Multi-split control method and system for logistics packet supply station based on vision and cloud collaboration

Abstract

The invention discloses a logistics package supply station multi-split control method and system based on vision and cloud collaboration, and the method comprises the steps: independently collecting package bar codes and image information by each package supply station, and transmitting task data to a server after local verification; the server performs priority ranking and resource allocation on the tasks according to a dynamic scoring mechanism; reasoning the batch processing tensor by using a detection model, and outputting information of packages in all the images; performing multi-dimensional rule filtering on an original result output by the detection model, and starting a processing flow for a triggered abnormal condition; and the packet supply station verifies and executes the control instruction, and feeds back confirmation information to the server after completing the packet loading action. Through a multi-split intensive processing mode, the visual computing units originally needing to be independently deployed on each packet supply platform are concentrated to the server for unified processing, the hardware cost is greatly reduced, meanwhile, only the packet supply platforms need to be added during expansion, and computing resources do not need to be repeatedly configured.

Description

Technical Field

[0001] This invention relates to the field of logistics automation technology, and more specifically, to a method and system for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration. Background Technology

[0002] Currently, the package feeding solutions used in logistics systems mainly fall into two categories: detection solutions based on photoelectric sensors (light curtains) and recognition solutions based on local computer vision. The former achieves automatic detection and control of package conveying by installing one or more sets of light curtain sensors at the entrance of the package feeding station, while the latter uses cameras and other visual sensors to identify information such as the position, size, and orientation of the package, thereby achieving automated and intelligent package feeding. However, both solutions are susceptible to environmental influences during package feeding. For example, strong light or dusty conditions can affect the normal operation of the sensors. Furthermore, light curtain sensors cannot effectively detect transparent or highly reflective packages, impacting detection results. For irregularly shaped or significantly sized packages, customized solutions are required.

[0003] At the same time, neither of these two solutions effectively prevents improper handling of packages by the supply personnel, which can easily lead to packages being lost or damaged during the process from the supply station to the unloading compartment, resulting in economic losses. It also increases the probability of hardware problems at the supply station, reducing customer satisfaction.

[0004] Therefore, a new solution is needed to address this problem. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a one-to-many control method and system for logistics parcel delivery stations based on vision and cloud collaboration.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration includes the following steps:

[0008] Step S1: Each parcel station independently collects the parcel barcode and image information, verifies it locally, and then transmits the task data to the server.

[0009] Step S2: The server receives and verifies task data from multiple task providers, prioritizes and allocates resources to the tasks according to the dynamic scoring mechanism, and forms a batch processing tensor.

[0010] Step S3: The server uses the detection model to infer the batch tensor and outputs the information of the package in all images at once.

[0011] Step S4: Perform multi-dimensional rule filtering on the raw results output by the detection model, start the processing flow for triggered abnormal situations, and send control commands to the corresponding package supply station.

[0012] Step S5: The package supplier verifies and executes the control command, completes the package loading action, and sends a confirmation message back to the server. The server then updates the task status.

[0013] Furthermore, step S1 includes the following steps:

[0014] Step S11: The parcel supply station reads the parcel barcode and performs local validity verification. If the barcode is valid, it triggers the acquisition of the original image of the parcel arriving at the parcel supply station area, and at the same time associates the original image with the corresponding verification code.

[0015] Step S12: Encapsulate the valid barcode, timestamp, image data with verification code, and the identifier of the corresponding supply station into a task data packet and send it to the server.

[0016] Furthermore, step S2 includes the following steps:

[0017] Step S21: The server receives and verifies the task data. Tasks that pass the verification are parsed into internal objects and enter the scheduling queue.

[0018] Step S22: Calculate a dynamic priority score for each internal object based on task urgency, waiting time, and estimated processing efficiency. The calculation formula is as follows:

[0019]

[0020] In the formula, Basic priority constant; , and These are weighting coefficients, corresponding to waiting time, urgency, and processing efficiency, respectively. This refers to the actual waiting time of the task in the queue. This is the maximum allowed waiting time threshold; This is an emergency mission indicator; The historical average processing time for tasks of the same package type; This is the maximum allowed processing time threshold for a single task;

[0021] Step S23: Periodically scan the scheduling queue based on dynamic priority scores. Sort all internal objects in descending order, and select image data from multiple tasks from the head of the sorting queue, stack them in the batch processing dimension to form a batch processing tensor.

[0022] Furthermore, step S3 includes the following steps:

[0023] Step S31: Use the full intersection-union loss supervised detection model for bounding box regression learning to obtain accurate bounding box localization. The calculation formula is as follows:

[0024]

[0025]

[0026] In the formula, This is the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. Center point of the prediction box Center point of the real frame The Euclidean distance between them; The diagonal length of the smallest closed region that can simultaneously contain both the predicted bounding box and the ground truth bounding box; These are the weighting coefficients used to balance the aspect ratio loss term; This is a parameter used to measure the aspect ratio similarity between the predicted bounding box and the ground truth bounding box; and These are the width and height of the prediction box, respectively; and These are the width and height of the actual bounding box, respectively;

[0027] Step S32: The detection model integrates a convolutional block attention module, which adaptively reconstructs the input feature map by sequentially calculating channel and spatial attention.

[0028] Furthermore, in step S32, the feature map reconstruction includes the following steps:

[0029] Step S321, calculate channel attention weights:

[0030]

[0031] In the formula, The feature map is input to the attention module of the convolutional block; and These represent global average pooling and global max pooling operations, respectively. For multilayer perceptrons with shared parameters; Use the Sigmoid activation function; This is the calculated channel attention weight vector;

[0032] Step S322, calculate spatial attention weights:

[0033]

[0034] In the formula, This is an intermediate feature map obtained through channel attention; For the core size is Convolution operations; This is a splicing operation along the channel dimension; The calculated spatial attention weight map;

[0035] Step S323: Reconstruct the feature map using attention weights:

[0036]

[0037] In the formula, This is the enhanced feature map output after dual attention modulation; This is for element-wise multiplication.

[0038] Furthermore, step S4 includes the following steps:

[0039] Step S41: The server performs compliance judgment on each raw result output by the detection model according to preset rules. When the rule filtering triggers an anomaly, hierarchical closed-loop processing is performed.

[0040] Step S42: The control commands obtained through filtering and rescheduling are bound to their corresponding tasks and sent to the corresponding package supply station.

[0041] Furthermore, in step S41, the compliance judgment includes baggage handling, out-of-field of view, and posture compliance judgment.

[0042] Furthermore, step S5 includes the following steps:

[0043] Step S51: The package station verifies and executes the control command to accurately place the package into the target cart slot.

[0044] Step S52: After execution is complete, the supply station sends an execution confirmation message to the server. The server updates the task status to complete and releases the resources it occupies.

[0045] This invention also provides a vision-based and cloud-based collaborative logistics parcel supply station control system, the system including a server and multiple parcel supply stations communicatively connected to the server, each parcel supply station comprising:

[0046] The data acquisition and verification module is used to acquire package barcodes and original images, and perform data verification locally;

[0047] The human-machine interaction module provides a local operating interface and audio-visual indicators, receives instructions from on-site personnel, and provides feedback on system status and alarm information.

[0048] The motion control module is used to receive and parse control commands from the server and drive the actuator to complete the package loading operation;

[0049] The status feedback module is used to send task execution confirmation messages to the server;

[0050] The server includes:

[0051] The communication and verification module is used to establish network connections with each package supply station and to receive and verify task data packets.

[0052] The task scheduling and resource management module is used to dynamically prioritize, sort, and allocate resources for all concurrent tasks.

[0053] Visual processing module: Utilizing the detection model optimized by the cross-multiplication-over-contraction loss function and the convolutional block attention module, high-performance parallel inference is performed on batch images to achieve package recognition and localization;

[0054] The rule filtering and anomaly handling module filters the raw results output by the detection model through preset rules and initiates hierarchical closed-loop processing for triggered anomalies.

[0055] The result routing module is used to accurately bind the processed control commands to the requesting packet supply station and send them to the corresponding packet supply station.

[0056] This invention also provides a vision-based and cloud-based collaborative logistics parcel delivery station multi-tasking control device, comprising:

[0057] Memory, used to store computer programs and data;

[0058] The processor is used to execute computer programs to implement the steps of the vision- and cloud-based collaborative logistics parcel delivery station one-to-many control method described above.

[0059] The beneficial effects of this invention are:

[0060] 1. This invention adopts a centralized processing mode of one-to-many, which centralizes the visual computing units that originally needed to be deployed independently at each packaging station to the server for unified processing, greatly reducing hardware costs. At the same time, when expanding, only the packaging station needs to be added, without the need to repeatedly configure computing resources.

[0061] 2. This invention adopts a dynamic priority task scheduling mechanism, which realizes intelligent sorting and batch processing reasoning under multi-task concurrency, and at the same time achieves the effect of rapid response to urgent tasks.

[0062] 3. In this invention, the detection model integrates the intersection-union loss and the convolutional block attention module, which can still achieve high-precision bounding box regression and feature focusing in real-world scenarios such as complex lighting, occlusion, and background interference, thereby improving the adaptability and recognition stability of the package.

[0063] 4. This invention monitors the package supply status through multi-dimensional rule filtering and automatically triggers a graded processing flow for abnormal situations, thereby improving the fault tolerance of the system.

[0064] 5. This invention automates the entire process from data acquisition, verification, and transmission to visual recognition, command issuance, action execution, and status feedback, reducing the need for manual intervention. Attached Figure Description

[0065] Figure 1 This is a flowchart of a one-to-many control method for a logistics parcel supply station based on vision and cloud collaboration in this embodiment.

[0066] Figure 2 This is a structural framework diagram of a one-to-many logistics parcel supply control system based on vision and cloud collaboration in this embodiment.

[0067] Figure 3 This is a structural framework diagram of a multi-device control device for logistics parcel delivery based on vision and cloud collaboration in this embodiment.

[0068] Reference numerals: Packing station 1, Data acquisition and verification module 11, Human-computer interaction module 12, Motion control module 13, Status feedback module 14, Server 2, Communication and verification module 21, Task scheduling and resource management module 22, Vision processing module 23, Rule filtering and exception handling module 24, Result routing module 25, Memory 3, Processor 4. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Example: A one-to-many control method for logistics parcel delivery stations based on vision and cloud collaboration. This method, through architectural innovation, centralizes the visual computing tasks of multiple parcel delivery stations 1 onto a single server 2 for processing, constructing a "one-to-many" intensive processing mode; simultaneously, it improves parcel recognition and positioning in complex scenarios through algorithm optimization, and utilizes intelligent scheduling to ensure concurrent execution of multiple tasks. Specifically, such as... Figure 1 As shown, it includes the following steps:

[0071] Step S1, Parallel Data Acquisition and Transmission of Multiple Packaging Stations 1: Each packaging station 1 independently collects the package barcode and image information, and after local verification, transmits the task data to server 2.

[0072] Step S1 specifically includes the following steps:

[0073] In step S11, the barcode scanner of the package supply station 1 reads the package barcode and performs local validity verification. If the barcode is valid, the industrial camera bound to the package supply station 1 is triggered to capture the original image of the package arriving at the package supply station 1 area. If the barcode is invalid, the barcode data is discarded, and a local prompt can be issued through the human-machine interaction module 12, and then the system continues to wait to read the next valid barcode.

[0074] For valid barcodes, to prevent data errors at the acquisition end, a cyclic redundancy check (CRC-32) is immediately performed on the binary data of the acquired original image to generate a check code and associate it with the original image.

[0075] Step S12: Encapsulate the valid barcode, timestamp, image data with verification code, and the unique identifier of the corresponding packet supply station 1 into a task data packet, and send it to the server 2 using the TCP protocol.

[0076] It should be noted that server 2 monitors multiple packet supply stations 1 in real time to determine whether there are any network connection requests sent to it. If there are, server 2 establishes a connection with the packet supply station 1 that initiated the network connection request and interacts with it; otherwise, it continues to monitor.

[0077] Further, in step S2, server 2 dynamically prioritizes task scheduling: server 2 receives and verifies task data from multiple packet supply stations 1, and prioritizes and allocates resources to tasks according to a dynamic scoring mechanism to achieve optimal system throughput and response time.

[0078] Step S2 specifically includes the following steps:

[0079] Step S21: After receiving the task data packet, server 2 verifies its TCP integrity and CRC checksum; the task that passes the verification is parsed as an internal object and enters the scheduling queue.

[0080] In step S22, to optimize overall throughput and response time, and to prevent tasks from getting stuck in long waits due to improper scheduling strategies, the scheduler of server 2 calculates a dynamic priority score for each internal object. This is to achieve a comprehensive optimization of task waiting time, urgency, and processing efficiency. The calculation formula is as follows:

[0081]

[0082] In the formula, This is a preset basic priority constant used to ensure that all tasks have a basic scheduling weight, avoiding division by zero or extremely small values. , and These are preset weighting coefficients, corresponding to waiting time, urgency, and processing efficiency, respectively, and can be dynamically adjusted according to operational strategies. =0.3, =0.5, =0.2; This refers to the actual waiting time of the task in the queue. The preset maximum allowable waiting time threshold can be set through historical data statistics; This is an emergency task indicator, set when submitting a task; 1 indicates emergency, and 0 indicates normal. The historical average processing time for tasks of the same package type is calculated in real time by the system. This is the preset maximum allowed processing time threshold for a single task, which can be set through historical data statistics.

[0083] The aforementioned scoring mechanism aims to achieve multi-objective optimized scheduling. The numerator integrates basic priority, normalized waiting time, urgency, and a processing efficiency factor (the shorter the historical average processing time, the greater the contribution of this factor). The denominator is the historical average processing time. Given similar numerators, tasks with shorter historical processing times (higher efficiency) receive a higher dynamic priority score, thus gaining priority in scheduling. This scoring mechanism effectively prevents task backlog and provides rapid responses to urgent tasks.

[0084] In step S23, the scheduler periodically (e.g., every 10 milliseconds) scans the scheduling queue, sorts all internal objects in descending order according to the dynamic priority score, selects multiple tasks from the head of the sorting queue, and stacks the image data of these tasks in the batch processing dimension to form a four-dimensional batch processing tensor (dimensions are batch size, number of channels, height and width).

[0085] By constructing a batch processing tensor, multiple preprocessed images are stacked in the batch dimension to achieve multi-task parallel inference.

[0086] Further, in step S3, package recognition and localization based on the improved detection model: server 2 uses the detection model to infer the batch tensor and outputs the precise bounding box, type and key point information of the packages in all images at once.

[0087] Step S3 specifically includes the following steps:

[0088] Step S31: Use the Complete Intersection over Union (CLOU) Loss supervised detection model for bounding box regression learning to obtain accurate bounding box localization. The calculation formula is as follows:

[0089]

[0090]

[0091] In the formula, The intersection-union ratio (IUU) of the predicted bounding box and the ground truth bounding box is calculated by dividing the area of ​​their intersection by the area of ​​their union. Center point of the prediction box Center point of the real frame The Euclidean distance between them; The diagonal length of the smallest closed region that can simultaneously contain both the predicted bounding box and the ground truth bounding box; These are the weighting coefficients used to balance the aspect ratio loss term, and are preset hyperparameters for model training; This is a parameter used to measure the aspect ratio similarity between the predicted bounding box and the ground truth bounding box; and These are the width and height of the prediction box, respectively; and These are the width and height of the actual bounding box, respectively.

[0092] By employing the cross-union loss function, the model can predict bounding boxes that better match the actual outline of the package, thus facilitating the robotic arm's grasping and positioning of the package.

[0093] Step S32: The detection model integrates a convolutional block attention module (CBAM) to adaptively reconstruct the input feature map by sequentially calculating channel and spatial attention.

[0094] Feature map reconstruction includes the following steps:

[0095] Step S321, calculate channel attention weights:

[0096]

[0097] In the formula, The feature map is input to the attention module of the convolutional block; and These represent global average pooling and global max pooling operations, respectively. For multilayer perceptrons with shared parameters; Use the Sigmoid activation function; This is the calculated channel attention weight vector.

[0098] First, by analyzing the input feature map Global average pooling and global max pooling operations are performed separately to aggregate global information for each channel in the spatial dimension. Then, the two pooling structures are input into a shared multilayer perceptron for nonlinear transformation and information interaction between channels. Finally, the outputs of the two multilayer perceptrons are summed and mapped through an activation function to generate channel attention weight vectors. .

[0099] Step S322, calculate spatial attention weights:

[0100]

[0101] In the formula, This is an intermediate feature map obtained through channel attention; For the core size is Convolution operations; This is a splicing operation along the channel dimension; This is the calculated spatial attention weight map.

[0102] First, analyze the intermediate feature maps respectively. Perform global average pooling and global max pooling operations to obtain two spatial feature maps; then, concatenate these two feature maps and feed them into a... The convolutional layers are used for fusion and feature extraction; finally, a spatial attention weight map is generated using the activation function. .

[0103] Step S323: Reconstruct the feature map using attention weights:

[0104]

[0105] In the formula, This is the enhanced feature map output after dual attention modulation; This is for element-wise multiplication.

[0106] First, the generated channel attention weights are compared with the original input feature map. Channel-wise element-wise multiplication is performed to achieve global scaling of different feature channels; then, the generated spatial attention weights are multiplied element-wise at spatial locations with the scaled feature map, finally outputting the reconstructed enhanced feature map. .

[0107] By employing a convolutional block attention module, the model automatically focuses on key feature channels and spatial regions in the image that are crucial for identification, while suppressing irrelevant background and noise interference. This improves the model's feature discrimination ability in scenarios with varying lighting, partial occlusion, and complex backgrounds. Specifically:

[0108] When faced with changes in lighting and reflective packaging materials, channel attention is used to suppress the responses of feature channels that are sensitive to light or contaminated by reflections, while enhancing the reliance on feature channels related to material texture and stable shape. For example, when strong light causes local overexposure of the package, the channel weights for color information in that area are reduced, while the channel weights for contour gradient information are increased. In the original image captured by the packaging station, the package is generally located in the center of the image, but there are distractions such as the conveyor belt in the background. Spatial attention can soften the background and highlight the package. By combining channel attention and spatial attention, the model's ability to extract features from the package and its robustness against interference in noisy environments are significantly improved.

[0109] Further, in step S4, multi-dimensional rule filtering and anomaly handling: multi-dimensional rule filtering is performed on the original results output by the detection model, the processing flow is started for triggered anomalies, and control commands are sent to the corresponding package supply station 1.

[0110] Step S4 specifically includes the following steps:

[0111] Step S41: Server 2 performs compliance checks on each raw result output by the detection model according to preset rules, determining whether it meets the conditions for uploading, including:

[0112] Packet Detection: To prevent efficiency from being affected by packages being stuck, a package detection mechanism is set up. The package detection mechanism is as follows: if packages at the same location are detected in consecutive images and the displacement of each package is less than a preset value, then the package is determined to be a stuck package.

[0113] Beyond-field detection: To ensure operational safety and compliance, the package must be within the permitted working area. Specifically, the coordinates of the center point of each detection frame are calculated, and it is determined whether the point falls within the pre-marked area. If the coordinates are outside the area, it is considered an invalid beyond-field detection.

[0114] Attitude compliance judgment: In order to adapt to automated grasping or placement, the attitude of the package must meet the requirements. The rule is: using the coordinates of the key points of the package output in step S3, calculate the attitude angle representing the orientation of the package. If the absolute value of the attitude angle exceeds the preset value, it is judged as an abnormal attitude.

[0115] When rule filtering triggers an exception, a tiered closed-loop process is executed, including:

[0116] Real-time alarm: After determining an anomaly, server 2 immediately sends an alarm command to the corresponding supply station 1 for the abnormal task. At the same time, server 2 marks the task status as "awaiting manual handling" and starts a counter. Upon receiving the command, supply station 1 triggers its local audible and visual alarm device, prompting staff to handle the situation.

[0117] Manual intervention: Staff handle abnormal situations, and after handling, staff send a "retry request" signal to server 2 through the package supply station 1.

[0118] Task rescheduling: If server 2 receives a "retry request" before the timer expires, it immediately clears the alarm status and waiting for manual processing mark of the task, then re-inserts the task into the task scheduling queue, recalculates the priority according to the scheduling algorithm, and enters a new round of processing.

[0119] If server 2 does not receive a "retry request" before the timer expires, it is determined that manual intervention has failed or timed out. The task status is automatically marked as "failed", a detailed exception log is recorded, and all resources it occupies are released.

[0120] Step S42: The control commands obtained through filtering and rescheduling are bound to their corresponding tasks, encapsulated into a result data packet, and a CRC checksum is attached. Then, the packet is sent to the corresponding packet supply station 1.

[0121] Furthermore, in step S5, the actions of the packet supply platform 1 are synchronized with the status of server 2:

[0122] After the package feeding station 1 completes the verification, the drive actuator (such as a robotic arm) completes the package feeding action and sends a confirmation message to the server 2. The server 2 then updates the task status.

[0123] Step S5 specifically includes the following steps:

[0124] Step S51: After receiving the result data packet, the package supply station 1 verifies its CRC check code. After the verification is successful, the drive actuator accurately places the package into the target cart slot.

[0125] Step S52: After execution, the supply station 1 sends a confirmation message with a verification code to the server 2. After the server 2 verifies the confirmation message, it updates the status of the corresponding task to "completed" and releases all the resources it occupies.

[0126] This embodiment also provides a one-to-many control system for a logistics parcel delivery station based on vision and cloud collaboration, such as... Figure 2 As shown, the system includes a server 2 and multiple supply stations 1 that are connected to the server 2 via a switch; each supply station 1 includes a data acquisition and verification module 11, a human-computer interaction module 12, a motion control module 13, and a status feedback module 14.

[0127] The data acquisition and verification module 11 is responsible for acquiring the package barcode and original image, and performing data verification locally to ensure the integrity and accuracy of the acquired data.

[0128] The human-machine interaction module 12 provides a local operation interface and audio-visual indicators to receive instructions from on-site personnel and to provide feedback on system status and alarm information to them.

[0129] The motion control module 13 is used to receive and execute control commands from the server 2 to complete package loading operations such as package grabbing, moving and placing.

[0130] After the task is completed, the status feedback module 14 sends a task execution confirmation message to the server 2, providing a basis for the server 2 to release resources.

[0131] Server 2 includes a communication and verification module 21, a task scheduling and resource management module 22, a visual processing module 23, a rule filtering and exception handling module 24, and a result routing module 25.

[0132] Among them, the communication and verification module 21 is responsible for establishing network connections with each packet supply station 1 and receiving and verifying task data packets.

[0133] The task scheduling and resource management module 22 is responsible for dynamically prioritizing, sorting, and allocating resources for all concurrent tasks. Its core function is to maximize the overall system throughput through scheduling algorithms and ensure that critical tasks are responded to in a timely manner.

[0134] The visual processing module 23 is equipped with a detection model optimized by the cross-union loss function and the convolutional block attention module. This model performs high-performance parallel inference on batch images to achieve package recognition and localization.

[0135] The rule filtering and anomaly handling module 24 filters the original results output by the detection model through preset rules and initiates hierarchical closed-loop processing for triggered anomalies.

[0136] The routing module 25 is responsible for accurately binding the processed control command to the requesting packet supply station 1, and sending the command precisely to the corresponding packet supply station 1.

[0137] This embodiment also provides a vision-based and cloud-based collaborative logistics parcel delivery station 1 multi-device control device, such as... Figure 3 As shown, it includes a memory 3 and a processor 4; wherein, the memory 3 is used to store computer programs and data; the processor 4 is used to implement the steps of the above-mentioned vision and cloud-based collaborative logistics parcel station 1 one-to-many control method when executing the computer program.

[0138] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration, characterized in that, Includes the following steps: Step S1: Each parcel station (1) independently collects parcel barcode and image information, and after local verification, transmits the task data to the server (2). Step S2, the server (2) receives and verifies the task data from multiple package supply stations (1), prioritizes and allocates resources for the tasks according to the dynamic scoring mechanism, and forms a batch processing tensor; Step S3, the server (2) uses the detection model to infer the batch tensor and outputs the information of the package in all images at once; Step S4: Perform multi-dimensional rule filtering on the original results output by the detection model, start the processing flow for triggered abnormal situations, and send control instructions to the corresponding package supply station (1); In step S5, the package supply station (1) verifies and executes the control command, and after completing the package loading action, it sends a confirmation message to the server (2). The server (2) then updates the task status.

2. The method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration as described in claim 1, characterized in that, Step S1 includes the following steps: Step S11: The package supply station (1) reads the package barcode and performs local validity verification. If the barcode is valid, it triggers the acquisition of the original image of the package arriving at the package supply station (1) area, and the original image is associated with the corresponding verification code. Step S12: Encapsulate the valid barcode, timestamp, image data with verification code, and the identifier of the corresponding supply station (1) into a task data packet and send it to the server (2).

3. The method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration as described in claim 1, characterized in that, Step S2 includes the following steps: Step S21, the server (2) receives and verifies the task data. The tasks that pass the verification are parsed into internal objects and enter the scheduling queue. Step S22: Calculate a dynamic priority score for each internal object based on task urgency, waiting time, and estimated processing efficiency. The calculation formula is as follows: In the formula, Basic priority constant; , and These are weighting coefficients, corresponding to waiting time, urgency, and processing efficiency, respectively. This refers to the actual waiting time of the task in the queue. This is the maximum allowed waiting time threshold; This is an emergency mission indicator; The historical average processing time for tasks of the same package type; This is the maximum allowed processing time threshold for a single task; Step S23: Periodically scan the scheduling queue based on dynamic priority scores. Sort all internal objects in descending order, and select image data from multiple tasks from the head of the sorting queue, stack them in the batch processing dimension to form a batch processing tensor.

4. The method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration as described in claim 1, characterized in that, Step S3 includes the following steps: Step S31: Use the full intersection-union loss supervised detection model for bounding box regression learning to obtain accurate bounding box localization. The calculation formula is as follows: In the formula, This is the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. Center point of the prediction box Center point of the real frame The Euclidean distance between them; The diagonal length of the smallest closed region that can simultaneously contain both the predicted bounding box and the ground truth bounding box; These are the weighting coefficients used to balance the aspect ratio loss term; This is a parameter used to measure the aspect ratio similarity between the predicted bounding box and the ground truth bounding box; and These are the width and height of the prediction box, respectively; and These are the width and height of the actual bounding box, respectively; Step S32: The detection model integrates a convolutional block attention module, which adaptively reconstructs the input feature map by sequentially calculating channel and spatial attention.

5. The method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration according to claim 4, characterized in that, In step S32, the feature map reconstruction includes the following steps: Step S321, calculate channel attention weights: In the formula, The feature map is input to the attention module of the convolutional block; and These represent global average pooling and global max pooling operations, respectively. For multilayer perceptrons with shared parameters; Use the Sigmoid activation function; This is the calculated channel attention weight vector; Step S322, calculate spatial attention weights: In the formula, This is an intermediate feature map obtained through channel attention; For the core size is Convolution operations; This is a splicing operation along the channel dimension; The calculated spatial attention weight map; Step S323: Reconstruct the feature map using attention weights: In the formula, This is the enhanced feature map output after dual attention modulation; This is for element-wise multiplication.

6. The method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration according to claim 1, characterized in that, Step S4 includes the following steps: Step S41, the server (2) performs compliance judgment on each original result output by the detection model according to preset rules. When the rule filtering triggers an anomaly, it performs hierarchical closed-loop processing. Step S42: The control commands obtained through filtering and rescheduling are bound to the corresponding tasks and sent to the corresponding package supply station (1).

7. The method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration according to claim 6, characterized in that, In step S41, the compliance judgment includes baggage handling, out-of-field of view, and posture compliance judgment.

8. The method for controlling multiple logistics parcel delivery stations based on vision and cloud collaboration according to claim 1, characterized in that, Step S5 includes the following steps: Step S51, the package supply station (1) verifies and executes the control command to accurately place the package into the target cart slot; Step S52: After execution is completed, the supply station (1) sends an execution confirmation message to the server (2), the server (2) updates the status of the task to complete, and releases the resources it occupies.

9. A vision- and cloud-based collaborative logistics parcel delivery station multi-tasking control system for implementing the method of claim 1, characterized in that, The system includes a server (2) and multiple package supply stations (1) communicatively connected to the server (2), wherein each package supply station (1) includes: The data acquisition and verification module (11) is used to acquire the package barcode and the original image, and to perform data verification locally; The human-machine interaction module (12) is used to provide a local operation interface and sound and light indication, receive instructions from on-site personnel and provide feedback on system status and alarm information to personnel; The motion control module (13) is used to receive and parse control commands from the server (2) and drive the actuator to complete the package loading operation; The status feedback module (14) is used to send a task execution confirmation message to the server (2); The server (2) includes: The communication and verification module (21) is used to establish a network connection with each package supply station (1) and to receive and verify task data packets; The task scheduling and resource management module (22) is used to dynamically prioritize, sort, and allocate resources for all concurrent tasks; Visual processing module (23): Utilizes the detection model optimized by the cross-multiplex loss function and the convolutional block attention module to perform high-performance parallel inference on batch images, thereby realizing the identification and localization of packages; The rule filtering and anomaly handling module (24) filters the original results output by the detection model through preset rules and initiates hierarchical closed-loop processing for triggered anomalies. The result routing module (25) is used to accurately bind the processed control command to the requesting packet supply station (1) and send it to the corresponding packet supply station (1).

10. A multi-device control device for a logistics parcel supply station based on vision and cloud collaboration, characterized in that, include: Memory (3) is used to store computer programs and data; The processor (4) is used to implement the steps of the vision-based and cloud-based collaborative logistics parcel station (1) one-to-many control method as described in any one of claims 1-8 when executing a computer program.

Images