Four-way vehicle path dynamic binding system and method based on sortation level

Abstract

The application relates to the technical field of industrial sorting, in particular to a four-way vehicle path dynamic binding system and method based on a sorting grade, which comprises the following steps: in the process of transporting a material frame by a four-way vehicle, fusing an edible fungus image in the frame and material frame traceability information into a lightweight identification model to output a sorting grade and a confidence level frame by frame to form a grade judgment sequence; analyzing corresponding destinations of each sorting grade into candidate road sections according to the order of each turning node; judging whether the highest confidence level reaches a binding threshold before reaching the turning node; when the highest confidence level reaches the binding threshold, binding the sorting grade with the highest confidence level with the candidate road section after the turning node and generating a turning and unloading instruction; when the highest confidence level does not reach the binding threshold, eliminating the sorting grade which is continuously lower than an elimination threshold, and sequentially extending the binding to the next turning node along the direction in which the uneliminated destination can reach. The application integrates grading into the transportation link, and continuous sorting can be realized in the transportation.

Description

Technical Field

[0001] This application relates to the field of industrial sorting technology, specifically a four-way vehicle path dynamic binding system and method based on sorting level. Background Technology

[0002] Existing edible mushroom sorting equipment typically first grades the edible mushrooms in the material crates at a fixed testing station, and then uses a four-way conveyor to transport the material crates to the corresponding unloading point based on the grading results. Some equipment also installs a visual recognition device on the four-way conveyor to capture images of the edible mushrooms in the crates during transport and identify their quality grade to determine the unloading point. However, in the actual process of the four-way conveyor traveling along the crisscrossing track network, the time when the four-way conveyor arrives at the track junction does not always coincide with the time when the quality grade of the edible mushrooms in the crate is identified. The four-way conveyor often reaches the junction before the quality grade is determined, requiring immediate turning towards a specific unloading point.

[0003] For example, in the transportation and sorting of button mushrooms, a four-way vehicle carrying a crate of mushrooms exits from the entrance and reaches the first fork in the road. At this point, the quality grade identification result of the mushrooms in the crate is still fluctuating between level one and level two. If the four-way vehicle turns into the corresponding branch road in the level one direction at the fork, and the subsequent identification result stabilizes at level two, the material crate cannot return to the branch road leading to the level two unloading point and can only be sent to the level one unloading point, resulting in mis-sorting due to mismatched grades. As another example, if a four-way vehicle stops before the fork while waiting for the identification result to stabilize, it occupies the fork and the preceding section of the track, forcing other four-way vehicles following behind to queue on the track.

[0004] If the mismatch between the time when the four-way vehicle selects its turning direction at the junction and the time when the quality grade is stably identified cannot be effectively resolved, the material crates will be sent to unloading points that do not match their quality grade, requiring manual re-sorting. Alternatively, the four-way vehicle may delay transportation time and obstruct the passage of subsequent vehicles due to stopping and turning back before the junction. Therefore, it is necessary to address the problem of difficulty in determining the transportation and sorting destination caused by the time when the four-way vehicle needs to select its turning direction at the track junction is earlier than the time when the quality grade of the edible fungi in the crate is stably identified.

[0005] In view of this, this application proposes a four-way vehicle path dynamic binding system and method based on sorting level. Summary of the Invention

[0006] To achieve the above objectives, this application provides a four-way vehicle path dynamic binding system and method based on sorting level, the specific technical solution of which is as follows:

[0007] Firstly, this application provides a method for dynamic binding of four-way vehicle paths based on sorting level, including:

[0008] During the transportation of edible fungi material frames in a four-way vehicle, images of the surface of the edible fungi in the frames are collected and the variety and harvesting traceability information of the material frames are read. The visual features of the images and the traceability information are fused and input into a lightweight recognition model, and the sorting level and credibility are output frame by frame, forming a level determination sequence that is updated progressively with the transportation process.

[0009] According to the order of each reversing node on the driving path, the unloading destination corresponding to each sorting level is parsed into sequentially connected candidate road segments, and a binding mapping between sorting level and candidate road segments is established. Before the four-way vehicle arrives at each reversing node, it is determined whether the highest confidence in the level judgment sequence has reached the set binding threshold.

[0010] When the highest confidence level reaches the binding threshold, the sorting level with the highest confidence level is bound to the candidate road segment after the reversing node, and converted into reversing and unloading control instructions; if the binding threshold is not reached, the sorting level with a confidence level that is continuously lower than the elimination threshold is eliminated, and the travel direction that can be reached by all non-eliminated destinations is selected. The binding is then carried over to the next reversing node, and the four-way vehicle is driven to complete the transportation and sorting along the bound road segment.

[0011] Preferably, the step of fusing image visual features and traceability information into the lightweight recognition model includes: extracting image visual feature vectors from the backbone of the lightweight recognition model, wherein the backbone includes alternately stacked depthwise separable convolutional layers and downsampling layers, as well as channel attention units located between the layers; encoding the variety, storage time converted from harvest time, and temperature and humidity levels corresponding to mushroom house number into traceability information encoding vectors; and concatenating the image visual feature vectors with the traceability information encoding vectors weighted by the fusion weight coefficients to form a fused feature vector and mapping it to the single-frame credibility of the material frame at each sorting level.

[0012] Preferably, the frame-by-frame output of sorting level and credibility includes: smoothing the single-frame credibility output by the lightweight recognition model frame by frame to obtain the cumulative credibility of each sorting level; renormalizing the cumulative credibility of each sorting level after each frame update; and taking the cumulative credibility of the starting frame of the recursion as the proportion of each sorting level in the historical sorting results of the same cultivation batch; and forming a record by combining the cumulative credibility of each sorting level, frame number, collection time, and the track node vertex number of the four-way vehicle, and retaining the most recent set number of records in the level determination sequence in a first-in-first-out manner.

[0013] Preferably, the step of parsing the unloading destinations corresponding to each sorting level into sequentially connected candidate road segments includes: abstracting the track network into a directed graph composed of vertices and edges; taking the current position of the four-way vehicle as the starting point, obtaining a reference path to the unloading point corresponding to each sorting level on the directed graph; dividing the reference path into a first segment, a middle segment, and a last segment according to the order in which the reversing nodes appear on the reference path; the first segment extends from the current position of the four-way vehicle to the first reversing node; the middle segment is located between two adjacent reversing nodes; and the last segment extends from the last reversing node to the unloading point.

[0014] Preferably, establishing the binding mapping from sorting level to candidate road segment includes: creating a row for each sorting level in tabular form, with the row containing fields such as sorting level number, candidate road segment sequence, start and end vertex numbers and road segment length of each road segment in the candidate road segment sequence, corresponding unloading point number, and binding status field; initially setting the binding status field of each sorting level in the binding mapping to pending binding, and storing a copy of the binding mapping on the scheduling server.

[0015] Preferably, the binding mapping includes: as the four-way vehicle passes through a reversing node, re-analyzing the remaining candidate road segments for each sorting level according to the position after passing, updating the candidate road segment sequence, removing the candidate road segments that have already been passed during the re-analysis, and removing the temporarily closed track segments in the self-directed graph issued by the scheduling server; the candidate road segment sequence contains common road segments composed of road segment intervals that are traversed by the reference paths of each sorting level.

[0016] Preferably, whether the highest confidence level in the judgment sequence reaches the set binding threshold includes: triggering a judgment when the distance between the four-way vehicle and the forward reversing node is reduced to the advance judgment distance, the advance judgment distance being jointly determined by the four-way vehicle's running speed, the link delay time from the initiation of the judgment to the completion of the reversing wheel assembly's action preparation, and the safety margin distance; reading the latest record from the judgment sequence, taking the maximum value of the cumulative confidence level of each sorting level in the record and comparing it with the binding threshold, and adopting a unified binding threshold for each sorting level.

[0017] Preferably, the step of binding the sorting level with the highest credibility to the candidate road segment after the reversing node includes: rewriting the binding status field of the sorting level with the highest credibility in the binding mapping to "bound", marking the row containing the remaining sorting levels as invalid, and using the candidate road segment sequence of the bound sorting level as the running path of the four-way vehicle from the current reversing node to the corresponding unloading point; after the binding is completed, the level determination sequence continues to be updated with image acquisition, and the running path is not rewritten with subsequent fluctuations in credibility.

[0018] Preferably, the conversion into reversing and unloading control commands includes: generating reversing control commands one by one according to each reversing node in the running path, wherein the reversing control command includes the reversing node vertex number, the entry direction, the exit direction, and the action type of the reversing wheel set; generating unloading control commands, wherein the unloading control commands include the target unloading point vertex number and the unloading trigger condition, wherein the unloading trigger condition is that the four-way vehicle's arrival detection signal at the target unloading point is valid; and writing the reversing control commands and unloading control commands into the vehicle-side control queue in the execution order.

[0019] Preferably, during the driving process, the four-way vehicle retrieves and executes the reversing control command and unloading control command one by one from the vehicle-side control queue in the execution order; after each command is executed, the execution confirmation is sent back to the dispatch server via the wireless communication module; if the execution of the command times out and no execution confirmation is received, the four-way vehicle performs parking protection and reports to the dispatch server.

[0020] Preferably, the sorting levels whose elimination credibility is continuously lower than the elimination threshold include: for sorting levels that are in the pending binding state and whose cumulative credibility obtained by recursive smoothing of credibility output frame by frame has been continuously lower than the elimination threshold for a set number of frames, the corresponding row is set to invalid in the binding mapping, and the corresponding unloading point is removed from the scope of subsequent judgment; the re-inspection level used to collect abnormal material boxes is used as a fallback level and does not participate in elimination, and sorting levels that are set to invalid are not restored in the current transportation task.

[0021] Preferably, the selection of travel directions that are accessible to all non-eliminated destinations includes: for each drivable outbound direction of the current reversing node, determining on the directed graph after removing the temporarily closed track segment whether there is still a feasible path to the unloading point corresponding to each non-eliminated sorting level; retaining the outbound directions for which there is a feasible path to the unloading point corresponding to all non-eliminated sorting levels.

[0022] Preferably, the departure direction includes: when there is more than one retained departure direction, selecting the departure direction with the largest sum of the number of remaining reversing nodes on the reference path of each non-eliminated sorting level; when the sum of the number of remaining reversing nodes is the same, selecting the departure direction with the smallest sum of the reference path length to the unloading point corresponding to each non-eliminated sorting level.

[0023] Preferably, the step of extending the binding to the next reversing node includes: during the extension period, the four-way vehicle continues to travel and does not turn back, and the grade determination sequence continues to be updated with subsequent image acquisition; when the four-way vehicle arrives at a reversing node in the outgoing direction where there is no unloading point that can be simultaneously reached by all non-eliminated sorting grades, the binding is no longer extended, but instead the binding is based on the sorting grade with the highest current credibility; when only the re-inspection grade remains of the sorting grades to be bound, the re-inspection grade is used as the binding grade.

[0024] Preferably, the process of driving the four-way vehicle to complete transportation and sorting along the bound road section includes: after the four-way vehicle arrives at the target unloading point and the arrival detection signal is valid, the vehicle-end control queue releases the unloading control command, the transfer rollers of the loading platform rotate forward and transfer the material box to the roller conveyor line of the target unloading point; the identification information of the material box, the bound sorting level, the unloading point number and the completion time are written into the task record, and the task record is archived together with the traceability information.

[0025] Secondly, this application also provides a four-way vehicle path dynamic binding system based on sorting level, which is used to implement the four-way vehicle path dynamic binding method based on sorting level, including: a sequence generation module, a mapping determination module, and a binding control module.

[0026] The sequence generation module is used to collect images of the surface of edible fungi in the four-way vehicle transporting the edible fungi material frame and read the variety and harvesting traceability information of the material frame. The image visual features and traceability information are fused and input into a lightweight recognition model, and the sorting level and credibility are output frame by frame to form a level determination sequence that is updated progressively with the transportation process.

[0027] The mapping determination module is used to resolve the unloading destination corresponding to each sorting level into sequentially connected candidate road segments according to the order of each reversing node on the driving path, establish a binding mapping between sorting level and candidate road segments, and determine whether the highest confidence in the level determination sequence reaches the set binding threshold before the four-way vehicle arrives at each reversing node.

[0028] The binding control module is used to bind the sorting level with the highest credibility to the candidate road segment after the reversing node when the highest credibility reaches the binding threshold, and convert it into a reversing and unloading control command; if the binding threshold is not reached, the sorting level with credibility continuously lower than the elimination threshold is eliminated, the travel direction that can be reached by all non-eliminated destinations is selected, the binding is extended to the next reversing node, and the four-way vehicle is driven to complete the transportation and sorting along the bound road segment.

[0029] The beneficial effects of this application are as follows: This application collects images of edible fungi within the frame during the transportation of edible fungi materials in a four-way vehicle and reads traceability information. The lightweight recognition model outputs the sorting level and credibility frame by frame to form a level judgment sequence. By fusing visual features with traceability information and recursively updating the frame-by-frame results, recognition and transportation are carried out simultaneously, solving the problem of misjudgment caused by fluctuations in lighting and occlusion in single-frame recognition.

[0030] This application resolves the destinations corresponding to each sorting level into sequentially connected candidate road segments according to the order of each reversing node, establishes binding mappings, and determines whether the highest confidence level has reached the binding threshold before arriving at the reversing node. By advancing the path resolution and mapping maintenance to be completed during the driving process, the path decision at the reversing node only needs to look up a table and compare, thus solving the problem of four-way vehicles slowing down and lingering at the reversing node due to waiting.

[0031] This application generates control instructions by binding the sorting level with the candidate road segment after the reversal node when the highest confidence level reaches the binding threshold. If the highest confidence level is not reached, sorting levels below the elimination threshold are eliminated and the travel direction that can reach all non-eliminated destinations is selected and bound in sequence. By giving up the decision-making opportunity in a way that preserves accessibility, the application solves the problem of misclassification and parking back-turn caused by the direction selection at the reversal node being earlier than the stable determination of the quality level.

[0032] The technical solution of this application overlaps the three stages of edible fungi quality identification, route decision-making and transportation in time. The material box does not need to stop at the inspection station from the entrance to the unloading point. At the same time, it does not turn back because the sorting grade is only determined after the four-way vehicle has passed the reversing node. Without adding independent inspection stations and sites, the quality grading is integrated into the transportation process, shortening the process time from harvesting to grading and warehousing, and keeping the mis-sorting ratio within the range that the re-sorting station can bear. Attached Figure Description

[0033] Figure 1 Flowchart of the four-way vehicle path dynamic binding method based on sorting level provided in this application;

[0034] Figure 2 This application provides a schematic diagram of a real-world track network scenario.

[0035] Figure 3 A schematic diagram of the lightweight recognition model architecture provided in this application;

[0036] Figure 4 A schematic diagram illustrating the binding mapping between sorting levels and candidate road segments provided in this application;

[0037] Figure 5 This application provides a schematic diagram of the advance distance triggering method.

[0038] Figure 6 The structural diagram of the four-way vehicle path dynamic binding system based on sorting level provided in this application. Detailed Implementation

[0039] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0040] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0041] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0042] Example 1

[0043] Reference Figures 1 to 5 This is the first embodiment of the present application, such as Figure 1 As shown, a method for dynamically binding four-way vehicle paths based on sorting level is provided.

[0044] The path dynamic binding method provided in this embodiment is applied to the post-harvest transportation and grading stages of industrialized edible fungi production. After the edible fungi are harvested and loaded into material crates in the mushroom house, the material crates are transferred to a four-way vehicle at the entrance docking station. The four-way vehicle then transports the material crates along the track network to various unloading points classified according to quality grades. In traditional operations, the material crates need to undergo static grading at a separate inspection station before being transferred by the conveyor equipment according to the grading results. Inspection and transportation are sequential in time, and the inspection station is prone to backlog during peak harvesting periods. This embodiment integrates quality identification and path decision-making into the transportation journey of the four-way vehicle. At the start of transportation, the sorting grade of the material crates has not yet been determined, and at the end of transportation, the material crates have been delivered to the unloading point that matches the sorting grade.

[0045] The track network consists of longitudinal tracks laid along the length of the storage area and transverse tracks laid along the width of the storage area. Reversing nodes are set at the intersections of the longitudinal and transverse tracks. At these reversing nodes, the four-way vehicle switches its travel direction between longitudinal and transverse by raising and lowering reversing wheel sets. For example, the track network includes 4 longitudinal tracks and 6 transverse tracks, with a total of 24 reversing nodes. An entrance connection point is set on one side of the track network, and along the end of the transverse tracks on the other side, unloading points for premium grade products, grade 1 products, grade 2 products, and re-inspection products are sequentially set. Each unloading point is equipped with a roller conveyor line and a buffer position. The roller conveyor line receives the material frames removed from the four-way vehicle.

[0046] Figure 2The diagram illustrates the track network layout, where longitudinal and transverse tracks intersect to form the network, with reversing nodes at the intersections. An entrance connection point is located on one side of the track network, where material boxes are transferred to the four-way trolley. Along the other side, at the end of the transverse tracks, unloading points are sequentially located for premium, first-grade, second-grade, and re-inspection products. Each unloading point is equipped with a roller conveyor and a buffer. The four-way trolley carries the material boxes within the track network, switching between longitudinal and transverse travel at the reversing nodes to transport the material boxes to the unloading point matching the sorting grade. Figure 2 The label indicates the longitudinal track, transverse track, reversing node, entrance connection point, four unloading points, four-way vehicle and material box.

[0047] The four-way vehicle comprises a body, a set of traveling wheels, a set of reversing wheels, a lifting mechanism, a loading platform, an onboard controller, a positioning unit, and a wireless communication module. The traveling wheels are used for travel along the longitudinal track, while the reversing wheels are used for travel along the transverse track. The lifting mechanism switches the contact state between the two sets of wheels and the track. The loading platform carries the material frame and has built-in reversible transfer rollers for removing the material frame at the unloading point. The positioning unit obtains the vertex number of the track node where the four-way vehicle is located by reading the landmark tags laid at the track nodes. The wireless communication module is responsible for the status and command interaction between the onboard controller and the scheduling server.

[0048] The vehicle-mounted identification hardware is installed on a bracket above the loading platform, including an industrial area scan camera and two sets of strip lights. The camera's field of view covers the opening of the material frame, and the lights provide stable illumination for the edible fungi inside the frame. An edge computing unit is installed inside the vehicle body below the bracket. The edge computing unit is connected to the vehicle controller via an internal bus, and all identification calculations are completed locally in the edge computing unit. A wireless RFID reader is installed on the side of the loading platform, and electronic tags are affixed to the side walls of the material frames. During the harvesting and loading process, the electronic tags are written with four traceability information items—variety code, harvest time, mushroom house number, and cultivation batch number—by the warehousing station.

[0049] In this embodiment, button mushrooms are used as the sorting object. Based on the diameter of the cap, the degree of opening of the cap, the color of the mushroom body and the mechanical damage, the quality is divided into four sorting grades: special grade, first grade, second grade and re-inspection grade. The re-inspection grade is used to collect material boxes that do not meet the judgment conditions of special grade, first grade and second grade in appearance or have abnormalities in the identification process. The four sorting grades correspond one-to-one with four unloading points.

[0050] Step S1: During the transportation of edible fungi material frames in the four-way vehicle, images of the surface of edible fungi in the frames are collected and the variety and harvesting traceability information of the material frames are read. The visual features of the images and the traceability information are fused and input into a lightweight recognition model. The sorting level and credibility are output frame by frame, forming a level determination sequence that is updated progressively as the transportation progresses.

[0051] Specifically, image acquisition and traceability information reading are completed first. After the four-way vehicle carrying the material frame drives out of the inlet docking position, the on-board controller sends a recognition start signal to the edge computing unit. The industrial area array camera then continuously acquires surface images of the edible fungi within the frame at a fixed frame rate. The acquisition is synchronized with the movement of the four-way vehicle, and the change in the position of the four-way vehicle naturally creates differences in the shooting angle between adjacent frames. For example, the acquisition frame rate is set to 10 frames per second. The frame rate is set based on the fact that sufficient frames need to be accumulated during the travel time between adjacent reversal nodes for subsequent recursive updates. At the same time, the product of the single-frame inference time and the frame rate must not exceed 1, so that the recognition calculation does not exceed the real-time processing capability of the edge computing unit.

[0052] The RFID reader reads the electronic tag once upon activation, caching the variety code, harvest time, mushroom house number, and cultivation batch number in the edge computing unit. This process avoids repeated readings during transit. This single read and cache operation reduces the number of interactions between the reader and the electronic tag, and vibrations caused by vehicle movement prevent the loss of traceability information en route.

[0053] Before entering the recognition process, the original image undergoes preprocessing, which includes three operations: performing fixed white balance correction according to factory calibration parameters; cropping the effective area within the frame based on the frame edge positioning results to eliminate interference from the track and shelf background; and scaling the cropped image to the input size of the lightweight recognition model. Frame edge positioning is achieved using edge detection with a fixed threshold combined with maximum connected component extraction. Since the relative position of the material frame on the loading platform is fixed and the position range of the frame edge in the image is stable, a fixed threshold is sufficient for positioning. After preprocessing, the edge computing unit calculates a sharpness evaluation value for each frame. The sharpness evaluation value is represented by image gradient statistics, a standard method in image quality evaluation, which will not be elaborated upon here. Frames with sharpness evaluation values ​​below the lower limit are directly discarded and not included in the recognition process. The lower limit is set based on the following: statistically analyzing the sharpness evaluation value distribution of image frames acquired under normal production line operation conditions and blurred frames acquired under artificially created shaking conditions, and taking the lower 5th percentile of the normal frame distribution as the lower limit. Discarding blurred frames avoids contamination of subsequent reliability updates by motion blur.

[0054] Before being used for identification, traceability information is encoded into a fixed-length vector. Variety coding uses unique thermal coding, with the coding dimension consistent with the number of product varieties on the production line. The harvest time is converted into the storage time from harvest to the current frame acquisition time, and then linearly normalized according to a preset upper limit for storage time. This upper limit is set based on the longest allowable residence time of button mushrooms under normal temperature circulation conditions; for example, the upper limit is 8 hours. The mushroom house number is mapped to the mushroom house temperature and humidity level code. The temperature and humidity control range for the same mushroom house is fixed, and the level code is taken from the preset temperature and humidity levels of the mushroom house environmental control system. All components are combined to form the traceability information coding vector. The reason for incorporating traceability information into the identification criteria is that different varieties of mushroom caps have different color and morphological benchmarks. Storage time is related to the degree of browning of the mushroom body; mushrooms stored for longer periods show deeper browning and more rapid cap opening, making them more likely to be classified as lower grade. The mushroom house environment affects the cap opening speed. Relying solely on appearance characteristics has limited ability to distinguish between samples with similar appearances. Traceability information provides a priori distinguishing factor beyond appearance. After the storage time normalization value is included in the fusion, for mushroom frames with longer storage times but appearances between adjacent grades, the identification results tilt towards the lower grade, consistent with the actual trend of edible fungi declining in quality over time.

[0055] Visual feature extraction is accomplished by the backbone of a lightweight recognition model. The backbone consists of alternating stacks of depthwise separable convolutional layers and downsampling layers, with lightweight channel attention units embedded between layers. The network terminates with a fixed-length image visual feature vector output via global average pooling. Depthwise separable convolution breaks down standard convolution into two steps: channel-wise convolution and pointwise convolution. Channel-wise convolution applies a spatial convolution kernel to each input channel individually, while pointwise convolution uses a kernel of size 1 to perform a linear combination of channels. This two-step breakdown significantly reduces computational cost compared to standard convolution. The channel attention unit first performs global average pooling on the features of each channel to obtain channel descriptors, then generates channel weights through two fully connected layers and a gating function. Channel weights are then multiplied sequentially to enhance feature channels related to mushroom cap lesions and umbrella edges, resulting in a small number of additional parameters. For example, the input image is uniformly scaled to 320×320 pixels, the dimension of the image visual feature vector is 128, the dimension of the source information encoding vector is 16, and the number of parameters in the backbone is approximately 1.5 million. The model size is determined by a trade-off between the validation set hierarchical accuracy and the computing power of the vehicle edge computing unit. The lightweight design aims to ensure that the entire recognition process is executed locally on the vehicle, and the recognition results do not need to be uploaded to the server via a wireless link and then transmitted back. Link fluctuations will not affect the judgment cycle.

[0056] Visual features and source features are merged in the fusion layer, and the fusion method can be represented as follows: ,in, Represents the fused feature vector. Represents the visual feature vector of an image. This represents the source information encoding vector. This represents the fusion weight coefficient used to balance the contributions of the source information encoding vector. Indicates will With scriptures Weighted Join the beginning and end. Integrate the weighting coefficients. The setting is based on: When the size is too large, the prior knowledge has too strong an influence on the identification results, and the hierarchical features that are clearly reflected in the appearance are masked by the prior knowledge. When the sample size is too small, the traceability information becomes almost invalid, and the ability to distinguish between samples with similar appearances declines. During setup, training is performed on several candidate values ​​on the validation set, and the grading accuracy is compared. The value whose accuracy falls within the peak range is selected. For example, The value is set to 0.6. The fused feature vector is mapped through a fully connected layer and then connected to a normalized exponential function to output the single-frame confidence of the material box at the four sorting levels. The sum of the four single-frame confidences is 1.

[0057] Figure 3 This demonstrates the structural composition of a lightweight recognition model. The input image enters the backbone, which consists of alternating stacks of depthwise separable convolutional layers and downsampling layers, with channel attention units embedded between layers. Figure 3 The stacking relationship is represented by side-by-side frames. The backbone section ends with global average pooling, and its output is an image visual feature vector. The traceability information encoding vector is fed into the fusion layer from below, where it is merged with the image visual feature vector. The merged result is then passed through a fully connected layer and followed by a normalized exponential function to output the credibility of the material frame at four sorting levels: special grade, first grade, second grade, and re-inspection. Figure 3 The annotations include the input image, backbone, depthwise separable convolutional layer, downsampling layer, channel attention unit, global average pooling, image visual feature vector, source information encoding vector, fusion layer, fully connected layer, normalized exponential function, and sorting level confidence.

[0058] The training samples for the lightweight recognition model are in-frame images collected and stored during historical transportation on the production line. These images are independently labeled by two quality inspectors according to sorting level. Samples with inconsistent labeling are submitted to a third party for arbitration to determine their level, ensuring consistent labeling standards. During training, three types of data augmentation are applied: brightness perturbation, partial occlusion, and viewpoint shift. These three types of augmentation simulate illumination attenuation, mushroom stacking occlusion, and changes in shooting angle caused by vehicle movement, respectively, enabling the model to adapt to imaging conditions during transportation. Cross-entropy with class weights is used as the loss function during training. The number of samples for each sorting level is unbalanced; sorting levels with low frequency are assigned higher class weights to prevent them from being overwhelmed during training. After training, the model undergoes 8-bit integer quantization and is redeployed to the edge computing unit. Quantization converts network weights and activations from floating-point representation to 8-bit integer representation, reducing the model's storage space and inference time. The accuracy loss on the validation set before and after quantization is confirmed to be within an acceptable range. For example, the inference time of a single frame of the quantized model does not exceed 30ms, which meets the acquisition cycle of 10 frames per second.

[0059] The reliability of a single frame fluctuates due to instantaneous changes in illumination and local occlusion. Directly using the result of a single frame as the basis for path decision-making can easily lead to misjudgment. This embodiment performs frame-by-frame smoothing on the reliability of a single frame to form a cumulative reliability score. The smoothing can be accomplished using the following recursive form: .in, The sequence number indicating the sorting level. Indicates the frame number. This indicates that the lightweight recognition model is in the first... Frame pair sequence number is The reliability of a single frame outputting the sorting level. Indicates the sequence number is The sorting level is at the 1st Cumulative credibility after frame update Indicates the sequence number is The sorting level is at the 1st Cumulative credibility after frame update This represents the smoothing coefficient.

[0060] Smoothing coefficient The basis for this setting includes: When the value is too large, the cumulative confidence level responds slowly to new observations, and the number of frames required for the identification conclusion to converge to the true level increases. When the value is too small, single-frame perturbations are directly transmitted to the cumulative confidence level, exacerbating sequence fluctuations. Smoothing coefficient The setup uses a verification video sequence as the object, and calculates the number of frames required for the cumulative credibility to stabilize to the true level under different values, as well as the number of false positives before stabilization. The value representing a compromise between these two metrics is then selected. For example, We set the value to 0.7. The cumulative confidence of the four sorting levels is renormalized after each frame update so that the sum of the four is kept at 1. Normalization avoids the drift of the numerical benchmark during the recursion process.

[0061] Cumulative confidence level of the recursive start frame Given by the prior distribution. Edible fungi from the same cultivation batch in the same mushroom house grow under consistent conditions. Therefore, the grade proportions of historical sorting results are valuable for reference in subsequent material frames from the same batch. The proportion of each sorting grade in the historical sorting results of the same batch is taken; if historical data is missing, it is initialized with a uniform distribution. For example, when there is no historical data, the initial cumulative confidence of each of the four sorting grades is 0.25. The advantage of batch prior initialization is that the recognition sequence has a reasonable tendency from the beginning stage, and the number of frames required for convergence is compressed. The cultivation batch number is used to retrieve the historical proportion of the same batch in the initial frame initialization stage, and the storage time converted from the harvest time participates in the recognition in the frame-by-frame fusion stage. The four traceability information each play a role in the recognition process.

[0062] After each frame update, the edge computing unit assembles a record by combining the cumulative reliability of the four sorting levels, the frame number, the acquisition time, and the vertex number of the track node where the four-way vehicle is currently located. This record is then appended and stored frame by frame. The record accumulates continuously as the vehicle travels, forming a recursively updated level determination sequence that progresses with the transportation process. The level determination sequence retains the most recent record in a first-in, first-out manner. For example, the most recent 50 records are retained. The number of records to retain is set based on the product of the travel time between adjacent reversing nodes and the acquisition frame rate, plus a margin of 100%, ensuring that any path determination can retrieve the update result covering the previous route segment.

[0063] The processing in step S1 is all completed within the travel distance of the four-way vehicle. The material box does not pass through a separate static detection station, and the identification does not take up time outside of the conveying process. Multi-frame recursion smoothly absorbs occasional single-frame misjudgments, and the traceability prior supplements the discrimination criteria for samples with insufficient appearance differentiation. The grade judgment sequence is continuously output in a credible form, providing quantifiable input for path determination.

[0064] Step S2: According to the order of each reversing node on the driving path, the unloading destination corresponding to each sorting level is parsed into sequentially connected candidate road segments, and a binding mapping between sorting level and candidate road segments is established. Before the four-way vehicle arrives at each reversing node, it is determined whether the highest confidence in the level judgment sequence has reached the set binding threshold.

[0065] Step S2 focuses on establishing and maintaining the mapping relationship between sorting levels and operating paths. The track network is abstracted as a directed graph in the scheduling server. The vertices of the directed graph correspond to reversing nodes, unloading points, and entrance docking points, while the edges correspond to track segments between adjacent vertices. Edge attributes record the track segment length and allowed travel direction. The directed graph is imported once based on the track construction drawings and revised synchronously during track modifications. Each time a four-way vehicle's positioning unit passes a landmark tag, it reports the vertex number it is currently at. The onboard controller synchronously saves the current position of the four-way vehicle in the directed graph. Landmark positioning is a standard method in the four-way vehicle field and will not be elaborated upon here.

[0066] The correspondence between sorting grades and unloading points is maintained on the scheduling server in the form of a configuration table: premium grade corresponds to premium grade unloading point, grade 1 corresponds to grade 1 unloading point, grade 2 corresponds to grade 2 unloading point, and re-inspection grade corresponds to re-inspection unloading point. The configuration table is modified by the management end when the production line switches product types or adjusts the grading standards. It remains unchanged during the execution of a single transportation task and is synchronized to the vehicle controller when the task is issued. The data is consistent between the vehicle end and the server.

[0067] The analysis of candidate road segments begins with the current position of the four-way vehicle. The onboard controller calculates a reference path for each sorting level's corresponding unloading point. This reference path is obtained on a directed graph using the shortest path algorithm; shortest path solving is a standard method in path planning and will not be elaborated upon here. After obtaining the reference path, it is divided into sequentially connected candidate road segments according to the order in which the reversing nodes appear: the first segment starts from the current position of the four-way vehicle and ends at the first reversing node along the route; the middle segment starts from one reversing node and ends at the next reversing node; and the final segment starts from the last reversing node and ends at the unloading point.

[0068] For example, the reference path corresponding to the primary product unloading point is a sequential connection of the entrance dock, reversing node J1, reversing node J2, and the primary product unloading point. After segmentation, three candidate road segments are obtained: from the entrance dock to reversing node J1, from reversing node J1 to reversing node J2, and from reversing node J2 to the primary product unloading point. The reference paths of the four sorting levels often overlap with each other in a section near the entrance dock, and the overlapping part constitutes a common road segment. After leaving the common road segment, each reference path separates step by step at different reversing nodes. The existence of the common road segment provides the spatial conditions for continuing to drive when the sorting is not yet mature.

[0069] Figure 4 The correspondence between sorting levels and candidate routes in the track network is shown in the form of a directed graph. Figure 4 The left side is the entrance connection point. On the main line, reversing nodes J1, J2, J3, and J4 are distributed in sequence. Each reversing node leads upwards to the unloading point for premium products, first-grade products, second-grade products, and re-inspection products, respectively. Figure 4The reference route for the first-grade product is shown in bold. This route starts from the entrance docking point, passes through J1 and J2, and leads to the first-grade product unloading point. It is divided into the first segment, the middle segment, and the last segment according to the sequence of reversal nodes. The segment from the entrance docking point to J1 is the overlapping point of all reference routes and is marked as the common route segment. Figure 4 The map shows the entrance connection point, reversal nodes J1 to J4, four unloading points, reference route, first segment, middle segment, last segment, and common road segment.

[0070] The parsing results are organized into a binding mapping from sorting level to candidate road segments. This binding mapping is stored in a table on the vehicle controller. Each row in the table corresponds to a sorting level, and the fields within each row are, in order: sorting level number, candidate road segment sequence, start and end vertex numbers and road segment length for each segment within the candidate road segment sequence, corresponding unloading point number, and binding status field. The binding status field is initially set to "pending binding." The scheduling server retains a copy of the binding mapping for task monitoring, and this copy is updated synchronously with any updates on the vehicle.

[0071] Each time the four-way vehicle passes a reversal node, the onboard controller re-analyzes the remaining candidate road segments for each sorting level based on the new position after the vehicle's departure. Road segments that have already been passed are removed from the binding mapping, and temporary closed track segments issued by the scheduling server are removed from the directed graph during re-analysis. Re-analysis only involves the remaining parts, limiting the computational load. For example, with an adjacent reversal node spacing of 3m and a four-way vehicle speed of 1.5m / s, the travel time between two nodes is 2s, and the re-analysis time is no more than 50ms, sufficient to complete within the travel time. Through continuous maintenance, the binding mapping always reflects the actual road segment composition from the current position of the four-way vehicle to each unloading point. Decisions at reversal nodes only require reading the table, eliminating the need for temporary path calculations.

[0072] The reversing wheel assembly requires time from lifting and starting to being ready for reversing. Path determination must be initiated before the four-way vehicle reaches the reversing node, allowing sufficient preparation time for the reversing mechanism. Therefore, the trigger position for the determination needs to be determined in advance. The advance determination distance can be determined by the following formula: .in, This indicates that the distance should be determined in advance. The speed of the four-way vehicle is measured in real time by the onboard encoder. This represents the link delay time from the initiation of the judgment to the completion of the reversing wheel assembly's action preparation. This indicates the safety margin distance. When the distance between the four-way vehicle and the forward reversing node decreases to the pre-determined distance, the vehicle controller triggers a determination.

[0073] Figure 5This demonstrates the spatial relationship for a four-way vehicle to trigger a judgment before a reversing node. The four-way vehicle travels along the longitudinal track in its direction of travel, with the reversing node ahead. When the distance between the four-way vehicle and the reversing node decreases to the pre-judgment distance, a judgment is triggered at the judgment triggering position. Figure 5 The trigger position is marked with a vertical dotted line. The distance between the trigger position and the reversing node is the advance judgment distance, which is used for the preparation of the reversing wheel assembly movement. Figure 5 The preparation interval is indicated in brackets above. The distance to the reversing node is marked as a safety margin distance. Figure 5 The label includes information on the four-way vehicle, driving direction, longitudinal track, reversing node, trigger position, advance judgment distance, safety margin distance, and reversing wheel assembly preparation.

[0074] Link latency The coverage determination calculation time, the instruction transmission time within the vehicle controller, and the time for the reversing wheel assembly lifting action are all considered. The values ​​are based on the link delay statistics obtained from multiple actual measurements of the prototype, and to account for coverage delay fluctuations, the upper quantile value of the measured distribution is used. For example, Take 0.8s. Safety margin distance. This is used to absorb positional deviations caused by landmark positioning errors and wheel-rail slippage. The value is determined by the sum of the measured upper limit of the positioning error and the measured upper limit of the slippage. For example, Take 0.3m. For example, when a four-way vehicle is traveling at a speed of 1.5m / s, the advance judgment distance is calculated as 1.5×0.8+0.3, which is 1.5m. That is, the judgment is triggered when the four-way vehicle travels to a distance of 1.5m from the next reversal node.

[0075] After the trigger is determined, the vehicle controller reads the latest record from the classification sequence, takes the maximum value among the cumulative confidence scores of the four classification levels in the record, and then compares the maximum value with the bound threshold. Comparison: The maximum value is reached The binding conditions are met; the maximum value has not been reached. If the binding conditions are not met, the determination will be made. Each determination consists of only one read and one comparison, and the time consumed is negligible compared to the link latency. The determination process will not interfere with the preparation time of the switching mechanism.

[0076] Binding threshold The setting is based on: The conditions for initiating the binding action are determined by several factors. A low value indicates that insufficiently reliable judgments are allowed to proceed, resulting in incorrect binding and the material boxes being sent to mismatched unloading points, requiring manual re-sorting. A high value generally delays binding, compressing the decision margin for subsequent reversal nodes. The process involves: statistically analyzing the cumulative reliability distribution of correctly judged and incorrectly judged samples at the judgment time on the validation set; determining the upper limit of the allowable misbinding ratio based on the processing capacity of the re-sorting station; and selecting a threshold value from the two distributions where the proportion of incorrect samples meeting the binding conditions does not exceed the upper limit of the misbinding ratio, using this threshold value as the binding threshold. For example, the upper limit of the misbinding ratio is set to 2%, and the corresponding binding threshold... Set the threshold to 0.85. The threshold is set to a uniform value for all four sorting levels, avoiding ambiguity caused by inconsistent judgment criteria between different sorting levels.

[0077] Step S2 moves the path parsing and mapping maintenance to the driving process, and the path determination at the reversing node is compressed into two actions: table lookup and comparison. The initiation position of the determination is constrained by the pre-determined distance, and the reversing mechanism obtains a certain action preparation time. When the four-way vehicle passes through the reversing node, it does not slow down or linger due to waiting for the determination result, and the continuity of track passage is maintained.

[0078] Step S3: When the highest confidence level reaches the binding threshold, the sorting level with the highest confidence level is bound to the candidate road segment after the reversing node, and converted into reversing and unloading control instructions; if the binding threshold is not reached, the sorting level with a confidence level that is continuously lower than the elimination threshold is eliminated, and the travel direction that can be reached by all non-eliminated destinations is selected. The binding is then extended to the next reversing node, and the four-way vehicle is driven to complete the transportation and sorting along the bound road segment.

[0079] Step S3 is executed along two branches based on the judgment result, corresponding to the execution of the binding and the continuation of the binding respectively.

[0080] When the binding conditions are met, the vehicle controller rewrites the binding status field of the sorting level corresponding to the highest cumulative confidence level as "bound" in the binding mapping. At the same time, it marks the row containing the remaining sorting levels as invalid and releases the candidate road segment record occupied by the invalid row. The candidate road segment sequence of the bound sorting levels constitutes the subsequent running path of the four-way vehicle, starting from the current reversing node and ending at the unloading point corresponding to the bound sorting level.

[0081] The onboard controller generates two types of control commands based on the bound candidate road segment sequence. Reversing control commands are generated sequentially for each reversing node within the sequence. Each reversing control command includes the reversing node vertex number, entry direction, exit direction, action type of the reversing wheel set, command sequence number, and verification field. The action type distinguishes between lifting and lowering. Unloading control commands include the target unloading point vertex number, unloading action type, and unloading trigger condition. The unloading trigger condition is set to a valid arrival detection signal from the four-way vehicle at the target unloading point, which is provided by the position switch at the unloading point. All commands are written into the vehicle-side control queue in execution order and are retrieved and executed sequentially by the onboard controller during operation.

[0082] After each instruction is executed, the on-board controller sends an execution confirmation back to the dispatch server via the wireless communication module. The dispatch server updates the task status of the four-way vehicle based on the confirmation. If an instruction execution times out without confirmation, the on-board controller initiates a parking protection procedure and reports it to the dispatch server. Parking protection is a standard safety interlock for four-way vehicle control and will not be discussed further here. This instruction-level confirmation mechanism allows the dispatch server to keep track of the execution progress of each four-way vehicle, and troubleshooting can be done by locating the specific instruction when an anomaly occurs.

[0083] Once the binding is complete, the level determination sequence continues to update with image acquisition, but the bound running path is no longer rewritten due to subsequent fluctuations in the accumulated credibility. The binding retention process avoids the four-way vehicle repeatedly rewriting instructions due to small fluctuations in credibility when approaching the reversing node. The number of times the reversing wheel set is raised and lowered is controlled, and the wear of the mechanism and driving disturbance are reduced accordingly. The four-way vehicle's movement rhythm is stable when passing through subsequent reversing nodes.

[0084] When the binding conditions are not met, the vehicle controller first performs an elimination. The vehicle controller traverses the sorting levels in the pending binding state in the binding mapping, and continuously checks the cumulative trustworthiness. Frames below the elimination threshold The sorting level is eliminated, and the eliminated sorting level is marked as invalid in the binding mapping, and the corresponding unloading point is removed from the subsequent judgment range. Elimination only applies to the three sorting levels: Special Grade, Level 1, and Level 2. The re-inspection level serves as a fallback level and is not involved in elimination. The fallback arrangement ensures that at least one usable sorting level always exists within the judgment range. The invalidated state will not be restored within the current transportation task. The reason for not restoring is that the elimination conditions have already constrained the probability of erroneous elimination to a low level. If a restoration mechanism were set, the binding mapping would repeatedly change with the reliability fluctuations, which would actually impair the stability of the judgment.

[0085] Elimination threshold The setting is based on: statistically analyzing the lower bound of the cumulative confidence fluctuation of true grade samples during the identification process on the validation set, and selecting a low value that the cumulative confidence of the true grade is unlikely to be reached as the elimination threshold, ensuring that the elimination action only applies to sorting grades that are truly invalid. For example, Set to 0.1. Number of consecutive frames. Used to filter out short-term interference, which includes two typical situations: reflections from adjacent storage locations sweeping across the field of view and shadows from the frame edge obstructing the view. The value is determined by rounding up the product of the acquisition frame rate and the measured short-term interference duration, leaving a margin. For example, if the acquisition frame rate is 10 frames per second and the measured short-term interference duration does not exceed 0.3 seconds, it is converted to 3 frames, leaving a margin. Take 5.

[0086] After the elimination process is completed, the onboard controller verifies each feasible outbound direction of the current reversing node. The verification method is as follows: assuming the four-way vehicle exits along the verified direction, a reachability search is performed on the directed graph after removing temporary closed track sections. This determines whether a feasible path still exists for the unloading point corresponding to each non-eliminated sorting level. Reachability search is a conventional method in graph theory and will not be elaborated here. After verification, the outbound directions that ensure all non-eliminated unloading points remain reachable are retained. By selecting outbound directions based on the overall reachability of non-eliminated unloading points, it avoids prematurely entering branches and isolating unloading points corresponding to sorting levels that have not yet been eliminated from the feasible range.

[0087] When there are more than one exit direction that meets the reachability conditions, the onboard controller determines a unique exit direction according to two levels of rules. The first level rule calculates the sum of the number of remaining reversing nodes on the reference paths of each non-eliminated sorting level after traveling along each candidate exit direction, and selects the exit direction with the largest sum. The larger the sum of remaining reversing nodes, the more decision positions the four-way vehicle can bind to in the subsequent journey, and the lower the risk of being forced to a dead end at a branch. If the results of the first level are the same, the second level rule is activated: calculates the sum of the reference path lengths from each candidate exit direction to each non-eliminated unloading point, and selects the exit direction with the smallest sum. The smaller the sum of reference path lengths, the lower the detour cost for the four-way vehicle to reach each candidate unloading point. Both levels of rules are deterministic, and the selection result is unique under the same state; the direction determination process does not contain any random components. This unique direction selection ensures that different four-way vehicles make consistent decisions under the same track state, facilitating the scheduling server's prediction and coordination of traffic flow.

[0088] After determining the exit direction, the onboard controller generates only one reversing control command for the current reversing node and writes it to the vehicle-side control queue. The binding action is then carried over to the next reversing node along the exit direction. After the four-way vehicle passes the current reversing node, the level determination sequence continues to be updated with subsequent data collection. The binding mapping is refreshed with location according to the maintenance method in step S2, and the determination process is initiated again at the pre-determination distance before the next reversing node. During the extension period, the four-way vehicle remains in a driving state without stopping or turning back. As the four-way vehicle travels on public road sections, the number of frames required for identification continuously accumulates, the cumulative reliability increases accordingly, and the probability that the binding conditions will be met at subsequent reversing nodes increases accordingly.

[0089] In the track network, there exists a type of reversing node where, after a four-way vehicle leaves this node, there is no longer a direction from which all non-eliminated unloading points can be simultaneously reached; all non-eliminated reference paths completely separate at the reversing node. When a four-way vehicle arrives at a completely separated reversing node and the binding conditions are still not met, the onboard controller no longer follows the previous path but instead performs binding according to the sorting level with the highest current cumulative confidence. The command generation method after binding is consistent with that of the branch that meets the binding conditions. When a four-way vehicle arrives at a completely separated reversing node, all three sorting levels—Special Grade, Grade 1, and Grade 2—have been eliminated, or the highest cumulative confidence among the three is below the elimination threshold. At this time, the on-board controller sets the re-inspection level to the binding level, and the four-way vehicle drives to the re-inspection unloading point, where the manual workstation re-grades the edible fungi in the box. Boundary handling ensures that any material box will inevitably obtain a definite running path within a limited number of reversing nodes, the judgment process will not be suspended, and no vehicles with uncertain destinations will be stranded on the track.

[0090] The four-way vehicle travels along the designated candidate route sequence to the target unloading point. Once the arrival detection signal is valid, the vehicle-side control queue releases the unloading control command. The transfer rollers on the loading platform rotate forward, moving the material frame to the roller conveyor line at the unloading point. An unloading completion signal is then transmitted back to the scheduling server. The scheduling server writes the material frame's electronic tag number, the associated sorting level, the unloading point number, and the completion time into the task record. This task record, along with traceability information, is archived to form a quality file for the cultivation batch. This quality file can be used for subsequent planting management and process adjustments. After unloading, the four-way vehicle's status is set to idle, and the scheduling server assigns the next transportation task.

[0091] When multiple four-way vehicles operate in parallel on the production line, each four-way vehicle independently executes the identification, judgment, and binding processes of steps S1 to S3. The mutual exclusion of reversing nodes and track segments among multiple vehicles is coordinated by the scheduling server using conventional traffic control methods. Traffic control is not an improvement in this embodiment and will not be discussed here. Since each four-way vehicle obtains a consistent departure direction based on two levels of deterministic rules under the same track conditions, the scheduling server can predict its subsequent direction based on the current position of each vehicle and its bound mapping copy, thereby pre-arranging the occupancy order of reversing nodes, and the path conflicts when multiple vehicles are operating in parallel are correspondingly reduced.

[0092] To facilitate understanding, a typical transportation and sorting process will be used as an example to illustrate the complete workflow. For instance, in... Figure 2 In the grid-like track network shown, each unloading point has a feasible path via different reversing nodes. A four-way vehicle carrying a frame of button mushrooms drives out from the entrance docking position, passing through reversing nodes J1, J2, and J3 in sequence. The binding threshold is set to 0.85, and the elimination threshold is set to 0.1. When the four-way vehicle reaches a distance of 1.5m from reversing node J1, the first judgment is triggered. In the latest record of the level judgment sequence, the cumulative confidence of level 1 is 0.62, which is the maximum value among the four sorting levels. It does not reach the binding threshold, and no sorting level meets the elimination condition of 5 consecutive frames below 0.1. Reversing node J1 has two possible exit directions: straight and reversing. The straight direction makes all four unloading points reachable and the total number of remaining reversing nodes is greater. The on-board controller selects the straight direction, and the four-way vehicle drives straight through reversing node J1. When the four-way vehicle reaches a distance of 1.5m from the reversing node J2, the second judgment is triggered. The cumulative confidence of Level 1 rises to 0.79, but still does not reach the binding threshold. The cumulative confidence of Level 2 has been below 0.1 for 5 consecutive frames, so Level 2 is eliminated and set to invalid in the binding mapping. The remaining three unloading points, namely Special Level, Level 1 and Re-inspection, are all accessible in the straight direction, and the four-way vehicle passes through the reversing node J2 in the straight direction. The four-way vehicle triggers the third judgment when it reaches 1.5m from the reversing node J3. The cumulative confidence of the first level reaches 0.90, exceeding the binding threshold. After the four-way vehicle passes the reversing node J2, the remaining candidate road segments have been re-analyzed according to the new position. The reference path corresponding to the first level is updated to lead to the first-level product unloading point via the reversing node J3. The on-board controller binds the first level with the candidate road segments leading to the first-level product unloading point after the reversing node J3, and generates the reversing control command at the reversing node J3 and the unloading control command at the first-level product unloading point. The four-way vehicle completes the lifting and reversing at the reversing node J3, enters the transverse track, and is triggered to unload by the arrival signal after arriving at the first-level product unloading point. The material box is transferred to the roller conveyor line of the first-level product unloading point, and the status is returned to the scheduling server.

[0093] Step S3 ensures that the determination of insufficient credibility does not come at the cost of stopping and waiting or turning back, but instead postpones the decision-making opportunity by preserving accessibility; the elimination mechanism continuously narrows the candidate range during the postponement process, and the two-level deterministic rules for direction selection eliminate execution ambiguity; the fallback level and boundary handling close all abnormal exits, and the transportation and sorting process of any material box can converge within a certain number of steps.

[0094] In summary, steps S1 to S3 above are used in this embodiment. The identification, judgment, and transportation stages overlap in time. The material frame does not need to stop at the inspection station from the entrance docking point to the unloading point, and there is no backtracking caused by the sorting grade being determined only after the four-way vehicle has passed the reversing node. The time occupied by the reversing node and track section is reduced, ensuring the production line's conveying capacity during peak harvesting periods. The combination of multi-frame recursion and binding thresholds and elimination thresholds keeps the proportion of incorrect sorting within the acceptable range of the re-sorting station. Identification is completed locally on the vehicle-mounted edge computing unit, and the wireless link only handles the transmission of judgment results and status information, resulting in a light communication burden and low computational pressure on the scheduling server. This method integrates quality grading into the transportation stage without increasing the investment in independent inspection stations and facilities, shortening the process time from harvesting to grading and warehousing of edible fungi. The sorting grade, path destination, and traceability information of the material frame form a closed loop in the task record, facilitating production line traceability and statistics.

[0095] The above embodiments are merely illustrative examples of this application. Equivalent substitutions and conventional adjustments made by those skilled in the art based on the embodiments should all fall within the protection scope defined by the claims of this application.

[0096] Example 2

[0097] Reference Figure 6 This is the second embodiment of the present application, which provides a four-way vehicle path dynamic binding system based on sorting level.

[0098] The system includes: a sequence generation module, a mapping determination module, and a binding control module;

[0099] The sequence generation module is used to collect images of the surface of edible fungi in the four-way transport of edible fungi material frames and read the variety and harvesting traceability information of the material frames. The image visual features and traceability information are fused and input into a lightweight recognition model, and the sorting level and credibility are output frame by frame to form a level determination sequence that is updated progressively with the transportation process.

[0100] The mapping determination module is used to resolve the unloading destination corresponding to each sorting level into sequentially connected candidate road segments according to the order of each reversing node on the driving path, establish a binding mapping between sorting level and candidate road segments, and determine whether the highest confidence in the level determination sequence reaches the set binding threshold before the four-way vehicle arrives at each reversing node.

[0101] The binding control module is used to bind the sorting level with the highest credibility to the candidate road segment after the reversing node when the highest credibility reaches the binding threshold, and convert it into a reversing and unloading control command; if the binding threshold is not reached, the sorting level with credibility continuously lower than the elimination threshold is eliminated, the travel direction that can be reached by all non-eliminated destinations is selected, the binding is extended to the next reversing node, and the four-way vehicle is driven to complete the transportation and sorting along the bound road segment.

[0102] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0103] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of this application without departing from the spirit and scope of protection of the claims. All of these variations are within the protection scope of this application.

[0104] It should be noted that the collection, storage, and use of all personal information involved in the technical solution of this application must be carried out only after obtaining the explicit authorization and separate consent of the information subject. The processing of personal information-related data strictly complies with the requirements of currently effective national laws, regulations, and relevant standards and specifications. The collected personal information is limited to use within the specific purpose necessary to achieve the technical solution of this application, and no processing beyond that purpose will be performed. Necessary technical and management measures will be taken to ensure the security of personal information.

Claims

1. A method for dynamic binding of four-way vehicle paths based on sorting level, characterized in that, include: During the transportation of edible fungi material frames in a four-way vehicle, images of the surface of the edible fungi in the frames are collected and the variety and harvesting traceability information of the material frames are read. The visual features of the images and the traceability information are fused and input into a lightweight recognition model, and the sorting level and credibility are output frame by frame, forming a level determination sequence that is updated progressively with the transportation process. According to the order of each reversing node on the driving path, the unloading destination corresponding to each sorting level is parsed into sequentially connected candidate road segments, and a binding mapping between sorting level and candidate road segments is established. Before the four-way vehicle arrives at each reversing node, it is determined whether the highest confidence in the level judgment sequence has reached the set binding threshold. When the highest confidence level reaches the binding threshold, the sorting level with the highest confidence level is bound to the candidate road segment after the reversing node, and converted into reversing and unloading control instructions; if the binding threshold is not reached, the sorting level with a confidence level that is continuously lower than the elimination threshold is eliminated, and the travel direction that can be reached by all non-eliminated destinations is selected. The binding is then carried over to the next reversing node, and the four-way vehicle is driven to complete the transportation and sorting along the bound road segment.

2. The four-way vehicle path dynamic binding method based on sorting level according to claim 1, characterized in that, The process of fusing image visual features with source tracing information and inputting it into the lightweight recognition model includes: Visual feature vectors of images are extracted from the backbone of a lightweight recognition model. The backbone contains alternating stacked depthwise separable convolutional layers and downsampling layers, as well as channel attention units located between the layers. The storage time converted from variety and harvest time, as well as the temperature and humidity level corresponding to the mushroom house number, are encoded into traceability information encoding vectors. The image visual feature vector is concatenated with the traceability information encoding vector after being weighted by the fusion weight coefficient to form a fusion feature vector, which is then mapped to the single-frame credibility of the material frame at each sorting level.

3. The four-way vehicle path dynamic binding method based on sorting level according to claim 1, characterized in that, The frame-by-frame output sorting level and reliability include: The single-frame confidence of the lightweight recognition model is recursively smoothed frame by frame to obtain the cumulative confidence of each sorting level. After each frame is updated, the cumulative confidence of each sorting level is renormalized, and the cumulative confidence of the starting frame of the recursion is taken as the proportion of each sorting level in the historical sorting results of the same cultivation batch. The cumulative reliability of each sorting level, frame number, collection time, and the vertex number of the track node where the four-way vehicle is located are combined into a record, and the most recent set number of records are retained in the level determination sequence in a first-in-first-out manner.

4. The four-way vehicle path dynamic binding method based on sorting level according to claim 1, characterized in that, The process of resolving the unloading destinations corresponding to each sorting level into sequentially connected candidate routes includes: The track network is abstracted as a directed graph consisting of vertices and edges. Starting from the current position of the four-way vehicle, the reference path to the unloading point corresponding to each sorting level is obtained on the directed graph. Based on the order in which the reversing nodes appear on the reference path, the reference path is divided into a first segment, a middle segment, and a last segment. The first segment extends from the current position of the four-way vehicle to the first reversing node, the middle segment is located between two adjacent reversing nodes, and the last segment extends from the last reversing node to the unloading point.

5. The four-way vehicle path dynamic binding method based on sorting level according to claim 1, characterized in that, The process of establishing a binding mapping between sorting levels and candidate road segments includes: Create a row for each sorting level in tabular form. The row fields include sorting level number, candidate route sequence, start and end vertex numbers and route length of each route in the candidate route sequence, corresponding unloading point number, and binding status field. The binding status field of each sorting level in the binding mapping is initially set to pending binding, and a copy of the binding mapping is stored on the scheduling server.

6. The four-way vehicle path dynamic binding method based on sorting level according to claim 5, characterized in that, The binding mapping includes: As the four-way vehicle passes a reversal node, the remaining candidate road segments for each sorting level are re-analyzed according to the position after passing, and the candidate road segment sequence is updated. During the re-analysis, the candidate road segments that have been passed are removed, and the temporary closed track segments issued by the scheduling server are removed from the self-directed graph. The candidate road segment sequence contains common road segments formed by road segment intervals traversed by the reference paths of each sorting level.

7. The four-way vehicle path dynamic binding method based on sorting level according to claim 1, characterized in that, Whether the highest confidence level in the judgment level determination sequence reaches the set binding threshold includes: The determination is triggered when the distance between the four-way vehicle and the forward reversing node is reduced to the advance determination distance. The advance determination distance is determined by the running speed of the four-way vehicle, the link delay time from the initiation of the determination to the completion of the action preparation of the reversing wheel group, and the safety margin distance. The latest record is read from the sorting level determination sequence. The maximum value of the cumulative credibility of each sorting level in the record is compared with the binding threshold. A unified binding threshold is adopted for each sorting level.

8. The four-way vehicle path dynamic binding method based on sorting level according to claim 1, characterized in that, The process of binding the sorting level with the highest credibility to the candidate road segment after the reversal node includes: Rewrite the binding status field of the sorting level with the highest confidence in the binding mapping as "bound", mark the row containing the remaining sorting levels as "invalid", and use the candidate road segment sequence of the bound sorting levels as the running path of the four-way vehicle from the current reversing node to the corresponding unloading point. After binding is completed, the level determination sequence continues to be updated as images are acquired, and the running path is not rewritten with subsequent fluctuations in credibility.

9. The method for dynamic binding of four-way vehicle paths based on sorting level according to claim 1, characterized in that, The conversion into reversing and unloading control commands includes: Each reversing node in the running path is generated one by one to generate a reversing control command. The reversing control command includes the reversing node vertex number, the entry direction, the exit direction, and the action type of the reversing wheel set. Generate unloading control instructions, which include the target unloading point vertex number and unloading trigger conditions. The unloading trigger conditions are that the four-way vehicle's arrival detection signal at the target unloading point is valid. Write the reversing control instructions and unloading control instructions into the vehicle-side control queue in the execution order.

10. The four-way vehicle path dynamic binding method based on sorting level according to claim 9, characterized in that, The vehicle-side control queue includes: During the journey, the four-way vehicle is taken out from the vehicle-side control queue one by one in the execution order and executes the reversing control command and the unloading control command. After each instruction is executed, the wireless communication module sends an execution confirmation back to the dispatch server; if the instruction execution times out and no execution confirmation is received, the four-way vehicle performs parking protection and reports to the dispatch server.

11. The method for dynamic binding of four-way vehicle paths based on sorting level according to claim 1, characterized in that, The sorting levels where the elimination confidence level remains below the elimination threshold include: If the number of frames in the pending binding state and the cumulative confidence obtained by the confidence level obtained by the frame-by-frame output is continuously lower than the elimination threshold reaches the set number of frames for the sorting level, the corresponding row will be set to invalid in the binding mapping and the corresponding unloading point will be removed from the scope of subsequent judgment. The re-inspection level used to collect abnormal materials is treated as a fallback level and will not be eliminated. Sorting levels that are set to invalid will not be restored in the current transportation task.

12. The four-way vehicle path dynamic binding method based on sorting level according to claim 1, characterized in that, The selected travel directions that are accessible to all non-eliminated destinations include: For each possible outgoing direction of the current reversing node, determine whether there is still a feasible path on the directed graph after removing the temporary closed track segment; All unload points corresponding to the sorting grades that have not been phased out have feasible outbound routes.

13. The four-way vehicle path dynamic binding method based on sorting level according to claim 12, characterized in that, The direction of departure includes: When there is more than one retained departure direction, select the departure direction with the largest sum of remaining reversing nodes on the reference path of each non-eliminated sorting level; When the total number of remaining reversing nodes is the same, the departure direction with the smallest total reference path length to the unloading point corresponding to each non-eliminated sorting level is selected.

14. The method for dynamic binding of four-way vehicle paths based on sorting level according to claim 1, characterized in that, The step of extending the binding to the next reversing node includes: During the extension period, the four-way vehicle continues to travel and does not turn back, and the level determination sequence continues to be updated with subsequent image acquisition; When a four-way vehicle arrives at a reversing node in the outgoing direction where there is no unloading point that can be simultaneously reached by all non-eliminated sorting grades, the process will not be delayed. Instead, the sorting grade with the highest current credibility will be used for binding. If the only sorting grade left to be bound is the re-inspection grade, the re-inspection grade will be used as the binding grade.

15. The method for dynamic binding of four-way vehicle paths based on sorting level according to claim 1, characterized in that, The four-way drive vehicle completes transportation and sorting along the designated road section, including: After the four-way vehicle arrives at the target unloading point and the arrival detection signal is valid, the vehicle-end control queue releases the unloading control command, the transfer rollers of the loading platform rotate forward and transfer the material frame to the roller conveyor line of the target unloading point; Write the material box's identification information, the bound sorting level, the unloading point number, and the completion time into the task record, and archive the task record along with the traceability information.

16. A four-way vehicle path dynamic binding system based on sorting level, used to implement the four-way vehicle path dynamic binding method based on sorting level as described in any one of claims 1 to 15, characterized in that, include: Sequence generation module, mapping determination module, and binding control module; The sequence generation module is used to collect images of the surface of edible fungi in the four-way vehicle transporting the edible fungi material frame and read the variety and harvesting traceability information of the material frame. The image visual features and traceability information are fused and input into a lightweight recognition model, and the sorting level and credibility are output frame by frame to form a level determination sequence that is updated progressively with the transportation process. The mapping determination module is used to resolve the unloading destination corresponding to each sorting level into sequentially connected candidate road segments according to the order of each reversing node on the driving path, establish a binding mapping between sorting level and candidate road segments, and determine whether the highest confidence in the level determination sequence reaches the set binding threshold before the four-way vehicle arrives at each reversing node. The binding control module is used to bind the sorting level with the highest credibility to the candidate road segment after the reversing node when the highest credibility reaches the binding threshold, and convert it into a reversing and unloading control command; if the binding threshold is not reached, the sorting level with credibility continuously lower than the elimination threshold is eliminated, the travel direction that can be reached by all non-eliminated destinations is selected, the binding is extended to the next reversing node, and the four-way vehicle is driven to complete the transportation and sorting along the bound road segment.

Images