing safe operation methods for unmanned agricultural machinery mainly include the following categories:

[0006] The first category is safety identification methods based on fixed-category object detection. These methods train object detection models to identify known categories of objects such as people, vehicles, obstacles, and crop boundaries, and determine whether to slow down or stop based on the object's location and detection confidence.

[0007] The second category is safety assessment methods based on semantic segmentation or passable area segmentation. These methods extract field roads, crop areas, obstacle areas, and restricted areas using image segmentation models, and generate control commands based on the relationship between the current location of the agricultural machinery and the passable area.

[0008] The third category is obstacle avoidance methods based on multi-sensor ranging or local map construction. These methods typically use lidar, millimeter-wave radar, depth cameras, or visual depth estimation to obtain obstacle distances and perform obstacle avoidance control based on local grid maps or cost maps.

[0009] The fourth category is safety control methods based on rule-based thresholds. These methods set fixed thresholds based on obstacle distance, detection confidence, vehicle speed, heading deviation, or operating mode. When a risk indicator exceeds the threshold, it triggers a stop, speed limit, or manual intervention.

[0010] Existing methods for safe operation of unmanned agricultural machinery largely rely on fixed-category target detection, single-view image judgment, or simple threshold judgment. In situations with strong light, dust, occlusion, crop morphology changes, and numerous temporary foreign objects in the field, perception results are prone to fluctuation. When unknown targets or uncommon obstacles outside the training categories enter the operating area, existing models struggle to determine in a timely and accurate manner whether they will affect safe operation. Furthermore, existing methods typically output stop and speed limit commands directly based on detection confidence, obstacle distance, or fixed thresholds, failing to adequately consider the reliability of perception results from different perspectives, conflicts between perception results, and whether the current speed and braking distance meet safety requirements. This can easily lead to problems such as failure to stop when necessary, frequent stops when not needed, or uneven transitions between speed limits, detours, and braking.

[0011] Therefore, there is a need for a method that can comprehensively utilize multi-view images, agricultural machinery status, and safety control conditions to improve the operational safety of unmanned agricultural machinery in complex field environments. Summary of the Invention

[0012] Objective: To overcome the following shortcomings in existing technologies: First, to address the problem of unstable safety perception for unmanned agricultural machinery in complex field environments. Specifically, this means improving the stability of identifying personnel, vehicles, crop boundaries, ditches, implements, and uncommon obstacles through cross-verification and time-series tracking of multi-camera images under conditions of strong light, dust, obstruction, crop morphology changes, and numerous temporary foreign objects. Second, to address the problem of insufficient basis for safety control judgments of unmanned agricultural machinery. Specifically, this means using multi-view perception results, view reliability, unknown target scores, risk locations, and conditions such as the current speed, braking distance, and implement outline of the agricultural machinery together for safety action judgments, enabling the agricultural machinery to output hold, speed limit, detour, or braking commands more stably. This invention provides a cross-view unmanned agricultural machinery safety operation method and system to improve the safety perception and control capabilities of unmanned agricultural machinery in complex field environments.

[0013] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0014] Firstly, a method for safe operation of unmanned agricultural machinery from multiple perspectives specifically includes:

[0015] Step 1: Acquire multi-camera images from different perspectives, positioning and inertial navigation data, and agricultural machinery operation status data to obtain time-synchronized multi-source operation data and predicted driving corridors. Based on camera calibration parameters and positioning and inertial navigation data, establish coordinate transformation relationships from each camera coordinate system to the agricultural machinery body coordinate system.

[0016] Step 2: Input multi-camera images from different perspectives in the multi-source operation data into the basic vision model to obtain candidate instances from different perspectives. Based on the candidate instances and the predicted driving corridor, generate candidates for passable areas and candidates for restricted areas.

[0017] Step 3: Project candidate instances from different perspectives to a unified coordinate system based on the coordinate transformation relationship. Perform cross-perspective association based on the spatial position, appearance features, scale changes and motion direction of the candidate instances in the unified coordinate system. Combine the images at adjacent time points to perform temporal tracking and generate evidence for the detection of known targets and evidence for the discovery of unknown targets.

[0018] Step 4: Based on the viewpoint reliability factor, perform reliability discounting and fusion on the evidence of known target detection and the evidence of unknown target discovery to obtain joint evidence, uncertainty value and unknown target score.

[0019] Step 5: Construct a spatiotemporal probability occupancy risk field based on joint evidence, uncertainty values, unknown target scores, passable area candidates, prohibited area candidates, predicted driving corridors, and agricultural machinery movement constraints.

[0020] Step 6: Based on the spatiotemporal probability of occupying the risk field, predict the reachable trajectory of the agricultural machinery after executing different candidate actions, and calculate the minimum distance between each reachable trajectory and the high-risk area. Based on the minimum distance, use control barrier constraints to eliminate candidate actions that cannot meet the dynamic safety distance requirements, obtain the safe action feasible region, and output hold, speed limit, or detour control commands based on the safe action feasible region.

[0021] Optionally, it also includes: Step 7: When the feasible domain for safe actions is not obtained, output braking control and manual takeover instructions, and perform safety event caching and parameter calibration.

[0022] Optionally, the method for obtaining the time-synchronized multi-source operation data and predicting the driving corridor specifically includes:

[0023] The clock deviation is calculated based on the camera timestamp and the inertial navigation timestamp. Then, according to the agricultural machinery control cycle, the multi-camera images, positioning inertial navigation data and agricultural machinery operation status data are aligned to obtain multi-source operation data with a unified time reference.

[0024] Based on the agricultural machinery speed and steering status in multi-source operation data, combined with the outline of the agricultural machinery implements, maximum steering capacity and braking capacity, the reachable trajectory of the agricultural machinery within a preset time range is predicted, and the reachable trajectory is expanded by the outline of the vehicle body and agricultural machinery implements to generate a predicted driving corridor.

[0025] Optionally, step 2 specifically includes:

[0026] Drive the underlying visual model using point hints, box hints, text hints, or combinations of point hints, box hints, and text hints to generate candidate masks that include target category probabilities and region contours.

[0027] Candidate instances are selected based on the class relevance, boundary stability, and cross-frame preservation rate of the candidate masks, resulting in a list of candidate instances including target class, segmentation contour, location, and recognition confidence.

[0028] The ground area that is connected to the current location of the unmanned agricultural machinery and whose width meets the passage requirements of the agricultural machinery and implements is identified as a candidate passable area.

[0029] Regions that protrude from the ground, occupy passage space, or are continuous but whose category cannot be determined are identified as obstacle candidate regions.

[0030] The crop area, ditch area, preset no-entry area, and the area formed by safely expanding the obstacle candidate area according to the width of the agricultural machinery body, the outline of the agricultural machinery implements, and the positioning error in the candidate instances are identified as no-entry area candidates.

[0031] Optionally, step 3 specifically includes:

[0032] Based on the coordinate transformation relationship, candidate instances from different perspectives are mapped to the agricultural machinery body coordinate system to obtain the position, occupied contour and spatial scale of each candidate instance in the agricultural machinery body coordinate system.

[0033] Based on the position of candidate instances in the agricultural machinery body coordinate system, the similarity of appearance features between the occupied contour and spatial scale, the overlap of bird's-eye view projection, parallax consistency, adjacent frame displacement, scale change and optical flow residual, it is determined whether candidate instances from different viewpoints or different times correspond to the same target, and the relationship between candidate instances is established.

[0034] Based on the aforementioned correlation, observations of the same target from different perspectives and at different times are merged to form a target trajectory that includes target category, location, occupied contour, movement speed, movement direction, and trajectory continuity. For targets with clear categories, consistent cross-perspective correlation, and continuous trajectories, known target detection evidence is generated. For targets whose categories cannot be determined, whose identification results from different perspectives conflict, but which appear continuously or overlap with the predicted driving corridor, unknown target discovery evidence is generated.

[0035] Optionally, step 4 specifically includes:

[0036] Based on the category confidence, spatial features, motion features, and unknown target discovery evidence of known target detection evidence, a basic trust assignment is constructed for known target detection evidence and unknown target discovery evidence, which serves as the initial trust level of known target detection evidence and unknown target discovery evidence for known targets, unknown targets, and undetermined sets.

[0037] Calculate the reliability factors for evidence of known target detection and evidence of unknown target discovery.

[0038] When the reliability factor is lower than a preset threshold, the basic trust allocation of known target detection evidence and unknown target discovery evidence is discounted based on the reliability factor to obtain the discounted trust level of known targets, unknown targets and undetermined set.

[0039] By fusing evidence of known target detection and evidence of unknown target discovery, the amount of evidence conflict is determined when different pieces of evidence support mutually exclusive target hypotheses.

[0040] When the amount of evidence conflict reaches a preset conflict threshold, the amount of evidence conflict is allocated to the unknown target hypothesis and the undetermined set according to the degree of overlap between the candidate instance and the predicted driving corridor.

[0041] The degree of trust corresponding to each objective hypothesis after combination is used as the joint evidence.

[0042] The trust level corresponding to the undetermined set is used as the uncertainty value, and the score of the unknown target is determined based on the trust level corresponding to the unknown target hypothesis.

[0043] Optionally, step 5 specifically includes:

[0044] The area surrounding the unmanned agricultural machinery is divided into a spatial grid, and the preset prediction time domain is divided into multiple prediction times.

[0045] Based on the target's position, velocity, direction of motion, and trajectory continuity, predict the probability that the target will occupy each spatial grid at different prediction times.

[0046] The confidence level of the target's existence and potential collision consequences are determined based on joint evidence, and the predicted occupancy range of the corresponding target is expanded based on the uncertainty level or the score of the unknown target.

[0047] Based on the spatial relationship between the predicted occupied area of ​​the target and the predicted driving corridor, passable area candidates, and restricted area candidates, the spatial risk value of each spatial grid at different prediction times is determined.

[0048] Based on the spatial risk value, a spatiotemporal probability occupancy risk field is generated, representing the target occupancy probability, collision probability, and potential collision consequences.

[0049] Optionally, step 6 specifically includes:

[0050] Based on the spatiotemporal probability of occupying the risk field, the reachable trajectory of the agricultural machinery after performing different candidate actions is predicted. Combined with the current speed, braking distance, steering ability and the working envelope of the agricultural machinery, the minimum distance between each reachable trajectory and the high-risk area is calculated.

[0051] Based on the minimum distance, candidate actions that cannot meet the dynamic safety distance requirements are eliminated using control barrier constraints, thus obtaining the feasible domain of safe actions;

[0052] When the feasible domain of safe actions is not empty, select the optimal safe action from it and output the maintain, speed limit or detour control command.

[0053] Among them, the feasible domain of the safety action The expression is as follows:

[0054]

[0055] In the formula, For the set of candidate actions, This is the current status of the agricultural machinery. For candidate control actions, To control barrier constraints.

[0056] The expression for the control barrier constraint is as follows:

[0057]

[0058] In the formula, For the minimum distance, To match vehicle speed Machinery outline and positioning error Relevant security thresholds.

[0059] The optimal safety action The expression is as follows:

[0060]

[0061] in, To predict the time domain, To perform the action Later at the predicted time The corresponding predicted driving corridor, Indicates the probability of occupancy. To understand uncertainty, For unknown scores, Indicates the probability of entering a restricted area. Indicates a reference action. , , , , For weight parameters, The target coordinates are given.

[0062] Optionally, step 7 specifically includes:

[0063] When manual takeover, high-risk braking, high uncertainty, or continuous false stops occur, the system caches multi-camera images, positioning and inertial navigation data, agricultural machinery operation status data, joint evidence, spatiotemporal probability occupancy risk field, and control results for the corresponding time period.

[0064] The cached security events are replayed for verification, and the view reliability weight, unknown target score threshold, risk field weight, and dynamic security distance parameter are updated based on the verification results to obtain the updated parameter package.

[0065] The updated parameter package is validated and distributed. If the validation is successful, the current parameter version is used; if the validation fails, it is rolled back to the previous parameter version.

[0066] Secondly, a cross-view unmanned agricultural machinery safety operation system includes: a multi-camera sensing unit, a status acquisition unit, a vehicle-side computing unit, an agricultural machinery controller, and a cloud platform.

[0067] The multi-camera sensing unit is used to collect images of the front, sides and rear of the unmanned agricultural machinery.

[0068] The status acquisition unit is used to acquire GNSS-IMU, wheel speed, steering status, equipment status, and operating mode.

[0069] The vehicle-side computing unit is used for spatiotemporal alignment, structural analysis, evidence fusion, risk field construction, and safety constraint decision-making of multi-source data.

[0070] The agricultural machinery controller is used to execute functions such as holding, speed limiting, detouring, braking, or manual takeover prompts based on the control commands output by the vehicle-side computing unit.

[0071] The cloud platform is used for security incident playback, parameter calibration, and version management.

[0072] Thirdly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a cross-view unmanned agricultural machinery safe operation method as described in any of the first aspects.

[0073] Fourthly, a computer device comprising:

[0074] Memory is used to store instructions.

[0075] A processor is configured to execute the instructions, causing the computer device to perform the operation of a cross-view unmanned agricultural machinery safety operation method as described in any of the first aspects.

[0076] Beneficial Effects: This invention provides a cross-view unmanned agricultural machinery safety operation method and system, involving intelligent agricultural equipment, machine vision, and unmanned agricultural machinery safety control technology. The method mainly includes: acquiring multi-channel camera images, positioning inertial navigation data, and agricultural machinery operation status data from different perspectives of the unmanned agricultural machinery; performing time synchronization, coordinate unification, and predictive driving corridor generation on the data; performing scene structure analysis on the camera images based on a basic vision model to obtain candidate instances, passable area candidates, prohibited area candidates, and obstacle candidate areas; performing cross-view correlation and temporal tracking on the candidate instances to generate known target detection evidence and unknown target discovery evidence, respectively; discounting and fusing semantic evidence, geometric evidence, and motion evidence according to the view reliability factor to obtain joint evidence, uncertainty value, and unknown target score; constructing a spatiotemporal probability occupancy risk field based on the joint evidence, uncertainty value, unknown target score, predicted driving corridor, and agricultural machinery motion constraints; determining the action feasible domain according to the risk field and control barrier constraints, and outputting at least one control command among maintaining, speed limiting, detouring, or braking. This method can improve the ability of unmanned agricultural machinery to identify unknown and uncommon obstacles in complex field environments, and improve the stability of safety control actions. Attached Figure Description

[0077] Figure 1 This is a framework diagram of a cross-view unmanned agricultural machinery safety operation system according to an embodiment of this application.

[0078] Figure 2 This is a flowchart of a cross-view unmanned agricultural machinery safe operation method according to an embodiment of this application.

[0079] Figure 3 This is a diagram of the cross-perspective evidence fusion and security constraint decision-making mechanism in an embodiment of this application. Detailed Implementation

[0080] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0081] The present invention will be further described below with reference to specific embodiments.

[0082] Example 1:

[0083] This embodiment provides a cross-view unmanned agricultural machinery safety operation system, such as Figure 1 As shown, it includes: a multi-camera sensing unit, a status acquisition unit, a vehicle-side computing unit, an agricultural machinery controller, and a cloud platform.

[0084] The multi-camera sensing unit is used to collect images of the front, sides and rear of the unmanned agricultural machinery.

[0085] The status acquisition unit is used to acquire GNSS-IMU (Global Navigation Satellite-Inertial Measurement Unit Combined Navigation System), wheel speed, steering status, equipment status, and operating mode.

[0086] The vehicle-side computing unit is used for spatiotemporal alignment, structural analysis, evidence fusion, risk field construction, and safety constraint decision-making of multi-source data.

[0087] The agricultural machinery controller is used to execute functions such as holding, speed limiting, detouring, braking, or manual takeover prompts based on the control commands output by the vehicle-side computing unit.

[0088] The cloud platform is used for security incident playback, parameter calibration, and version management.

[0089] This embodiment provides a cross-view unmanned agricultural machinery safety operation system. First, it acquires field environment and agricultural machinery status information through multiple cameras and a status acquisition unit. Then, it obtains candidate targets and multi-source evidence through a basic visual model and cross-view temporal correlation in the vehicle-side computing unit. Next, it obtains joint evidence, uncertainty values, and unknown target scores through reliability discount fusion and unknown target discovery. Finally, it generates safety control actions through spatiotemporal probability occupancy risk field and control barrier constraints, thereby improving the operational safety of unmanned agricultural machinery in complex field environments.

[0090] Example 2:

[0091] This embodiment provides a method for safe operation of unmanned agricultural machinery from multiple perspectives, such as... Figure 2 As shown, it specifically includes:

[0092] Step 1: Collect multi-camera images, positioning and inertial navigation data, and agricultural machinery operation status data from different perspectives. Synchronize these data with the agricultural machinery control cycle as the reference to obtain time-synchronized multi-source operation data. Based on camera calibration parameters and positioning and inertial navigation data, establish coordinate transformation relationships from each camera coordinate system to the agricultural machinery body coordinate system. Based on the agricultural machinery's position, speed, heading angle, steering angle, braking performance, and implement operating width, predict the reachable area of ​​the agricultural machinery within a preset time range, generating a predicted travel corridor that simultaneously covers the agricultural machinery body and implement operating envelopes.

[0093] Step 2: Input the multi-camera images from different perspectives and the predicted driving corridor from the multi-source operation data obtained in Step 1 into the basic visual model for scene structure analysis, obtain candidate instances from different perspectives, and generate passable area candidates and restricted area candidates based on the candidate instances from different perspectives. The basic visual model is a visual model pre-trained on large-scale image data and adapted using agricultural operation scene images. It can identify known targets based on category cues and region cues, and discover potential target areas whose specific categories cannot be determined. Among them, passable area candidates and restricted area candidates are used to limit the spatial range of the subsequent risk field, and candidate instances are used for cross-view correlation and temporal tracking in Step 3.

[0094] Step 3: Based on the coordinate transformation relationship obtained in Step 1, project the candidate instances extracted in Step 2 from different perspectives onto a unified coordinate system. Perform cross-view correlation based on the spatial location, appearance features, scale changes, and direction of motion among the candidate instances, and combine them with images from adjacent time points for temporal tracking. For targets with clear categories, consistent cross-view matching, and continuous trajectories, generate known target detection evidence containing target category, location, velocity, and trajectory; for targets that cannot be classified into known categories but appear stably in multiple perspectives or at consecutive time points, or overlap with the predicted driving corridor, generate unknown target discovery evidence.

[0095] Step 4: Further decompose the known target detection evidence generated in Step 3 into semantic evidence, geometric evidence, and motion evidence. Semantic evidence represents the target category and recognition confidence; geometric evidence represents the target's position, size, and spatial relationship with the predicted driving corridor; and motion evidence represents the target's speed, direction of motion, and trajectory continuity. Determine the viewpoint reliability factor based on image quality, target occlusion degree, cross-viewpoint matching consistency, and temporal tracking continuity. Perform reliability discounting and fusion on the semantic, geometric, and motion evidence and the unknown target discovery evidence. When different pieces of evidence conflict, incorporate the conflict information into the uncertainty value and the unknown target score, and output the joint evidence, uncertainty value, and unknown target score.

[0096] Step 5: Based on the joint evidence obtained in Step 4, determine the confidence level and danger level of the target's existence, and predict the probability of the target occupying various spatial positions at different future times based on the target's position, speed, and direction of motion. Use the passable area candidates and restricted area candidates obtained in Step 2 as spatial constraints, and the predicted driving corridor, agricultural machinery speed, steering ability, and braking distance obtained in Step 1 as motion constraints. For targets with high uncertainty values ​​or unknown target scores, expand their predicted occupancy range and increase the corresponding risk value, thereby generating a spatiotemporal probability occupancy risk field characterizing the target's occupancy probability, collision likelihood, and potential collision consequences.

[0097] Step 6: Based on the spatiotemporal probability risk field generated in Step 5, predict the reachable trajectory of the agricultural machinery after executing different candidate actions. Combine this with the current speed, braking distance, steering capability, and implement operation envelope of the agricultural machinery to calculate the minimum distance between each reachable trajectory and the high-risk area. Based on the minimum distance, use control barrier constraints to eliminate candidate actions that do not meet the dynamic safety distance requirements to obtain the safe action feasible region. When the safe action feasible region is not empty, select the action with the lowest risk value and the smallest deviation from the original operation trajectory, and output maintain, speed limit, or detour control commands. When the safe action feasible region is empty or the risk value exceeds the safety threshold, output braking control commands and prompt manual intervention.

[0098] Step 7: When manual takeover occurs, perform security event caching and parameter calibration.

[0099] Furthermore, the generation of the time-synchronized multi-source operation data and the predicted driving corridor specifically includes:

[0100] The clock deviation is calculated based on the camera timestamp and the inertial navigation timestamp. Then, according to the agricultural machinery control cycle, the multi-camera images, positioning inertial navigation data and agricultural machinery operation status data are aligned to obtain multi-source operation data with a unified time reference.

[0101] Based on the camera extrinsic parameters and the agricultural machinery attitude in the synchronization data frame, a coordinate transformation relationship is established from each camera coordinate system to the agricultural machinery body coordinate system. The coordinate transformation relationship is used in step 3 to map the candidate instances extracted in step 2 to the agricultural machinery body coordinate system, rather than directly mapping the instances before the candidate instances are extracted.

[0102] Based on the agricultural machinery speed and steering status in multi-source operation data, combined with the outline of the agricultural machinery implements, maximum steering capacity and braking capacity, the reachable trajectory of the agricultural machinery within a preset time range is predicted, and the reachable trajectory is expanded by the outline of the vehicle body and agricultural machinery implements to generate a predicted driving corridor.

[0103] Furthermore, the scene structure analysis includes:

[0104] The basic visual model is a visual model that has been pre-trained on large-scale general image data and adapted using agricultural operation scene images. It can respond to point cues, bounding box cues, and text cues to complete target segmentation and feature extraction. The basic visual model is driven by point cues, bounding box cues, text cues, or a combination of point cues, bounding box cues, and text cues to generate candidate masks containing target category probabilities and region contours.

[0105] Candidate instances are selected based on the class relevance, boundary stability, and cross-frame preservation rate of the candidate masks, resulting in a list of candidate instances including target class, segmentation contour, location, and recognition confidence.

[0106] The ground area that is connected to the current location of the unmanned agricultural machinery and whose width meets the passage requirements of the agricultural machinery and implements is identified as a candidate passable area.

[0107] Regions that protrude from the ground, occupy passage space, or are continuous but whose category cannot be determined are identified as obstacle candidate regions.

[0108] The crop area, ditch area, preset no-entry area, and the area formed by safely expanding the obstacle candidate area according to the width of the agricultural machinery body, the outline of the agricultural machinery implements, and the positioning error in the candidate instances are identified as no-entry area candidates.

[0109] Further, step 3 includes:

[0110] Using the coordinate transformation relationship between each camera coordinate system and the agricultural machinery body coordinate system established in step 1, the candidate instances from different perspectives extracted in step 2 are mapped to the agricultural machinery body coordinate system, so as to obtain the position, occupied contour and spatial scale of each candidate instance in the agricultural machinery body coordinate system.

[0111] Based on the position of candidate instances in the agricultural machinery body coordinate system, the similarity of appearance features between the occupied contour and spatial scale, the overlap of bird's-eye view projection, parallax consistency, adjacent frame displacement, scale change and optical flow residual, it is determined whether candidate instances from different viewpoints or different times correspond to the same target, and the relationship between candidate instances is established.

[0112] Based on the aforementioned correlation, observations of the same target from different perspectives and at different times are merged to form a target trajectory that includes target category, location, occupied contour, movement speed, movement direction, and trajectory continuity. For targets with clear categories, consistent cross-perspective correlation, and continuous trajectories, known target detection evidence is generated. For targets whose categories cannot be determined, whose identification results from different perspectives conflict, but which appear continuously or overlap with the predicted driving corridor, unknown target discovery evidence is generated.

[0113] Furthermore, the evidence for the discovery of the unknown target is determined by at least two of the following: the distance between the feature vector of the candidate region and the known security target category prototype, the category prediction entropy, the reconstruction residual, and the cross-perspective semantic conflict.

[0114] When a candidate region cannot be stably classified into a known target category, and the candidate region overlaps with the predicted driving corridor, the score of the unknown target in the candidate region is increased.

[0115] Furthermore, the viewpoint reliability factor is jointly determined by the visibility score, occlusion score, image sharpness score, cross-viewpoint consistency score, and temporal motion residual, and the specific expression is as follows:

[0116] No. Reliability factor of each candidate instance Represented as:

[0117]

[0118] in, Visibility score, To cover up the score, Image sharpness score For cross-perspective consistency score, For time-series motion residuals, To preset weights, This is the normalization function.

[0119] When the reliability factor is lower than a preset threshold, the semantic evidence, geometric evidence, motion evidence and unknown target discovery evidence generated from the corresponding perspective are attenuated according to the reliability factor, and the attenuated part is included in the uncertainty to obtain the target evidence with reliability discount, which is then used for subsequent evidence fusion.

[0120] Furthermore, the reliability discounting and fusion includes:

[0121] Based on the category confidence, spatial features, motion features, and unknown target discovery evidence of known target detection evidence, we construct basic trust assignments for semantic evidence, geometric evidence, motion evidence, and unknown target discovery evidence, and obtain the initial trust levels of semantic evidence, geometric evidence, motion evidence, and unknown target discovery evidence for known targets, unknown targets, and undetermined sets.

[0122] The basic trust assignment is discounted based on the perspective reliability factor to obtain the discounted trust level, as expressed below:

[0123]

[0124] in, and These are the target assumptions before and after the discount. Trust level, and For the undetermined set before and after the discount, respectively. Trust level, For the first The reliability factor of each candidate instance;

[0125] The discounted evidence is fused together, and the confidence level of unknown targets or uncertain sets is increased based on the amount of evidence conflict to obtain joint evidence. The confidence level of uncertain sets is used as an uncertainty value, and the score of unknown targets is determined based on the confidence level of unknown targets and the degree of overlap between unknown targets and predicted driving corridors.

[0126] Furthermore, the spatiotemporal probability occupancy risk field includes spatial risk values ​​at multiple future discrete moments. These spatial risk values ​​are related to the target occupancy probability, collision consequences, uncertainty values, unknown target scores, and the predicted overlap of driving corridors.

[0127] Location In the future Spatial risk value The expression is:

[0128]

[0129] in, To occupy the probability, For the consequences of the collision, To understand uncertainty, For unknown scores, For corridor overlap penalty items, To predict the time, For weights.

[0130] Furthermore, the control barrier constraints are constructed based on the agricultural machinery reachability set, braking distance, maximum lateral acceleration, implement outline, positioning error, and safety margin;

[0131] Control Barrier Constraints It can be represented as:

[0132]

[0133] Feasible domain of safe actions that meet control barrier constraints Represented as:

[0134]

[0135] in, For the set of candidate actions, This is the current status of the agricultural machinery. For candidate control actions, Indicates the predicted minimum distance. To match vehicle speed Machinery outline and positioning error Relevant security thresholds.

[0136] When there is no speed-limiting action or bypass action that satisfies the control barrier constraint, a braking command and a manual takeover prompt are output.

[0137] Further, step 7 specifically includes:

[0138] When manual takeover, high-risk braking, high uncertainty, or continuous false stops occur, the vehicle-side caches multi-camera images, positioning status, operational status, joint evidence, risk field, and control results for the corresponding time period.

[0139] The cloud platform replays and verifies cached events, and adjusts the perspective reliability weight, unknown target score threshold, risk field weight, and safety distance parameter based on the verification results.

[0140] Example 3:

[0141] This embodiment describes a computer-readable storage medium storing a computer program that, when executed by a processor, implements a cross-view unmanned agricultural machinery safe operation method as described in any of Embodiment 2.

[0142] Example 4:

[0143] This embodiment describes a computer device, including:

[0144] Memory is used to store instructions.

[0145] A processor is configured to execute the instructions, causing the computer device to perform the operation of a cross-view unmanned agricultural machinery safety operation method as described in any of Embodiment 2.

[0146] Example 5:

[0147] This embodiment describes the specific implementation process of a cross-view unmanned agricultural machinery safe operation method, such as... Figure 3 As shown, it includes:

[0148] Step 101: Collect multi-camera images, positioning inertial navigation data, and agricultural machinery operation status data from different perspectives.

[0149] In this step, the multi-camera images include at least two perspectives from the front, side, and rear of the unmanned agricultural machinery. The positioning and inertial navigation data includes GNSS position, IMU attitude, heading angle, and positioning error. The agricultural machinery's operational status data includes vehicle speed, steering status, braking status, operating mode, implement outline, and implement working status.

[0150] Step 102: Perform scene structure analysis on camera images based on the basic visual model.

[0151] In this step, the vehicle-mounted computing unit uses point cues, box cues, text cues, or combinations thereof to drive the basic visual model to generate candidate masks. Candidate instances are then filtered based on the category relevance, boundary stability, and cross-frame preservation rate of the candidate masks. Ground areas connected to the current location of the agricultural machinery are identified as passable candidate areas, while crop areas, obstacle areas, and their safety expansion boundaries are identified as prohibited candidate areas.

[0152] More specifically, candidate instances include people, vehicles, rocks, ditches, farm implements, crop boundaries, temporary stockpiles, and unknown targets that cannot be clearly classified. For passable area candidates, the system determines whether they meet the requirements of the outer width of the farm implements and the predicted driving corridor; for prohibited area candidates, the system performs expansion processing based on safety margins to avoid collisions between the edges of the farm implements and obstacles.

[0153] Step 103: Perform cross-perspective correlation and time-series tracking.

[0154] In this step, candidate instances are mapped to the agricultural machinery vehicle coordinate system, and an instance association graph is established based on appearance features, geometric overlap, parallax consistency, and optical flow residuals. , is represented as:

[0155]

[0156] in, For appearance similarity, To achieve a bird's-eye view of the projection overlap, For timing consistency, The weights are: The known target detection branch outputs the category, location, and confidence score; the open set anomaly branch generates evidence for unknown targets based on the category prototype distance, prediction entropy, reconstruction residual, and semantic conflict.

[0157] Step 104: Perform reliability discounting and fusion.

[0158] In this step, for each candidate target, a reliability factor determined by visibility, occlusion, ambiguity, cross-view consistency, and temporal residuals is calculated. Basic trust assignments for semantic, geometric, and motion evidence are constructed and discounted according to the reliability factor. The amount of evidence conflict is transformed into the trust level of the unknown set, yielding joint evidence and cognitive uncertainty.

[0159] More specifically, the first Reliability factor of each candidate instance Represented as:

[0160]

[0161] in, Visibility score, To cover up the score, Image sharpness score For cross-perspective consistency score, For time-series motion residuals, To preset weights, This is the normalization function.

[0162] Furthermore, based on evidence discovered from unknown targets, the first... Unknown target score of each candidate instance , Represented as:

[0163]

[0164] in, For class prototype distance, To predict entropy, To reconstruct the residuals, For cross-perspective semantic conflict quantity, to As weight.

[0165] Furthermore, cross-perspective semantic evidence, geometric evidence, motion evidence, and evidence for the discovery of unknown targets are jointly fused according to a reliability discount factor. A candidate instance for the proposition Joint trust Represented as:

[0166]

[0167] in, Indicates the first Similar evidence to propositions Basic trust allocation, This represents the reliability discount factor for the corresponding evidence. The amount of conflict between multiple sources of evidence. For an unknown set, The conflict transition coefficient, This indicates the number of evidence types involved in the fusion. This formula is used to uniformly calculate conflicts between different perspectives, uncertainties caused by low-quality images, and evidence of unknown targets, to obtain the fusion result required for subsequent security assessments.

[0168] Step 105: Construct the spatiotemporal probability occupation risk field.

[0169] In this step, the risk field construction module calculates the risk value of each spatial grid at multiple discrete future moments within the bird's-eye view coordinate system. The risk value is determined by the target occupancy probability, collision consequence weight, unknown target score, uncertainty value, and overlap ratio with the predicted driving corridor.

[0170] More specifically, location In the future risk value for:

[0171]

[0172] in, To occupy the probability, For the consequences of the collision, To understand uncertainty, For unknown scores, For corridor overlap penalty items, To predict the time, For weights.

[0173] Furthermore, control the barrier function It can be represented as:

[0174]

[0175] Feasible domain of safe actions that meet control barrier constraints Represented as:

[0176]

[0177] in, For the set of candidate actions, This is the current status of the agricultural machinery. For candidate control actions, Indicates the predicted minimum distance. To match vehicle speed Machinery outline and positioning error Relevant security thresholds.

[0178] Step 106: Make a safety action decision.

[0179] In this step, the safety constraint decision module constructs control barrier constraints based on the agricultural machinery's reachability set, braking distance, maximum lateral acceleration, implement outline, and safety margin. When there is an action that satisfies the constraints, it outputs a hold, speed limit, or detour command; when there is no safe action or the uncertainty exceeds the limit, it outputs a braking command under the minimum risk state and prompts for manual intervention.

[0180] Furthermore, among multiple safety candidate actions, the safety constraint decision module determines the optimal safe action based on the cumulative risk field value, the score of the unknown target, the uncertainty value, the probability of the restricted area, and the action smoothness constraint. Optimal safe action Represented as:

[0181]

[0182] in, This represents the set of actions that satisfy the control barrier constraints. To predict the time domain, To perform the action Later at the predicted time The corresponding predicted driving corridor, Indicates the probability of occupancy. To understand uncertainty, For unknown scores, Indicates the probability of entering a restricted area. Indicates a reference action. to These are the weight parameters.

[0183] Step 107: Perform security event caching and edge-cloud collaborative calibration.

[0184] The secure action generation algorithm for cross-perspective evidence fusion includes the following steps:

[0185] Step A1: Acquire multi-view images Agricultural machinery status and predicted driving corridor and multi-view images Perform scene structure analysis to obtain the scene structure. and candidate instances;

[0186] Step A2 involves associating candidate instances from different perspectives and at different times to obtain the association results. And calculate the first Perspective reliability factor of candidate instances Unknown target score and uncertainty values ;

[0187] Step A3: Based on the candidate instance's location, trajectory, joint evidence, unknown target score, uncertainty value, and predicted driving corridor, construct a spatiotemporal probability occupancy risk field. ;

[0188] Step A4, based on the condition of the agricultural machinery And the spacetime probability occupying the risk field Establish control barrier constraints and determine the feasible domain for safe actions:

[0189]

[0190] in, For agricultural machinery to be in good condition Next, execute candidate actions Control barrier function at time;

[0191] Step A5, when the safe action is feasible The risk is either empty, or the maximum risk value of the spatiotemporal probability in the risk field reaches the risk threshold. At that time, the braking action is determined to be a safe action. And cache the corresponding security events;

[0192] Step A6, when the safe action is feasible Not empty and the maximum risk value is lower than the risk threshold. When necessary, determine the safe action according to the following formula:

[0193]

[0194] in, To execute candidate actions The corresponding trajectory risk, Indicates the original planned operation. The penalty coefficient for trajectory deviation;

[0195] Step A7, output safety action When manual takeover, high-risk braking, or high-uncertainty events occur, the vehicle-side caches images, agricultural machinery operation status, joint evidence, risk fields, and control results for the corresponding time period. The cloud platform replays and verifies the cached events and updates relevant parameters. After version verification, the parameter package is distributed to the vehicle-side in a grayscale manner. If the vehicle-side verification fails, it rolls back to the previous parameter version.

[0196] As can be seen from the above embodiments, the present invention utilizes multi-view images, unknown target scores, view reliability, and agricultural machinery motion constraints to make safety judgments, which can improve the ability of unmanned agricultural machinery to judge obstacles and abnormal targets in complex field environments and reduce frequent false stops or unstable action switching.

[0197] Without departing from the principles of this invention, the multi-camera sensing unit can be replaced by a combination of a camera and a lidar, millimeter-wave radar, ultrasonic radar, or depth camera; the basic visual model can be replaced by a semantic segmentation model, an instance segmentation model, a visual language model, or a multi-task sensing model; the evidence fusion method can adopt DS evidence theory, Bayesian fusion, probabilistic graphical models, or weighted graphical optimization methods; the risk field can be represented by a grid risk map, a cost map, or a spatiotemporally reachable region; and the security decision-making method can adopt a control barrier function, model predictive control, rule constraint control, or a hierarchical security control method.

[0198] The above-mentioned alternative solutions can all achieve the purpose of multi-source perception, unknown target judgment and safe action control in this invention, and are equivalent replacements for the technical solutions of this invention.

[0199] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0200] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0201] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0202] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0203] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for safe operation of unmanned agricultural machinery from multiple perspectives, characterized in that: Specifically, it includes: Step 1: Acquire multi-camera images, positioning and inertial navigation data, and agricultural machinery operation status data from different perspectives to obtain time-synchronized multi-source operation data and predict the driving corridor; establish the coordinate transformation relationship from each camera coordinate system to the agricultural machinery body coordinate system based on the camera calibration parameters and positioning and inertial navigation data. Step 2: Input multi-camera images from different perspectives in the multi-source operation data into the basic vision model to obtain candidate instances from different perspectives. Based on the candidate instances and the predicted driving corridor, generate candidate passable areas and candidate restricted areas. Step 3: Project candidate instances from different perspectives to a unified coordinate system based on the coordinate transformation relationship. Perform cross-view correlation based on the spatial position, appearance features, scale changes and motion direction of the candidate instances in the unified coordinate system. Combine the images at adjacent time points to perform temporal tracking and generate evidence of known target detection and evidence of unknown target discovery. Step 4: Based on the viewpoint reliability factor, perform reliability discounting and fusion on the evidence of known target detection and the evidence of unknown target discovery to obtain joint evidence, uncertainty value and unknown target score; Step 5: Construct a spatiotemporal probability occupancy risk field based on joint evidence, uncertainty values, unknown target scores, passable area candidates, prohibited area candidates, predicted driving corridors, and agricultural machinery movement constraints; Step 6: Based on the spatiotemporal probability of occupying the risk field, predict the reachable trajectory of the agricultural machinery after executing different candidate actions, calculate the minimum distance between each reachable trajectory and the high-risk area; based on the minimum distance, use control barrier constraints to eliminate candidate actions that cannot meet the dynamic safety distance requirements, obtain the safe action feasible domain, and output the hold, speed limit or detour control commands based on the safe action feasible domain.

2. The method for safe operation of unmanned agricultural machinery across perspectives according to claim 1, characterized in that: Also includes: Step 7: When the feasible domain for safe actions is not obtained, output braking control and manual takeover commands, and perform safety event caching and parameter calibration.

3. The method for safe operation of unmanned agricultural machinery from a cross-view perspective according to claim 1, characterized in that: The method for obtaining the time-synchronized multi-source operation data and predicting the driving corridor specifically includes: The clock deviation is calculated based on the camera timestamp and the inertial navigation timestamp. Then, according to the agricultural machinery control cycle, the multi-camera images, positioning inertial navigation data and agricultural machinery operation status data are aligned to obtain multi-source operation data with a unified time reference. Based on the agricultural machinery speed and steering status in multi-source operation data, combined with the outline of the agricultural machinery implements, maximum steering capacity and braking capacity, the reachable trajectory of the agricultural machinery within a preset time range is predicted, and the reachable trajectory is expanded by the outline of the vehicle body and agricultural machinery implements to generate a predicted driving corridor.

4. The method for safe operation of unmanned agricultural machinery across perspectives according to claim 1, characterized in that: Step 2 specifically includes: Drive the base vision model using point hints, box hints, text hints, or a combination of point hints, box hints, and text hints to generate candidate masks that include target category probabilities and region contours; Candidate instances are selected based on the class relevance, boundary stability, and cross-frame preservation rate of the candidate masks to obtain candidate instances containing target class, segmentation contour, location, and recognition confidence. The ground area that is connected to the current location of the unmanned agricultural machinery and whose width meets the passage requirements of the agricultural machinery and implements is identified as a candidate passable area. Areas that protrude from the ground, occupy passage space, or are continuous but whose category cannot be determined are identified as obstacle candidate areas. The crop area, ditch area, preset no-entry area, and the area formed by safely expanding the obstacle candidate area according to the width of the agricultural machinery body, the outline of the agricultural machinery implements, and the positioning error in the candidate instances are identified as no-entry area candidates.

5. The method for safe operation of unmanned agricultural machinery across perspectives according to claim 1, characterized in that: Step 3 specifically includes: Based on the coordinate transformation relationship, candidate instances from different perspectives are mapped to the agricultural machinery body coordinate system to obtain the position, occupied contour and spatial scale of each candidate instance in the agricultural machinery body coordinate system; Based on the position of candidate instances in the agricultural machinery vehicle coordinate system, the similarity of appearance features between the occupied contour and spatial scale, the overlap of bird's-eye view projection, parallax consistency, adjacent frame displacement, scale change and optical flow residual, it is determined whether candidate instances from different viewpoints or different times correspond to the same target, and the relationship between candidate instances is established. Based on the aforementioned correlation, observations of the same target from different perspectives and at different times are merged to form a target trajectory that includes target category, location, occupied contour, movement speed, movement direction, and trajectory continuity. For targets with clear categories, consistent cross-perspective correlation, and continuous trajectories, known target detection evidence is generated. For targets whose categories cannot be determined, whose identification results from different perspectives conflict, but which appear continuously or overlap with the predicted driving corridor, unknown target discovery evidence is generated.

6. The method for safe operation of unmanned agricultural machinery across perspectives according to claim 1, characterized in that: Step 4 specifically includes: Based on the category confidence, spatial features, motion features of known target detection evidence and unknown target discovery evidence, a basic trust assignment is constructed for known target detection evidence and unknown target discovery evidence, which serves as the initial trust level of known target detection evidence and unknown target discovery evidence for known targets, unknown targets and undetermined sets; Calculate the reliability factors for evidence of known target detection and evidence of unknown target discovery; When the reliability factor is lower than a preset threshold, the basic trust assignment of known target detection evidence and unknown target discovery evidence is discounted based on the reliability factor to obtain the discounted trust level of known targets, unknown targets and undetermined set; By fusing known target detection evidence and unknown target discovery evidence, the amount of evidence conflict is determined when different pieces of evidence support mutually exclusive target hypotheses. When the amount of evidence conflict reaches a preset conflict threshold, the amount of evidence conflict is allocated to the unknown target hypothesis and the undetermined set according to the degree of overlap between the candidate instance and the predicted driving corridor. The degree of trust corresponding to each objective hypothesis after combination is used as the joint evidence; The trust level corresponding to the undetermined set is used as the uncertainty value, and the score of the unknown target is determined based on the trust level corresponding to the unknown target hypothesis.

7. The method for safe operation of unmanned agricultural machinery across perspectives according to claim 1, characterized in that: Step 5 specifically includes: The area surrounding the unmanned agricultural machinery is divided into a spatial grid, and the preset prediction time domain is divided into multiple prediction times. Based on the target's position, velocity, direction of motion, and trajectory continuity, predict the probability that the target will occupy each spatial grid at different prediction times; The confidence level of the target's existence and potential collision consequences are determined based on joint evidence, and the predicted occupancy range of the corresponding target is expanded based on the uncertainty value or the score of the unknown target. Based on the spatial relationship between the predicted occupied area of ​​the target and the predicted driving corridor, passable area candidates, and restricted area candidates, the spatial risk value of each spatial grid at different prediction times is determined. Based on the spatial risk value, a spatiotemporal probability occupancy risk field is generated, representing the target occupancy probability, collision probability, and potential collision consequences.

8. A method for safe operation of unmanned agricultural machinery from a cross-view perspective according to claim 1, characterized in that: Step 6 specifically includes: Based on the spatiotemporal probability of occupying the risk field, the reachable trajectory of the agricultural machinery after performing different candidate actions is predicted. Combined with the current speed, braking distance, steering ability and the working envelope of the agricultural machinery, the minimum distance between each reachable trajectory and the high-risk area is calculated. Based on the minimum distance, candidate actions that cannot meet the dynamic safety distance requirements are eliminated using control barrier constraints, thus obtaining the feasible domain of safe actions; When the feasible domain of safety actions is not empty, select the optimal safety action from it and output the maintain, speed limit or detour control command; Among them, the feasible domain of the safety action The expression is as follows: ； In the formula, For the set of candidate actions, This is the current status of the agricultural machinery. For candidate control actions, To control barrier constraints; The expression for the control barrier constraint is as follows: ； In the formula, For the minimum distance, To match vehicle speed Machinery outline and positioning error Relevant security thresholds; The optimal safety action The expression is as follows: ； in, To predict the time domain, To perform the action Later at the predicted time The corresponding predicted driving corridor, Indicates the probability of occupancy. To understand uncertainty, For unknown scores, Indicates the probability of entering a restricted area. Indicates a reference action. , , , , For weight parameters, The target coordinates are given.

9. A method for safe operation of unmanned agricultural machinery from a cross-view perspective, as described in claim 2, characterized in that: Step 7 specifically includes: When manual takeover, high-risk braking, high uncertainty or continuous false stop events occur, cache multi-camera images, positioning and inertial navigation data, agricultural machinery operation status data, joint evidence, spatiotemporal probability occupancy risk field and control results for the corresponding time period; The cached security events are replayed for verification, and the view reliability weight, unknown target score threshold, risk field weight and dynamic security distance parameter are updated based on the verification results to obtain the updated parameter package; The updated parameter package is validated and distributed. If the validation is successful, the current parameter version is used; if the validation fails, it is rolled back to the previous parameter version.

10. A cross-view unmanned agricultural machinery safety operation system, characterized in that: include: Multi-camera sensing unit, status acquisition unit, vehicle-side computing unit, agricultural machinery controller, and cloud platform; The multi-camera sensing unit is used to collect images of the front, side and rear of the unmanned agricultural machinery. The status acquisition unit is used to acquire GNSS-IMU, wheel speed, steering status, equipment status, and operation mode; The vehicle-side computing unit is used for spatiotemporal alignment, structural analysis, evidence fusion, risk field construction, and safety constraint decision-making of multi-source data. The agricultural machinery controller is used to execute control commands output by the vehicle-side computing unit to maintain speed, limit speed, detour, brake, or prompt manual intervention. The cloud platform is used for security incident playback, parameter calibration, and version management.

11. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements a cross-view unmanned agricultural machinery safe operation method as described in any one of claims 1-9.

12. A computer device, characterized in that: include: Memory, used to store instructions; A processor is configured to execute the instructions, causing the computer device to perform the operation of a cross-view unmanned agricultural machinery safety operation method as described in any one of claims 1-9.

Images