Multi-mode sensing and edge computing cooperative control method and system for intelligent gate

Abstract

The invention relates to a multi-modal sensing and edge computing cooperative control method and system for an intelligent gate. The method comprises the following steps that a gate edge node schedules a multi-modal sensing unit to perform cooperative data acquisition; dynamically determining a multi-sensor cooperation strategy according to the environment and the target, carrying out confidence fusion, and generating a comprehensive sensing result containing vehicle identity, container identity and container condition initial judgment information; according to the service type, the data complexity and the real-time load of the gate edge node, partial or all sensing data and analysis tasks are distributed to a station edge server or a cloud server for processing through a dynamic task unloading strategy; and receiving a processing result, making a decision in combination with a local sensing result, and generating a gate control instruction. According to the invention, all-weather, high-precision and automatic intelligent management and control of gate operation are realized through multi-modal sensing fusion and cloud side end cooperative computing, and the passing efficiency, the identification reliability, the safety and the operation economy are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and logistics management technology, and in particular to a multimodal sensing and edge computing collaborative control method and system for intelligent gates. Background Technology

[0002] Container ports, logistics parks, and yard gates serve as crucial links between land and sea transport, as well as warehousing and transportation. Their throughput efficiency and level of intelligent management directly impact the overall operational efficiency of the logistics chain. With the growth of global trade volume and the increasing demands for timeliness in the supply chain, achieving automation, unmanned operation, and intelligent management of gate operations has become an inevitable trend in the industry.

[0003] Traditional gate management relies primarily on manual verification of documents and inspection of vehicles and containers, which is inefficient and prone to errors. To improve efficiency, the industry has gradually introduced automation technologies. For example,

[0004] Existing technology 1: For example, Chinese invention patent application CN114330605A discloses a management method and system for vehicle passage gates. This method uses an electromagnetic induction coil to trigger a reader-controlled microwave antenna to read information from vehicle-mounted electronic tags and transmits the information to a control computer for comparison to control the gate. This solution achieves preliminary unmanned clearance, but its sensing method is singular, relying entirely on pre-installed electronic tags. It cannot identify untagged vehicles or the physical condition of containers (such as damage), and the system is prone to failure in inclement weather or when tags are damaged, lacking redundant sensing and verification mechanisms.

[0005] To further enhance recognition capabilities, existing technology two, such as the Chinese invention patent application CN115497057A, discloses an intelligent gate control system and method for container yards. This system acquires license plate and container number images through a front-end acquisition module, which are then identified by a gate control module (usually an industrial control computer) and integrated with the business system. Finally, the system completes the release or toll collection through a human-machine interaction module. This solution utilizes machine vision technology to enhance the dimensions of information acquisition. However, its system architecture is mostly centralized, with all recognition, comparison, and decision-making logic concentrated on the local industrial control computer or back-end server at the gate. This architecture has significant bottlenecks: First, high-resolution image processing and complex AI model inference require large amounts of computation, placing high demands on the performance of local computing units, resulting in high costs and difficult upgrades; second, when facing complex tasks such as multi-lane concurrency, high traffic volume, or the need to run high-precision damage detection, centralized nodes easily become performance bottlenecks, leading to processing delays and vehicle congestion; third, the system lacks the ability to dynamically schedule computing resources based on real-time load, network conditions, and task urgency.

[0006] In addition, there are other technical solutions: for example, Chinese invention patent CN118262302B discloses a 5G smart barrier gate management method and system based on binocular recognition. This approach attempts to manage barrier gates from a more macroscopic traffic flow perspective, utilizing binocular vision and other technologies to analyze road segment traffic flow and dynamically adjust barrier gate modes. While these solutions focus on regional traffic flow scheduling, they lack in-depth optimization for multi-dimensional, high-precision status perception and real-time business processing of individual passing vehicles and their carrying containers. In particular, they lack automated and intelligent detection capabilities for the critical inspection item of container damage.

[0007] In summary, existing smart gate technologies have the following limitations:

[0008] 1. Limited perception dimensions and insufficient collaboration: It relies heavily on a single or a few types of sensors and fails to fully utilize the advantages of multimodal sensors such as vision, radar, RFID, and inductive loops for collaborative complementarity and data fusion. The reliability, accuracy, and information completeness of identification in complex environments need to be improved.

[0009] 2. Bottlenecks exist in centralized computing architecture: Data processing and analysis tasks are concentrated on a single node, making it difficult to balance the relationship between computing real-time performance, accuracy and system cost, scalability, and unable to efficiently handle peak loads and complex computing tasks.

[0010] 3. Lack of real-time adaptive decision-making capability: The system usually adopts a fixed process and cannot dynamically adjust the perception strategy, calculation task allocation (unloading) and control commands according to environmental conditions (such as lighting and weather), vehicle behavior, business type and its own load status.

[0011] 4. Lack of in-depth container condition detection: Most solutions only focus on identification (license plate, container number), lacking a mechanism for real-time, automatic, preliminary screening of container condition (especially damage) and efficient linkage with detailed back-end inspection, making it difficult to meet the refined management needs of customs, insurance and other links for cargo condition monitoring.

[0012] Therefore, there is an urgent need for a new intelligent gate management method and system that can integrate multimodal perception, possess edge collaborative computing capabilities, and enable dynamic intelligent decision-making and control, in order to break through the above-mentioned technical bottlenecks and achieve a higher degree of automation, intelligence, and reliability in operation. Summary of the Invention

[0013] The present invention aims to address the shortcomings of the prior art by providing a multimodal sensing and edge computing collaborative control method and system for intelligent gates.

[0014] To achieve the above objectives, this invention employs the following technical solution: a multimodal sensing and edge computing collaborative control method for intelligent gates, applied to a system including gate edge nodes, site edge servers, and cloud servers, comprising the following steps:

[0015] S1: When a vehicle enters the gate's sensing area, the multimodal sensing units deployed at the gate's edge nodes coordinate data collection. The multimodal sensing units include at least a visual sensor, an RFID reader, a radar sensor, and a ground loop.

[0016] S2: The gate edge node dynamically determines the coordination strategy between the visual sensor, RFID reader and radar sensor based on the environmental conditions and the sensing target, and fuses the data collected by each sensor in real time according to the confidence level to generate a comprehensive sensing result that includes vehicle identity, container identity and preliminary container condition information.

[0017] S3: Based on the business type, data complexity, and real-time load of the gate edge nodes according to the comprehensive perception results, some or all of the perception data and analysis tasks are distributed to the site edge server or cloud server for processing through a preset dynamic task offloading strategy.

[0018] S4: Receives the processing results from the site edge server or cloud server, combines them with the locally generated comprehensive perception results, and makes decisions by the gate edge node or the cooperating site edge server to generate gate control commands. The control commands are used to control the operation of the barrier gate and human-machine interaction terminal.

[0019] The initial condition assessment information of the container is obtained by real-time analysis of the visual data stream by a lightweight damage detection model deployed at the edge node of the gate. If the initial assessment result indicates that there is an anomaly, a high-priority data upload and detailed analysis process is triggered.

[0020] Specifically, the cooperative strategy in step S2 includes:

[0021] Based on information on illumination, weather, and visibility, the confidence weights of visual sensors and radar sensors in data fusion are dynamically adjusted.

[0022] Based on the reading status of the RFID reader, the ground loop and visual sensor are dynamically switched as the initial sensing trigger source.

[0023] When different sensors produce conflicting identification results for the same target, an arbitration function, which is performed by a radar sensor or a designated vision sensor, is activated for verification, and the confidence weights of each sensor are dynamically updated based on historical data.

[0024] Specifically, the dynamic task unloading strategy in step S3 includes:

[0025] Based on the comprehensive perception results, perform intent recognition to predict the vehicle's passage or business intent;

[0026] If the intention is for regular fast passage and the load on the gate edge node is below the threshold, then all processing and control will be completed by the gate edge node.

[0027] If a complex business intent is identified or an abnormal container condition is detected, the high-resolution visual data and related information are packaged and uploaded to the terminal edge server via a high-speed communication link for detailed inspection of container damage and verification of business documents.

[0028] If the site edge server is overloaded or the model version needs to be updated, the task will be further offloaded to the cloud server.

[0029] Specifically, the generation and detailed analysis of initial container condition information are carried out in a coordinated manner, including:

[0030] The lightweight damage detection model performs real-time localization and initial screening of key parts of containers in video streams, and outputs the probability and region of anomalies.

[0031] When the initial screening abnormality probability exceeds the first threshold, the gate edge node controls the supplementary lighting equipment and multi-angle vision sensors to capture a high-definition multi-view image set of the target area.

[0032] The high-definition multi-view image set is bound with the vehicle and container identity information to form a task package to be inspected in detail, and sent to the site edge server according to the dynamic task unloading strategy.

[0033] The edge server at the site runs a high-precision damage identification model, processes the task package, generates a standardized report containing quantified damage type, location, and size, and feeds it back to the gate control system and business management platform.

[0034] Specifically, the method also includes:

[0035] S5: Abnormal Behavior Collaborative Processing Mechanism: The radar sensor and vision sensor work together to monitor vehicle behavior trajectory in real time; when following too closely, abnormal lingering, or driving in the wrong direction is detected, the calculation priority of the gate edge node is immediately increased and the local behavior analysis model is started; if it is determined to be a security threat, the gate edge node directly generates an alarm and sends an immediate control command to the barrier gate and the audible and visual alarm equipment, while the event snapshot and data are uploaded asynchronously.

[0036] Specifically, the method also includes:

[0037] S6: Predictive-based energy efficiency management strategy: Based on historical traffic data and real-time sensing information, predict the traffic flow status in future periods; during the predicted low traffic periods, the non-critical sensing units of the gate edge node control part enter a low-power sleep mode and extend the sensing cycle; when the sensing units are triggered again or the predicted peak period arrives, all sensing units are automatically woken up and restored to normal working mode.

[0038] A multimodal sensing and edge computing collaborative control system for intelligent gates, used to realize a method for multimodal sensing and edge computing collaborative control of intelligent gates, the system comprising:

[0039] The multimodal sensing unit, deployed at the gate, includes at least a visual sensor for visual information acquisition, an RFID reader for identification, a radar sensor for ranging and behavior perception, and a ground loop for sensing vehicle signals.

[0040] The gate edge computing node is deployed locally at the gate and connected to the multimodal sensing unit. It is equipped with a first processing module and a first communication module for performing sensing scheduling, data fusion, lightweight model inference, real-time control, and task offloading decisions.

[0041] The site edge server is deployed in the site-level network and is equipped with a second processing module and a second communication module. It is used to receive and process complex computing tasks offloaded from the gate edge computing node and run high-precision analysis models.

[0042] The cloud server is connected to the site edge server and the gate edge computing node network for big data analysis, model training and global system management.

[0043] The gate execution and interaction unit, including the barrier gate, information display screen, and voice broadcaster, is controlled by the gate edge computing node or the site edge server.

[0044] Specifically, the vision sensors include:

[0045] A binocular vision camera deployed at the top of the gate is used to collect multi-view images of the container and assist in generating depth information;

[0046] High-resolution area scan cameras deployed on the side are used to capture high-definition images of the sides of the container.

[0047] The gate edge computing node performs time synchronization and spatial alignment of image data collected from different vision sensors through the first processing module.

[0048] Specifically, the gateway edge computing node also includes a local consensus module, and the site edge server includes a corresponding remote consensus module. The local consensus module is used to generate an event declaration containing a data hash value when a critical event or data anomaly is detected. The remote consensus module is used to verify the received event declaration and compare it with the declarations from other gateway edge computing nodes for consensus confirmation. The confirmed data is recorded in a distributed ledger.

[0049] In particular, the system also includes an environmental perception module that integrates temperature, humidity, and illuminance sensors. Its output signal is connected to the first processing module of the gate edge computing node as one of the input parameters for dynamically determining the collaborative strategy.

[0050] The beneficial effects of this invention are:

[0051] 1. This invention completely changes the reliance on a single sensor by constructing a multimodal sensing unit composed of a visual sensor, an RFID reader / writer, a radar sensor, and a ground loop coil, and by implementing a dynamic collaborative strategy and confidence fusion based on gate edge nodes. First, the complementarity and fusion of multi-source information (visual features, RFID identification, radar trajectory) makes the perception information of vehicle identity, container identity, and preliminary container condition more comprehensive and accurate, significantly reducing the false recognition rate and the missed recognition rate. Second, the system can dynamically adjust the fusion weights according to environmental conditions such as lighting and weather, and can automatically switch to visual or ground loop triggering when RFID fails. It also initiates an arbitration mechanism when sensor results conflict, greatly enhancing the system's adaptability and robustness in complex and harsh environments, overcoming the shortcomings of poor environmental adaptability and insufficient reliability in existing technologies.

[0052] 2. This invention innovatively constructs a three-tiered collaborative computing architecture: "gate edge node - station edge server - cloud," supplemented by a dynamic task offloading strategy. On one hand, lightweight real-time sensing, fusion, and control tasks are retained at the gate edge node, ensuring ultra-low latency real-time performance of core functions such as gate control and abnormal behavior response. On the other hand, through task offloading, computationally intensive tasks such as high-resolution image analysis and complex model inference (e.g., detailed damage inspection) are flexibly allocated to the more powerful station edge server or the cloud, achieving on-demand and efficient utilization of computing resources and avoiding performance bottlenecks and vehicle congestion problems inherent in centralized computing architectures. This architecture can be flexibly expanded according to business growth, exhibiting excellent overall system elasticity and scalability.

[0053] 3. The dynamic task unloading strategy and abnormal behavior collaborative processing mechanism of this invention endow the system with a high degree of intelligent decision-making and proactive control capabilities. The system can intelligently decide the execution location of tasks based on business intent, data complexity, and its own load, optimizing the overall processing flow. More importantly, through radar and visual collaborative monitoring, it can proactively detect safety threats such as following too closely or abnormal delays in real time, and immediately initiate local high-priority responses (such as alarms and gate control), achieving a leap from passive identification to proactive safety prevention and control, significantly improving the operational safety and reliability of the gate area.

[0054] 4. This invention fills a gap in existing technologies for automated detection of the physical condition of containers. By deploying a lightweight damage detection model at edge nodes, it achieves real-time, automated initial screening of container conditions. If an anomaly is detected during the initial screening, the system automatically triggers high-definition image acquisition, task packaging, and, relying on a collaborative computing architecture, quickly initiates high-precision damage identification in the background, generating a standardized report. This highly efficient "front-end initial screening - back-end detailed inspection" process achieves, for the first time, intelligent and streamlined inspection of container damage at the gate, providing refined and visualized key data support for customs supervision, insurance claims, and terminal management, significantly improving the precision of logistics management.

[0055] 5. The predictive energy efficiency management strategy introduced in this invention enables the system to predict traffic flow based on historical and real-time data, and intelligently adjust the power consumption and operating mode of the sensing units during low-traffic periods. This predictive energy-saving mechanism effectively reduces the overall energy consumption of the system during long-term operation. Simultaneously, the cloud server is responsible for the global training and updating of the model, and can uniformly distribute it to the edge, achieving continuous algorithm optimization and intelligent operation and maintenance. In summary, this invention improves performance and functionality while optimizing the long-term operation and maintenance costs of the system through intelligent scheduling and energy efficiency management, resulting in significant overall benefits.

[0056] In summary, this invention, through the organic integration of core technologies such as multimodal perception collaboration, edge computing collaboration, and dynamic intelligent decision-making and control, systematically solves the limitations of existing intelligent gates in terms of perception capabilities, computing architecture, decision intelligence, and deep detection. It achieves a comprehensive improvement in gate operation efficiency, reliability, security, intelligence level, and economy, demonstrating outstanding technological innovation and broad industrial application prospects. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 This is a structural block diagram of the system of the present invention;

[0059] The following will describe in detail, with reference to the accompanying drawings, embodiments of the present invention. Detailed Implementation

[0060] The present invention will be further described below with reference to embodiments:

[0061] like Figure 1 As shown, the multimodal sensing and edge computing collaborative control method for intelligent gates is applied to a system including gate edge nodes, site edge servers, and cloud servers, and includes the following steps:

[0062] S1: When a vehicle enters the gate's sensing area, the multimodal sensing units deployed at the gate's edge nodes coordinate data collection. The multimodal sensing units include at least a visual sensor, an RFID reader, a radar sensor, and a ground loop.

[0063] Specifically, in order to achieve comprehensive and multi-dimensional perception of passing vehicles and containers, each sensing unit adopts the following optimized layout based on its physical characteristics and functional positioning:

[0064] Inductive loop detectors: Installation location: Pre-buried under the road surface approximately 5-8 meters before the entrance of each lane. Typically, two loops are installed: the first (trigger loop) is used to initially detect vehicle arrival, and the second (terminus loop) assists in determining whether the vehicle has completely passed. As the most reliable and direct physical contact sensor, the inductive loop is the initial trigger source. When a metal vehicle passes by, it causes a change in the coil's inductance, generating a stable trigger signal. Once the gate edge node receives this signal, it determines that "a vehicle has entered the sensing area" and immediately wakes up or schedules other sensing units to enter working mode. It has high priority and is unaffected by environmental factors such as sunlight, rain, or fog.

[0065] Radar Sensor: Installation Location: Deployed on the side gantry or pillar of the gate, at a height of approximately 3-4 meters, facing laterally towards oncoming traffic, covering the entire lane area from the inductive loop detector to the gate. Millimeter-wave radar is preferred due to its excellent ranging, speed measurement, and motion trajectory tracking capabilities. The radar provides all-weather information on object presence, speed, distance, and rough outline. After a vehicle is triggered by the inductive loop detector, the radar continuously scans, accurately measuring the vehicle's position and speed, and determining whether its motion trajectory is normal (e.g., whether it is reversing or parking in a designated area). Its data is a crucial supplement to visual perception in adverse conditions such as rain, snow, fog, and nighttime, and is also one of the core data sources for the subsequent "abnormal behavior collaborative processing mechanism."

[0066] RFID Reader and Antenna: Installation Location: A combination of top-mounted and side-mounted layouts is used. Top-mounted antenna: Installed vertically downwards on the gantry directly above the lane, primarily reading vehicle electronic tags (front-end tags) mounted on the windshield of trucks. Side-mounted antenna: Installed horizontally on a pillar on the side of the lane, pointing towards the center of the container body, primarily reading electronic tags (body / container tags) mounted on the side of the container body. The RFID system is used for contactless, rapid acquisition of authoritative identification information. When a vehicle enters the antenna's electromagnetic field range (typically 7-10 meters), the reader automatically reads the encrypted vehicle ID, container number, and other information stored within the tag. The gate edge node uses the RFID information as a high-confidence identity verification benchmark, comparing and fusing it with visual recognition results. If the RFID read is successful, it greatly simplifies the identity recognition process; if the read fails (e.g., due to tag damage), the system automatically increases the weight of visual recognition and records the anomaly.

[0067] Visual Sensors: Installation Location: A network of cameras from multiple angles and of various types is used. Top-mounted binocular vision camera: Deployed at the center of the top of the gate, overlooking the entire lane. Used to acquire top-view and oblique-view images of vehicles and containers. Its binocular structure can generate depth information to assist in judging container height, trailer separation status, etc. Lateral high-resolution area scan camera: Deployed on both sides of the lane, at the same height as the sides of the container body. Used to capture high-definition images of the sides of the container body, which is a key perspective for container number OCR recognition and initial screening of container damage. It may also include auxiliary capture cameras: These can be deployed to capture close-ups of license plates, vehicle front features, etc. The visual sensors are the core for acquiring rich texture information and high-precision details. The gate edge nodes control each camera to capture images synchronously or in a time-sharing manner based on trigger signals. For example, the side cameras first identify the container number, while the top camera monitors the overall status. All image data is timestamped and initially corrected at the edge nodes. Visual information is the main basis for initial judgment of container damage, license plate recognition, and container number recognition.

[0068] The gate edge node acts as a local scheduling center, and its collaborative scheduling logic is as follows:

[0069] Initial trigger: The inductive loop detects a vehicle and sends a hard trigger signal to the gate edge node.

[0070] Task wake-up and scheduling: The gate edge node immediately wakes up the radar sensor to start continuous scanning and tracking. At the same time, based on the built-in strategy (such as considering the historical success rate of RFID reading), it decides whether to prioritize starting the RFID reader for identity pre-reading or directly schedule the visual sensor for image capture.

[0071] Parallel data acquisition and feedback: Each sensor operates in parallel according to instructions. The radar continuously reports the vehicle's real-time position and speed; the RFID attempts to read and report electronic tag data; the vision sensor captures images at a preset angle and performs preliminary processing (such as compression and feature extraction).

[0072] Data aggregation: Raw data or preliminary processing results from all sensors are aggregated in real time to the gate edge node via wired (such as Ethernet, industrial bus) or wireless (such as 5G CPE) methods, and are accompanied by a precise timestamp.

[0073] In this step, the hard triggering of the inductive loop and all-weather monitoring by radar ensure extremely high reliability of vehicle perception, avoiding missed triggers caused by environmental factors in pure vision solutions. Identity information (RFID + visual OCR), visual texture information (multi-angle images), and spatial motion information (radar) are acquired simultaneously, forming a multi-dimensional, redundant data set, providing rich material for subsequent fusion decisions. When a sensor fails (e.g., a damaged RFID tag or a camera temporarily affected by glare), the system can automatically compensate by relying on information from other sensors (e.g., relying solely on visual identification or radar to determine vehicle position) according to a collaborative strategy, ensuring uninterrupted flow. The overall robustness of the system is significantly better than single-sensor solutions. The collaborative acquisition in this step is not simply parallel, but an ordered collaboration under intelligent scheduling of edge nodes. It dynamically adjusts based on the environment (e.g., strengthening supplemental lighting and adjusting visual confidence weights at night) and real-time feedback (e.g., immediately enhancing visual acquisition upon RFID reading failure), achieving a basic closed loop of "perception-feedback-adjustment," providing optimized input conditions for subsequent more advanced data fusion and task offloading decisions.

[0074] S2: The gate edge node dynamically determines the coordination strategy among visual sensors, RFID readers, and radar sensors based on the environmental conditions and the perceived target. It then fuses the data based on the real-time confidence levels of each sensor to generate a comprehensive perception result containing vehicle identity, container identity, and preliminary container condition information. The coordination strategy includes:

[0075] Based on information on illumination, weather, and visibility, the confidence weights of visual sensors and radar sensors in data fusion are dynamically adjusted.

[0076] Based on the reading status of the RFID reader, the ground loop and visual sensor are dynamically switched as the initial sensing trigger source.

[0077] When different sensors produce conflicting identification results for the same target, an arbitration function, which is performed by a radar sensor or a designated vision sensor, is activated for verification, and the confidence weights of each sensor are dynamically updated based on historical data.

[0078] Specifically, the sensor confidence weights are dynamically adjusted: a weight calculation engine can be built into the gate edge node. The engine's inputs include: environmental conditions: real-time data from the environmental perception module (such as illuminance (Lux value), visibility in meters, and weather classification codes); and real-time sensor performance metrics: such as image sharpness score and signal-to-noise ratio (SNR) for visual sensors; and SNR and target point cloud density for radar sensors. Its operating logic is as follows: during the day with good lighting and high visibility, the visual sensor weight is set to a high level (e.g., 0.8), and the radar sensor weight is set to a low level (e.g., 0.2), because visual information is abundant and accurate at this time. At night, or under low visibility conditions such as rain, snow, fog, or haze, the system automatically lowers the visual sensor weight (e.g., to 0.3) while significantly increasing the radar sensor weight (e.g., to 0.7), which is not dependent on optical conditions. For RFID, its weight is typically kept at a stable high level (e.g., 0.9) because its electromagnetic induction is less affected by the environment. This strategy ensures that the system always relies on the most reliable data source for decision-making in different environments, overcoming the drawback of fixed-weight fusion's performance plummeting in harsh environments.

[0079] Dynamic switching of initial sensing trigger source: The system presets a trigger source decision state machine. The default and optimal trigger source is the RFID reader / writer because of its fast recognition speed and extremely high accuracy. Its working logic is as follows: Normal process: A vehicle enters, the RFID successfully reads the tag, the system uses this as the primary identity source, and triggers the visual sensor for verification and container condition check. Abnormal switching process: If the RFID fails to read any valid tag within a preset time window (e.g., 2 seconds after the vehicle enters), the edge node immediately determines "RFID trigger failure". The system then checks the status of the inductive loop: If the inductive loop has been triggered, it immediately switches to the "untagged vehicle processing process" based on the inductive loop, and schedules the visual sensor for enhanced license plate and container number recognition. In extreme cases (e.g., inductive loop failure), the system can rely on the stable motion trajectory provided by the radar sensor as the trigger and vehicle presence criterion, and initiate visual recognition. This strategy solves the problem of over-reliance on RFID tags, ensuring that even in cases of tag damage, unauthorized vehicles, or no tags, the system can still sense vehicles and initiate processes through redundant triggering mechanisms, greatly improving the system's availability and generalization capabilities.

[0080] Conflict Arbitration and Confidence Dynamic Update: Conflict Detection: When the descriptions of the same target by different sensors are inconsistent, for example, the visual OCR recognizes the license plate number as "沪A12345", while the RFID reads the license plate number as "沪B67890", the system immediately detects an "identity conflict". Initiate Arbitration: The system initiates the arbitration process. Due to its precise ranging ability, the radar sensor is designated as the spatial reference arbitrator. The system traces back the radar trajectory and confirms that there is only one vehicle in the current lane, ruling out the possibility of data mismatch caused by vehicle intersection. Meanwhile, a high-definition visual sensor with a specified viewing angle can be dispatched to take a close-up shot and identify the license plate again. Verdict and Update: If the arbitration result (such as the second high-definition visual recognition) supports the RFID data, it is determined that the RFID is correct and the visual OCR's current recognition is incorrect. The system not only outputs the RFID data as the comprehensive result but also dynamically updates the confidence database: slightly reduces the confidence of "visual license plate recognition" for this lane under similar lighting conditions and consolidates the confidence of "RFID reading". In the long run, the system can learn that visual recognition is prone to errors in a specific lane during backlight periods, and thus automatically assigns a higher initial weight to RFID under future similar conditions. This strategy enables the system's self-learning and continuous optimization. It can not only solve single conflicts but also make the fusion strategy more accurate and adaptive through historical data accumulation, evolving from "rule-based fusion" to "experience-based fusion".

[0081] Confidence Fusion to Generate Comprehensive Perception Results: After applying the above collaborative strategy, dynamic weights are assigned to the data of each sensor and weighted fusion is performed: Vehicle / Container Identity Fusion: Adopt the "weighted voting" or "maximum a posteriori probability" model. Generation of Initial Judgment Information on Container Condition: The image stream collected by the visual sensor is input into the lightweight residual network model deployed on the edge node, and the abnormal probability scores and regional coordinates of the key parts of the container (corner fittings, container doors, side plates) are output in real time. This result is directly incorporated into the comprehensive perception result as the "initial judgment information on container condition". Result Encapsulation: The finally generated comprehensive perception result is a structured data packet, which at least includes: The fused vehicle identity (license plate number) and confidence. The fused container identity (container number) and confidence. The initial judgment information on container condition (normal / abnormal flag, abnormal area, initial judgment probability). Timestamp, lane number, and the original data indexes of each sensor. Step S2 transforms the original multi-modal data into a high-confidence, deep-semantic "comprehensive perception result" through a set of tightly coupled, dynamic collaborative strategies with feedback and learning capabilities.

[0082] S3: According to the service type, data complexity of the comprehensive perception result, and the real-time load of the gate edge node, through the preset dynamic task offloading strategy, allocate some or all of the perception data and analysis tasks to the yard edge server or cloud server for processing; The dynamic task offloading strategy includes:

[0083] Based on the comprehensive perception results, perform intent recognition to predict the vehicle's passage or business intent;

[0084] If the intention is for regular fast passage and the load on the gate edge node is below the threshold, then all processing and control will be completed by the gate edge node.

[0085] If a complex business intent is identified or an abnormal container condition is detected, the high-resolution visual data and related information are packaged and uploaded to the terminal edge server via a high-speed communication link for detailed inspection of container damage and verification of business documents.

[0086] If the site edge server is overloaded or the model version needs to be updated, the task will be further offloaded to the cloud server.

[0087] Specifically, vehicle passage / business intent recognition: A lightweight intent recognition model can be run at the gate edge node or the terminal edge server. The model's input features include: Vehicle / container identity: whether it belongs to a pre-booked or whitelisted vehicle. Initial container condition assessment: whether it is marked as "abnormal". Business database query results: based on the vehicle / container number, real-time query from the terminal business system to see if there are any pending transactions (such as "empty container pickup", "return of loaded container", "payment", "customs inspection"). Simple behavioral characteristics: whether the vehicle is stopped and waiting in a designated area (potentially indicating the need for manual interaction). Working logic and output: The model maps the above features to several preset intent categories, such as: Regular fast passage: whitelisted vehicles, normal container condition, no pending complex transactions; intent is to pass through the gate quickly. Complex transaction processing: vehicles have pending payments, document verification (such as checking the pickup voucher against the actual container number), special cargo registration, etc.; intent is to complete the transaction handover. Abnormal handling: initial container condition assessment is abnormal, or the system identifies vehicle information that does not match the reservation; intent is to initiate security checks and detailed recording. By anticipating business intent, the system can plan the allocation blueprint of computing resources before task execution, achieving efficient scheduling driven by objectives and avoiding resource waste or delays caused by indiscriminate data processing.

[0088] Multi-level, condition-triggered unloading decisions: Based on the identified intent, the system combines real-time status to execute the following hierarchical decisions:

[0089] Scenario 1: Local Processing (Edge Autonomy); Triggering Conditions: Intent = Regular Fast Passage AND Real-time Load of Gate Edge Node < Dynamic Threshold. Execution Action: All processing loops are completed locally. The gate edge node uses a local lightweight model to quickly complete final identity verification, directly generating a "allow" command to control the gate to lift. The entire process is completed in milliseconds. The dynamic threshold is not a fixed value but is dynamically calculated based on historical traffic predictions, current network latency, and the health status of the site server, ensuring that buffer computing power is reserved before peak periods.

[0090] Scenario 2: Offloading to the site edge server (near-end collaboration); Triggering conditions: Intent = Complex business processing OR Initial assessment of abnormal box condition OR Local load > dynamic threshold. Execution action: The gateway edge node binds the high-resolution original image, multi-modal fusion structured data, and coordinates of the initially identified abnormal area into a "task package to be detailed inspection". This data packet is marked as high priority and uploaded to the site edge server via a high-speed link such as 5G or fiber optic. The network queue manager prioritizes scheduling such data packets. The site edge server runs a high-precision damage recognition model, a complex OCR engine, or performs deep integration and verification with the core business system. Offloading computationally intensive, model-complex tasks requiring core data integration to a near-end server with stronger computing power and more comprehensive data not only frees up local resources and ensures the processing quality of complex tasks but also leverages the low-latency advantage of the edge intranet.

[0091] Scenario 3: Further offloading to cloud servers (remote protection and optimization); Triggering conditions: Edge server load > high threshold OR the model version required for the task is unavailable locally / nearby OR cross-regional data collaboration is required. Execution actions: When multiple gateways simultaneously initiate a large number of complex tasks, causing the edge server to overload, the system automatically routes some tasks to cloud elastic computing resources (such as cloud server clusters). When a new version of a model for damage identification is detected, the latest model service in the cloud can be directly called for processing, while the new model is silently deployed to the edge side for updates in the background. All detailed inspection reports and process data are synchronized to the cloud for macro-data analysis, model retraining, and long-term archiving. The cloud, as the final computing power pool, model factory, and data center, provides almost unlimited elastic scalability, ensuring system availability under extreme peak conditions and enabling centralized optimization and iteration of algorithms.

[0092] The coordination and guarantee mechanism of the unloading process: Gate edge nodes, site servers, and the cloud synchronize their respective load rates, network latency, and service availability in real time through heartbeat mechanisms and status publish / subscribe. Unloading decisions are executed by a distributed decision engine. If an unloading request to the site server times out or receives no response, the engine automatically triggers a fallback mechanism, such as attempting cloud unloading or downgrading to local simplified processing and recording alarms. Regardless of where the task is processed, the final processing result (such as "verification passed," "X type of damage found") is returned to the decision initiator (usually the gate edge node or a designated collaborative control node), which generates the final control command to ensure the uniformity of the control flow. The dynamic task unloading strategy defined in step S3 constructs a three-level collaborative elastic computing network of "local-near-cloud." Through intelligent intent recognition and real-time resource awareness, it achieves the optimal spatial distribution of computing tasks, fundamentally solving the computing bottleneck problem. This allows the system to meet the extreme real-time requirements of conventional operations while supporting the high-precision computing needs of complex businesses and deep detection, while ensuring the overall high availability and scalability of the system.

[0093] The initial condition assessment of the container is obtained through real-time analysis of the visual data stream by a lightweight damage detection model deployed at the gate edge nodes. If the initial assessment indicates an anomaly, a high-priority data upload and detailed analysis process is triggered. The generation of the initial condition assessment information and the detailed analysis process are carried out collaboratively, including:

[0094] The lightweight damage detection model performs real-time localization and initial screening of key parts of containers in video streams, and outputs the probability and region of anomalies.

[0095] When the initial screening abnormality probability exceeds the first threshold, the gate edge node controls the supplementary lighting equipment and multi-angle vision sensors to capture a high-definition multi-view image set of the target area.

[0096] The high-definition multi-view image set is bound with the vehicle and container identity information to form a task package to be inspected in detail, and sent to the site edge server according to the dynamic task unloading strategy.

[0097] The edge server at the site runs a high-precision damage identification model, processes the task package, generates a standardized report containing quantified damage type, location, and size, and feeds it back to the gate control system and business management platform.

[0098] Specifically, the first step is: real-time localization and initial screening of the lightweight model;

[0099] On the edge computing nodes of the gate, a specially optimized lightweight convolutional neural network model (such as a model based on a trimmed version of the MobileNet or ShuffleNet architecture) is deployed. The model's input is a medium-resolution video stream (e.g., 720p) from a fixed-viewpoint visual sensor (e.g., a side-view camera), rather than a full-size high-definition stream, to significantly reduce computational load. The model first identifies and locates key inspection areas of the container, such as the four corner fittings, front and rear doors, and the center area of ​​the side panels. These areas are high-risk areas for damage (e.g., deformation, corrosion, cracks). For each located area, the model performs fast binary or hierarchical classification inference, outputting an anomaly probability score P (e.g., 0.0 to 1.0) and the corresponding area bounding box coordinates. The lightweight model ensures extremely short single-frame processing time (e.g., <50ms), meeting the requirements of real-time video stream analysis, allowing multiple frames to be scanned while vehicles pass at a constant speed. With the limited computing power of the edge nodes, it can run concurrently with other tasks (e.g., license plate recognition). The design goal of this model is "better to report false positives than to miss false positives," meaning it is highly sensitive and ensures that potential anomalies can be detected initially, with the final determination left to a high-precision backend model.

[0100] Step 2: Threshold Triggering and Adaptive High-Definition Forensics;

[0101] Edge nodes continuously monitor the initial screening output P. A first threshold A1 (e.g., 0.65) is set for each key area. If P > A1 in any area, the system immediately identifies it as an "initial anomaly." Upon triggering, the gate edge nodes do not simply record the result but initiate a precise high-definition data acquisition sub-process: immediately illuminating the matrix LED lights for that area to eliminate shadows, reflections, and other interference, ensuring optimal image quality. Based on the coordinates of the anomaly area, visual sensors deployed in different locations (e.g., top camera, camera on the other side) are scheduled to simultaneously or rapidly capture images of that area, forming a high-definition image set (e.g., 1080p or higher) containing 2-3 different perspectives. Multi-angle images are crucial for determining the depth of the dent and the direction of the crack. This step separates "detection" from "evidence collection." The lightweight front-end model quickly identifies suspicious points, and upon triggering, immediately initiates "dedicated high-definition evidence collection," ensuring the timeliness of the overall process and providing high-quality, multi-dimensional first-hand image evidence for subsequent precise analysis.

[0102] Step 3: Task package building and intelligent uninstallation;

[0103] The edge node binds the high-definition multi-view image set, the coordinates and probabilities of the areas initially identified as anomalies, and the vehicle and container identities obtained from step S2, encapsulating them into a structured task package for detailed inspection. The handling of this task package fully follows the dynamic task offloading strategy described in S3. Since detailed container condition inspection is a computationally intensive and high-precision complex task, the system prioritizes offloading it to the terminal edge server. The task package is transmitted via a high-speed network link and marked as high priority to ensure rapid response.

[0104] Step 4: Detailed backend analysis and standardized report generation;

[0105] After receiving the task package, the edge server at the terminal utilizes its powerful computing resources to run a high-precision damage identification model (such as a large model based on ResNet, DenseNet, or Vision Transformer). This model can: accurately classify damage by determining the specific type of damage (e.g., dents, scratches, rust, cracks, holes); perform pixel-level segmentation to precisely outline the damage; and quantify dimensions by combining depth information from binocular vision or known container scale features to calculate the actual physical dimensions of the damage (length, width, area, dent depth). The server automatically generates a standardized damage inspection report based on the analysis results, which includes at least: container number, inspection time, damage type, specific location on the container (e.g., "lower left corner"), dimensional data, and attached high-resolution evidence images. This report is fed back in real time to two key systems: the gate control system, which uses this report as a decision-making basis to automatically determine "prohibit passage" or "pass but record," and triggers audible and visual alarms to alert staff. Business Management Platform: Reports are automatically archived into the container's logistics file, providing tamper-proof digital evidence for subsequent maintenance, insurance claims, liability determination, and customer reports.

[0106] S4: Receives the processing results from the site edge server or cloud server, combines them with the locally generated comprehensive perception results, and makes decisions by the gate edge node or the cooperating site edge server to generate gate control commands. The control commands are used to control the operation of the barrier gate and human-machine interaction terminal.

[0107] Specifically, the decision engine synchronizes two key inputs: local comprehensive perception results (from step S2), providing real-time on-site information such as vehicle identity, location, behavior, and initial assessment of container condition; and remote processing results (from step S3), providing authoritative backend conclusions such as business verification and detailed damage inspection reports. Then, flexible decision-making is employed: decision-making power is dynamically allocated according to the scenario: In normal situations (edge ​​decision-making), for simple scenarios such as "fast passage," the gate edge node directly utilizes local and remote results to quickly make a release / interception decision, achieving minimal latency. In complex situations (central decision-making), for scenarios involving complex business logic, severe damage, or security threats, the site edge server, possessing a global view, makes centralized decisions to ensure rigor. The decision results are transformed into specific, executable control command sequences and sent to the corresponding devices: for the barrier gate, "raise," "lower," or "emergency lock" commands are sent; for the human-machine interface terminal, the control display shows text / QR codes, the voice announcer plays prompts, and the audible and visual alarm sounds alarms. This step completes an intelligent closed loop from "perception-analysis" to "action," balancing response speed and processing complexity through the dynamic distribution of decision-making power, and accurately translating business conclusions into equipment actions, thereby achieving automated and flexible control of gate operations.

[0108] S5: Abnormal Behavior Collaborative Processing Mechanism: The radar sensor and vision sensor work together to monitor vehicle behavior trajectory in real time; when following too closely, abnormal lingering, or driving in the wrong direction is detected, the calculation priority of the gate edge node is immediately increased and the local behavior analysis model is started; if it is determined to be a security threat, the gate edge node directly generates an alarm and sends an immediate control command to the barrier gate and the audible and visual alarm equipment, while the event snapshot and data are uploaded asynchronously.

[0109] Specifically, collaborative monitoring involves data fusion between radar and vision sensors. Radar provides precise vehicle speed, distance, and motion vectors, while vision assists in vehicle classification and trajectory intent determination. The two complement each other, ensuring continuous and reliable tracking of vehicle behavior trajectories under all-weather conditions.

[0110] Anomaly detection and priority enhancement: The system's built-in rule engine analyzes the fused trajectory data in real time. Once a preset high-risk pattern is identified (such as following distance less than the safety threshold, vehicle remaining stationary for an extended period in a non-waiting area, or abnormal trajectory direction), the highest calculation priority is immediately triggered within the gate edge node, suspending some non-critical tasks and initiating a dedicated local lightweight behavior analysis model for rapid secondary confirmation and threat assessment.

[0111] Local instant response and asynchronous reporting: If the model determines that it is a real security threat (such as confirming an attempt to tailgate and break through the checkpoint), the system will bypass the regular task unloading and decision-making process, and directly generate the highest priority control command from the gate edge node, which will be executed synchronously: Control command: Immediately send an emergency lowering or locking command to the gate, and activate the audible and visual alarm for on-site deterrence and warning. Data retention: At the same time, snapshots of multi-sensor data (radar point cloud, video clips, timestamps) for a period of time before and after the event are cached locally, and then asynchronously uploaded to the site server and cloud via a background thread for post-event tracing and analysis, without affecting the execution of the current real-time control flow.

[0112] This step establishes a high-priority security loop independent of core business processes. By granting edge nodes autonomous decision-making and execution authority in emergency situations, a closed loop with extremely short latency from risk perception to physical intervention is achieved, greatly enhancing the gateway's proactive defense capabilities and overall security in response to sudden security incidents.

[0113] S6: Predictive-based energy efficiency management strategy: Based on historical traffic data and real-time sensing information, predict the traffic flow status in future periods; during the predicted low traffic periods, the non-critical sensing units of the gate edge node control part enter a low-power sleep mode and extend the sensing cycle; when the sensing units are triggered again or the predicted peak period arrives, all sensing units are automatically woken up and restored to normal working mode.

[0114] Specifically, traffic flow prediction: The system, located in the cloud or on the station server side, uses historical traffic records and real-time perceived traffic flow status, and employs time series analysis models (such as ARIMA) or machine learning models to predict the vehicle arrival rate and traffic volume of each gate in the next short period (such as the next 15 minutes).

[0115] Strategy Execution and Control: Prediction results are sent to each gate edge node. Based on the predicted traffic flow, nodes dynamically adjust their own and their subordinate sensing units' operating modes: Low Traffic / Idle Periods: Nodes control some non-critical sensing units (such as some supplementary lights and high-resolution scanning cameras) to enter low-power sleep or deep standby modes. Simultaneously, for continuously operating sensors (such as radar), their detection and scanning cycles can be lengthened (e.g., from 50ms to 200ms), thereby reducing overall power consumption. Peak Periods or Triggered Wake-up: When the predicted peak period arrives, or any sensor in operation (such as a ground loop coil) is triggered, the node immediately and automatically wakes up all sensing units and instantly restores them to normal operating modes and parameters, ensuring full system processing capacity.

[0116] This step, without sacrificing critical business responsiveness (ground sensor triggering can wake up immediately), intelligently reduces system energy consumption during low-load periods through accurate traffic prediction. This not only saves electricity costs but also effectively extends the service life of outdoor electronic equipment, demonstrating the system's high efficiency and sustainability.

[0117] like Figure 2 As shown, a multimodal sensing and edge computing collaborative control system for intelligent gates is used to realize a multimodal sensing and edge computing collaborative control method for intelligent gates. The system includes:

[0118] A multimodal sensing unit, deployed at the gate, includes at least a visual sensor for visual information acquisition, an RFID reader for identification, a radar sensor for ranging and behavior perception, and a ground loop coil for sensing vehicle signals. The visual sensor includes a binocular vision camera deployed at the top of the gate to acquire multi-view images of the container and assist in generating depth information; and a high-resolution area scan camera deployed on the side to capture high-definition images of the container's side. The gate edge computing node performs time synchronization and spatial alignment of the image data acquired from different visual sensors through a first processing module.

[0119] The gateway edge computing node, deployed locally at the gateway and connected to the multimodal sensing unit, is configured with a first processing module and a first communication module for performing sensing scheduling, data fusion, lightweight model inference, real-time control, and task offloading decisions. The gateway edge computing node also includes a local consensus module, and the site edge server includes a corresponding remote consensus module. The local consensus module is used to generate an event declaration containing a data hash value when a critical event or data anomaly is detected. The remote consensus module is used to verify the received event declaration and compare it with the declarations from other gateway edge computing nodes for consensus confirmation. The confirmed data is recorded in a distributed ledger.

[0120] The site edge server is deployed in the site-level network and is equipped with a second processing module and a second communication module. It is used to receive and process complex computing tasks offloaded from the gate edge computing node and run high-precision analysis models.

[0121] The cloud server is connected to the site edge server and the gate edge computing node network for big data analysis, model training and global system management.

[0122] The gate execution and interaction unit, including the barrier gate, information display screen, and voice broadcaster, is controlled by the gate edge computing node or the site edge server.

[0123] The system also includes an environmental sensing module that integrates temperature, humidity, and illuminance sensors. Its output signal is connected to the first processing module of the gate edge computing node as one of the input parameters for dynamically determining the collaborative strategy.

[0124] This system utilizes a multimodal sensing unit comprised of visual sensors (binocular cameras, high-speed scanning cameras), RFID readers, radar sensors, and inductive loops. Combined with real-time data input from an environmental sensing module, this establishes a redundant and complementary sensing network. This enables the system to simultaneously acquire high-definition textures, spatial depth, electronic identification, precise movement trajectories, and physical presence signals of containers. Furthermore, it adaptively adjusts its strategy in response to changes in lighting and weather, significantly overcoming the limitations of single sensors in complex scenarios and achieving all-weather, highly reliable panoramic perception.

[0125] The gate edge computing node acts as a local intelligent hub, executing lightweight model inference and real-time control, ensuring millisecond-level low latency for core functions such as passage and security response. The site edge server provides powerful near-end computing power to handle complex detailed inspection tasks. The cloud server is responsible for global optimization and model training. This three-tiered elastic architecture, combined with a dynamic task offloading strategy, achieves optimal allocation of computing tasks across the "edge-near-cloud" hierarchy, satisfying real-time requirements while efficiently utilizing computing resources, thus solving the performance bottleneck problem of centralized processing.

[0126] The system innovatively introduces local and remote consensus modules, forming a lightweight and trustworthy mechanism. When a critical event occurs (such as severe damage or card breach), edge nodes generate an event declaration with a hash value, which is then verified by the site server and stored in the distributed ledger. This process ensures the immutability and traceability of critical data, providing authoritative and credible technical evidence for business disputes, security audits, and liability determination—a core capability lacking in traditional gateway systems.

[0127] The introduction of the environmental perception module frees the system's decision-making from relying on fixed thresholds. Edge computing nodes at the gate can dynamically adjust perception fusion strategies and device parameters (such as supplemental lighting intensity) based on real-time lighting, temperature, and humidity. Simultaneously, the unified model training and management capabilities in the cloud support continuous algorithm iteration and smooth deployment to the edge. This endows the entire system with strong environmental adaptability and self-evolution capabilities, significantly reducing performance fluctuations and maintenance complexity caused by environmental changes.

[0128] The gate execution and interaction unit is directly controlled by the intelligent decision-making of edge nodes or station servers, and can receive precise command sequences (such as specific information display, voice broadcast, and gate action). This changes the traditional mechanical response mode of equipment, realizing refined and humanized control based on context understanding, which not only improves operational accuracy but also enhances the interaction experience for drivers and staff.

[0129] Through deep integration and collaborative design of hardware and software, this collaborative control system not only achieves full-link closed-loop enhancement in terms of functionality, including perception, computing, decision-making, control, and traceability, but also comprehensively improves system attributes such as reliability, real-time performance, trustworthiness, and adaptability, providing a solid technical foundation for building truly intelligent and unmanned modern gates.

[0130] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0131] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0132] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0133] The present invention has been described above by way of example. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any improvements made by adopting the inventive concept and technical solution of the present invention, or direct application to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A multimodal sensing and edge computing collaborative control method for intelligent gates, characterized in that, For systems that include gate edge nodes, site edge servers, and cloud servers, the following steps are included: S1: When a vehicle enters the gate's sensing area, the multimodal sensing units deployed at the gate's edge nodes coordinate data collection. The multimodal sensing units include at least a visual sensor, an RFID reader, a radar sensor, and a ground loop. S2: The gate edge node dynamically determines the coordination strategy between the visual sensor, RFID reader and radar sensor based on the environmental conditions and the sensing target, and fuses the data collected by each sensor in real time according to the confidence level to generate a comprehensive sensing result that includes vehicle identity, container identity and preliminary container condition information. S3: Based on the business type, data complexity, and real-time load of the gate edge nodes according to the comprehensive perception results, some or all of the perception data and analysis tasks are distributed to the site edge server or cloud server for processing through a preset dynamic task offloading strategy. S4: Receives the processing results from the site edge server or cloud server, combines them with the locally generated comprehensive perception results, and makes decisions by the gate edge node or the cooperating site edge server to generate gate control commands. The control commands are used to control the operation of the barrier gate and human-machine interaction terminal. The initial condition assessment information of the container is obtained by real-time analysis of the visual data stream by a lightweight damage detection model deployed at the edge node of the gate. If the initial assessment result indicates that there is an anomaly, a high-priority data upload and detailed analysis process is triggered.

2. The multimodal sensing and edge computing collaborative control method for intelligent gates according to claim 1, characterized in that, The cooperative strategies in step S2 include: Based on information on illumination, weather, and visibility, the confidence weights of visual sensors and radar sensors in data fusion are dynamically adjusted. Based on the reading status of the RFID reader, the ground loop and visual sensor are dynamically switched as the initial sensing trigger source. When different sensors produce conflicting identification results for the same target, an arbitration function, which is performed by a radar sensor or a designated vision sensor, is activated for verification, and the confidence weights of each sensor are dynamically updated based on historical data.

3. The multimodal sensing and edge computing collaborative control method for intelligent gates according to claim 1, characterized in that, The dynamic task unloading strategy in step S3 includes: Based on the comprehensive perception results, perform intent recognition to predict the vehicle's passage or business intent; If the intention is for regular fast passage and the load on the gate edge node is below the threshold, then all processing and control will be completed by the gate edge node. If a complex business intent is identified or an abnormal container condition is detected, the high-resolution visual data and related information are packaged and uploaded to the terminal edge server via a high-speed communication link for detailed inspection of container damage and verification of business documents. If the site edge server is overloaded or the model version needs to be updated, the task will be further offloaded to the cloud server.

4. The multimodal sensing and edge computing collaborative control method for intelligent gates according to claim 3, characterized in that, The generation and detailed analysis of initial container condition information are carried out in a coordinated manner, including: The lightweight damage detection model performs real-time localization and initial screening of key parts of containers in video streams, and outputs the probability and region of anomalies. When the initial screening abnormality probability exceeds the first threshold, the gate edge node controls the supplementary lighting equipment and multi-angle vision sensors to capture a high-definition multi-view image set of the target area. The high-definition multi-view image set is bound with the vehicle and container identity information to form a task package to be inspected in detail, and sent to the site edge server according to the dynamic task unloading strategy. The edge server at the site runs a high-precision damage identification model, processes the task package, generates a standardized report containing quantified damage type, location, and size, and feeds it back to the gate control system and business management platform.

5. The multimodal sensing and edge computing collaborative control method for intelligent gates according to claim 1, characterized in that, The method also includes: S5: Abnormal Behavior Collaborative Processing Mechanism: The radar sensor and vision sensor work together to monitor vehicle behavior trajectory in real time; when following too closely, abnormal lingering, or driving in the wrong direction is detected, the calculation priority of the gate edge node is immediately increased and the local behavior analysis model is started; if it is determined to be a security threat, the gate edge node directly generates an alarm and sends an immediate control command to the barrier gate and the audible and visual alarm equipment, while the event snapshot and data are uploaded asynchronously.

6. The multimodal sensing and edge computing collaborative control method for intelligent gates according to claim 1, characterized in that, The method also includes: S6: Predictive-based energy efficiency management strategy: Based on historical traffic data and real-time sensing information, predict the traffic flow status in future periods; during the predicted low traffic periods, the non-critical sensing units of the gate edge node control part enter a low-power sleep mode and extend the sensing cycle; when the sensing units are triggered again or the predicted peak period arrives, all sensing units are automatically woken up and restored to normal working mode.

7. A multimodal sensing and edge computing collaborative control system for an intelligent gate, used to implement the multimodal sensing and edge computing collaborative control method for an intelligent gate as described in any one of claims 1-6, characterized in that, The system includes: The multimodal sensing unit, deployed at the gate, includes at least a visual sensor for visual information acquisition, an RFID reader for identification, a radar sensor for ranging and behavior perception, and a ground loop for sensing vehicle signals. The gate edge computing node is deployed locally at the gate and connected to the multimodal sensing unit. It is equipped with a first processing module and a first communication module for performing sensing scheduling, data fusion, lightweight model inference, real-time control, and task offloading decisions. The site edge server is deployed in the site-level network and is equipped with a second processing module and a second communication module. It is used to receive and process complex computing tasks offloaded from the gate edge computing node and run high-precision analysis models. The cloud server is connected to the site edge server and the gate edge computing node network for big data analysis, model training and global system management. The gate execution and interaction unit, including the barrier gate, information display screen, and voice broadcaster, is controlled by the gate edge computing node or the site edge server.

8. The multimodal sensing and edge computing collaborative control system for intelligent gates according to claim 7, characterized in that, Visual sensors include: A binocular vision camera deployed at the top of the gate is used to collect multi-view images of the container and assist in generating depth information; High-resolution area scan cameras deployed on the side are used to capture high-definition images of the sides of the container. The gate edge computing node performs time synchronization and spatial alignment of image data collected from different vision sensors through the first processing module.

9. The multimodal sensing and edge computing collaborative control system for intelligent gates according to claim 7, characterized in that, The gateway edge computing node also includes a local consensus module, and the site edge server includes a corresponding remote consensus module. The local consensus module is used to generate an event declaration containing a data hash value when a critical event or data anomaly is detected. The remote consensus module is used to verify the received event declaration and compare it with the declarations from other gateway edge computing nodes for consensus confirmation. The confirmed data is recorded in a distributed ledger.

10. The multimodal sensing and edge computing collaborative control system for intelligent gates according to claim 7, characterized in that, The system also includes an environmental sensing module that integrates temperature, humidity, and illuminance sensors. Its output signal is connected to the first processing module of the gate edge computing node as one of the input parameters for dynamically determining the collaborative strategy.

Images