A deep learning-based port truck weighing judgment method

Abstract

The application provides a kind of port truck weighing judgment method based on deep learning, by introducing the vehicle state modeling mechanism based on visual perception, the spatial position, axle level structure and time stability of vehicle in the process of weighing are comprehensively perceived, and the visual state judgment model reflecting whether the vehicle is in the standard on-weighing state is constructed;On this basis, the visual state of vehicle and the weighing process are strongly coupled control, only when the vehicle meets the preset stable on-weighing condition, the weighing sampling is triggered and the effective weighing data is locked, the credibility of weighing result is guaranteed from the source, the application can effectively solve the problems of vehicle not fully on-weighing, weighing false triggering, non-truck parallel interference and artificial experience dependence in the process of port weighing, realize the closed-loop control of "state perception-weighing sampling-rule determination-control output" in the process of weighing, improve the accuracy, safety and unattended operation ability of port weighing operation, have good engineering application value and popularization prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle weighing, and particularly relates to a method for judging trucks passing the weighbridge in the port area based on deep learning. Background Art

[0002] Weighing trucks in the port area is a key business link in port logistics operation, billing settlement, and overloading supervision. Its essence is to determine whether the truck has completed the standardized weighing action and whether the weighing result is valid, and accordingly decide whether to allow the vehicle to enter the normal weighing process. In the actual port operation environment, the weighing result of the truck is not only related to the vehicle's own load, but also highly dependent on the vehicle's parking position, weighing integrity, and running state during the weighing process. Therefore, how to accurately judge the credibility of the weighing data under unattended conditions is the core technical problem restricting the reliable operation of the intelligent weighbridge system in the port area.

[0003] Existing port weighbridge systems mostly trigger the weighing process by means of inductive loop, pressure sensor or manual assistance. The typical technical route is: after detecting that the vehicle enters the weighing area, directly collect the weighbridge weighing data, and combine manual experience or simple rules to judge whether there is overloading. However, in practical applications, such solutions generally have the following deficiencies: (1) Insufficient ability to perceive the weighing state. Existing solutions usually can only perceive whether the vehicle enters the weighbridge area, and it is difficult to judge whether the vehicle is in a complete and standardized weighing state. When the truck is driving slowly, the front and rear axles are not fully on the weighbridge, the vehicle is pressing the edge or parked obliquely, etc., the weighing data is likely to have a large error, but the system still uses it as a valid result, affecting the accuracy of the weighing conclusion.

[0004] (2) Decoupling of weighing data and vehicle state. Most systems do not constrain the visual state of the vehicle in the weighing process. Once the weighing data is triggered, it is directly used for subsequent judgment, lacking a state verification mechanism for the validity of the weighing data, resulting in possible incorrect weighing control decisions under non-standard weighing conditions.

[0005] (3) Dependence on manual experience to complete key decisions. In the current port operation, the final decisions on vehicle type identification, axle number judgment, and whether it is overweight often still need to rely on manual experience for intervention. Especially in complex scenarios such as non-truck parallel driving, mis-triggering of weighing, or unstable vehicle weighing state, the system is difficult to independently complete reliable judgment, which is not conducive to the popularization and application of the unattended weighing mode.

[0006] (4) Difficulty in adapting to policy and port area rule changes. Existing weighbridge systems mostly use fixed thresholds or static rules for load determination, lacking a mechanism for dynamic association with vehicle types, axle-level structures, and port management policies. When relevant policies or management requirements change, the system has insufficient adaptability.

[0007] In summary, existing technologies generally lack a technical solution capable of deeply coupling vehicle visual status with the weighing process in complex real-world port environments, thereby automating weighing control decisions. Therefore, there is an urgent need to propose an intelligent weighing control method for unmanned port applications, enabling reliable perception of vehicle weighing status, automatic determination of weighing data validity, and intelligent control of the weighing process. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention proposes a deep learning-based method for judging truck weighing in port areas. The aim is to: introduce a vehicle state modeling mechanism based on visual perception to comprehensively perceive the spatial position, axle structure, and temporal stability of the truck during the weighing process, constructing a visual state judgment model that reflects whether the vehicle is in a compliant weighing state; based on this, the visual state of the vehicle is strongly coupled and controlled with the weighing process, triggering weighing sampling and locking valid weighing data only when the vehicle meets preset stable weighing conditions, thus ensuring the reliability of the weighing results from the source.

[0009] To achieve the above-mentioned objectives, this invention provides a method for judging truck weighing in port areas based on deep learning, the method comprising: S101. An image acquisition device is set up in the weighing area of ​​the port area to continuously acquire video of vehicles entering the weighing area and obtain a video image sequence. The video image sequence includes multiple original vehicle images. The video image sequence is preprocessed. S102. Identify the target vehicle in the preprocessed video image sequence, perform continuous tracking modeling, and obtain the target region of the image; S103. Based on the target identifier, crop the local image region of the vehicle from the target region of the image; S104. Detect the wheels in the local image region of the vehicle and project them in order along the direction of travel to obtain a sorted wheel set. Based on the sorted wheel set, determine the number of axles and the vehicle type, and generate a parameterized representation of the axle-level structure. S105. Based on the conditions of axle-level structural integrity, spatial consistency, motion stability and temporal continuity, jointly determine whether the target vehicle has entered a stable weighing state. S106. Collect vehicle weighing data under stable weighing conditions, combine vehicle type information and rule base to determine overweight vehicles, and execute weighing control decisions.

[0010] Furthermore, step S102 specifically includes the following operations: S201. Input the preprocessed original vehicle image into the deep learning-based target detection model to detect the target vehicle and output the detection result, which includes the candidate image target region in the current frame and the corresponding detection confidence. S202. Assign a unique target identifier to each detected target vehicle based on the detection results; S203. Continuously track and model the spatial position and motion state of the target vehicle in consecutive frames; S204. Based on the continuous tracking results, crop the image target region from the original vehicle image to obtain the target region that corresponds one-to-one with the target identifier.

[0011] Furthermore, step S203 specifically includes the following operations: S301. Establish a target vehicle trajectory set for the (t-1)th frame. The target vehicle trajectory set contains multiple vehicle trajectories, and each vehicle trajectory corresponds to a unique target identifier. S302. For any vehicle trajectory, define its state vector at frame t-1 based on the target vehicle center position, target scale parameters, and estimated motion speed. S303. Based on the state vector of the target vehicle in the previous frame, predict the state of the target vehicle in the t-th frame.

[0012] Furthermore, in step S203, a matching cost function based on spatial overlap relationship is constructed based on the intersection-union ratio function. The matching cost function is used to determine the association probability between the target vehicle and the existing vehicle trajectory. If the association probability is greater than a preset threshold, the state of the corresponding vehicle trajectory is updated. At the same time, the target vehicle inherits the target identifier of the associated vehicle trajectory. If the target vehicle is not associated with any of the existing vehicle trajectories, a new vehicle trajectory is created for the target vehicle and a target identifier is assigned. For vehicle trajectories whose state has not been updated for several consecutive frames, the life cycle of the corresponding vehicle trajectory is terminated.

[0013] Furthermore, in step S103, an image cropping operation is performed based on the target detection bounding box of the target vehicle in the target area of ​​the image to obtain a local image area of ​​the vehicle. The local image area of ​​the vehicle includes the wheel distribution information of the target vehicle and the spatial positional relationship between the target vehicle and the weighing area. The local image area of ​​the vehicle is then bound to the target identifier.

[0014] Furthermore, step S104 specifically includes the following operations: S401. Identify the wheels in the local image region of the vehicle to obtain the positional distribution information of the wheels, and further organize them to obtain a set of wheel targets. S402. Based on the position distribution information of the wheels in the wheel target set, perform parametric modeling of the wheel spatial distribution, and map the two-dimensional position distribution information in the wheel target set into a one-dimensional ordered representation along the direction of travel of the target vehicle, so as to determine the projected position of the wheels in the direction of travel. S403. Based on the projection position, sort and aggregate the one-dimensional ordered representation, group the wheels belonging to the same axle into a group, thereby dividing the wheel target set into several non-overlapping wheel subsets and constructing an axle-level structure sequence. S404. Determine the vehicle model based on the axle-level structure sequence; S405. Generate a parameterized representation of the axle-level structure describing the structural integrity of the truck based on the axle-level structure sequence.

[0015] Furthermore, step S105 specifically includes the following operations: S501. The visual state of the target vehicle during the weighing process is abstracted into discrete state variables, including the state of the vehicle not entering the weighing area, the state of the vehicle partially entering the weighing area, and the state of stable weighing. S502. Based on the integrity of the axle-level structure, spatial consistency, motion stability and temporal continuity, the stable weighing state of a vehicle is defined as a jointly established state under multiple conditions and temporal constraints. S503. Determine whether the target vehicle is in a stable weighing state; otherwise, the weighing data of the weighbridge shall not be accepted.

[0016] Furthermore, step S106 specifically includes the following operations: S601. When the target vehicle is in a stable weighing state, collect the original weighing data of the weighbridge within a continuous time window to obtain the weighing data sequence. S602. Based on the target vehicle model, obtain the maximum permissible weight threshold for the corresponding model from the rule base; S603. Based on the maximum permissible weight threshold obtained in the previous step and the weighing data sequence, determine whether the target vehicle is overweight. S604. Implement weighing control decisions based on the overweight determination results.

[0017] Furthermore, when collecting the original weighing data of the weighbridge, a weight change constraint is introduced. If the weight change constraint is not met, the weighbridge weighing data will not be accepted. The weighing data sequence obtained by S601 is locked only when the target vehicle is in a stable weighing state and the weight change constraint is met. The weighing data sequence is only output after the weighing data sequence is locked.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention elevates the determination of whether a vehicle is weighable from the traditional weighing threshold or single sensor trigger to a stable weighing status determination based on the vehicle's visual state. Through comprehensive analysis of the vehicle's structure, position, and motion state, it can accurately reflect whether the truck is truly and completely in a weighable state.

[0019] 2. This invention proposes a stable weighing state determination method based on the joint constraints of axle structure integrity, axle structure and weighing area spatial consistency, vehicle motion stability and time continuity, which effectively avoids the problem of false triggering of weighing in cases such as the vehicle not being fully on the weighbridge, stopping at the edge, or moving slowly.

[0020] 3. This invention uses a stable weighing state as the sole prerequisite for weighing behavior, and constructs a weighing trigger and data locking gating mechanism to ensure that weighing data is only accepted when the vehicle state meets the specification requirements, thereby significantly improving the accuracy and reliability of weighing results.

[0021] 4. By locking the weighing data after a stable weighing state is established, this invention can effectively suppress the impact of slight vehicle vibrations or short-term state fluctuations on the weighing results, ensuring the consistency and traceability of the weighing data.

[0022] 5. This invention features a strong coupling design between vehicle visual state perception, weighing data acquisition, and weighing control decision-making, enabling automated closed-loop control from vehicle state determination to whether weighing is permitted. This reduces reliance on human experience and intervention, making it suitable for complex unmanned passage environments in port areas.

[0023] 6. This invention combines vehicle model information determined based on axle-level structure with an updatable rule base to achieve flexible adaptation of overweight judgment rules for different vehicle models and under different policy conditions, and has good scalability and engineering practical value. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the overall process of a deep learning-based method for judging trucks weighing in port areas, provided by an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the image target region extraction process provided in an embodiment of the present invention.

[0027] Figure 3This is a schematic diagram of the continuous tracking modeling process provided in an embodiment of the present invention.

[0028] Figure 4 This is a schematic diagram of the process for generating a parameterized representation of a shaft-level structure provided in an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram of the target vehicle status determination process provided in an embodiment of the present invention.

[0030] Figure 6 This is a schematic diagram of the overweight determination and weighing control process provided in an embodiment of the present invention. Detailed Implementation

[0031] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0032] In port area truck weighing scenarios, existing weighing systems typically use weighbridge signals or simple position detection as trigger conditions, combined with human experience or fixed threshold rules to determine whether a vehicle has completed the weighing process. This approach generally suffers from the following problems in practical applications: First, the weighing behavior is decoupled from the vehicle's actual weighing status, making it prone to accidental weighing when the vehicle is not fully on the scale, is parked close to the edge, or is moving slowly. Second, in complex conditions such as multiple vehicles operating in parallel or interference from non-trucks, the system struggles to accurately distinguish between the target truck and interfering vehicles, relying heavily on human intervention. Third, the weighing data lacks effective state constraints and locking mechanisms, making it susceptible to vehicle vibrations or short-term state changes, resulting in unstable and unreliable weighing results.

[0033] To address the aforementioned problems, this invention proposes a deep learning-based method for judging truck weighing in port areas. Based on vehicle visual state perception and with weighing data acquisition and control as the core, it achieves automated and reliable control of the weighing process by jointly judging the truck's structural state, spatial relationships, and temporal stability during the weighing process. (Refer to...) Figure 1 As a preferred embodiment of the present invention, the method includes the following steps: S101. An image acquisition device is set up in the weighing area of ​​the port area to continuously acquire video of vehicles entering the weighing area and obtain a video image sequence, which includes multiple original vehicle images. The video image sequence is then preprocessed.

[0034] Industrial cameras or surveillance cameras are deployed at the weighbridge in the port area to continuously capture video of vehicles entering the weighbridge area. The image acquisition devices are preferably configured with a side-view or oblique side-view angle to ensure the visibility of the vehicle's overall outline and wheel structure in the image. The acquired video image sequences undergo preprocessing operations, including but not limited to image size normalization, noise suppression, and brightness and contrast adjustment, to improve the stability and robustness of subsequent visual perception processes under complex lighting, rain, fog, and nighttime conditions.

[0035] S102. Identify the target vehicle in the preprocessed video image sequence, perform continuous tracking modeling, and obtain the target region of the image.

[0036] S103. Based on the target identifier, crop the local image region of the vehicle from the target region of the image.

[0037] S104. Detect the wheels in the local image region of the vehicle and project and sort them along the direction of travel to obtain a sorted wheel set. Based on the sorted wheel set, determine the number of axles and the vehicle type, and generate a parameterized representation of the axle-level structure.

[0038] S105. Based on the conditions of axle-level structural integrity, spatial consistency, motion stability, and temporal continuity, jointly determine whether the target vehicle has entered a stable weighing state.

[0039] S106. Collect vehicle weighing data under stable weighing conditions, combine vehicle type information and rule base to determine overweight vehicles, and execute weighing control decisions.

[0040] The core innovation of this method lies in elevating the determination of vehicle weighability from a traditional weighing signal trigger or simple location judgment to a "stable weighing state" determination problem based on the vehicle's visual state. In its implementation, the system continuously acquires visual data from trucks entering the weighing lane and continuously identifies and perceives the state of the same target vehicle, obtaining state information reflecting the integrity of the vehicle's axle-level structure and its spatial relationship with the weighing area. Based on this, the system models vehicle weighability as a stable weighing state determination problem. By introducing constraints on structural integrity, spatial consistency, and temporal continuity, it determines whether the vehicle is truly, completely, and stably in a weighable state: when the vehicle has not yet reached a stable weighing state, the weighing data is not accepted; only when a stable weighing state is established does the system trigger weighing sampling and lock the weighing data, using it as the only valid weighing result. Subsequently, the system determines the vehicle type based on the vehicle's axle-level structure and completes overload determination and weighing control decisions based on an updatable rule base.

[0041] Through the above technical solution, the present invention realizes an unattended weighing control method with the vehicle's visually stable weighing state as the core driving condition. It organically integrates vehicle state determination, weighing data acquisition and weighing control decision-making, and constructs a complete state-driven weighing control closed loop, which significantly improves the reliability and anti-false triggering capability of the weighing system under complex working conditions in port areas.

[0042] As one possible implementation method, refer to Figure 2 Step S102 specifically includes the following operations: S201. Input the preprocessed original vehicle image into the deep learning-based target detection model to detect the target vehicle and output the detection result, which includes the candidate image target region in the current frame and the corresponding detection confidence.

[0043] Suppose that the i-th truck target detected in the current frame can be represented as:

[0044] in, This indicates the center position of the target vehicle in the image coordinate system. These represent the width and height of the target vehicle's bounding box, respectively. This represents the confidence level of the corresponding detection result. Based on the detection confidence level, the detection results are divided into a high-confidence detection set and a low-confidence detection set, as shown below:

[0045] in, These represent the high-confidence set and low-confidence set for the target vehicle detection, respectively. The preset detection confidence threshold is used to distinguish between the main target and potential occluded or blurred targets, and the preferred value range is 0.4–0.6.

[0046] S202. Assign a unique target identifier to each detected target vehicle based on the detection results.

[0047] S203. Continuously track and model the spatial position and motion state of the target vehicle in consecutive frames.

[0048] S204. Based on the continuous tracking results, crop the image target region from the original vehicle image to obtain the target region that corresponds one-to-one with the target identifier.

[0049] This implementation inputs preprocessed raw vehicle images into a deep learning-based target detection model to detect target vehicles appearing in the images, outputting candidate image target regions and corresponding detection confidence scores for the current frame. Based on the detection results, the system assigns a unique target identifier to each vehicle entering the monitoring field of view and continuously tracks the same target vehicle using time-series images. This continuous tracking process maintains stable identification and position awareness of the same target vehicle even when the vehicle is moving slowly, experiencing short-term occlusion, or has a complex background. By modeling the spatial position and motion state of the target vehicle in consecutive frames, the system can obtain state estimation results reflecting the vehicle's motion trend and position changes.

[0050] As a further preferred embodiment, refer to Figure 3 Step S203 specifically includes the following operations: S301. Establish a target vehicle trajectory set for the (t-1)th frame. The target vehicle trajectory set contains multiple vehicle trajectories, and each vehicle trajectory corresponds to a unique target identifier.

[0051] To achieve stable perception of the target vehicle across consecutive video frames, this implementation model the target vehicle's motion state over time. Let the set of target vehicle trajectories maintained by the system in frame t-1 be:

[0052] Among them, each truck trajectory Each corresponds to a unique target identifier, TruckID.

[0053] S302. For any vehicle trajectory, define its state vector at frame t-1 based on the target vehicle's center position, target scale parameters, and estimated motion speed.

[0054] For any vehicle trajectory, its state vector at frame t-1 is defined as:

[0055] in Indicates the center position of the target vehicle. Indicates the target scale parameter. This represents the estimated velocity of the target vehicle in the image plane.

[0056] S303. Based on the state vector of the target vehicle in the previous frame, predict the state of the target vehicle in frame t. The prediction process can be represented as:

[0057] Where F is the state transition matrix, used to describe the motion changes of the target vehicle between adjacent frames. The prediction process is not limited to specific filtering or estimation algorithms.

[0058] Meanwhile, to establish a correlation between the detected target and existing truck trajectories, this implementation constructs a matching cost function based on spatial overlap. Let the predicted bounding box of the m-th vehicle trajectory in frame t be... The bounding box of the i-th target vehicle is The matching cost between the two is defined as:

[0059] Where IoU(·) represents the intersection-union ratio function. A vehicle and its trajectory are considered correlated if both simultaneously meet the following conditions:

[0060] in, The preset spatial consistency threshold is preferably set to 0.5. Under the above conditions, the system completes the association between the target vehicle and the vehicle trajectory based on the principle of minimizing matching cost. Once the target vehicle is successfully associated with an existing vehicle trajectory, the system updates the state of that trajectory, and the update process can be represented as follows:

[0061] in, This represents the state update function, used to fuse the predicted state and detection results. Simultaneously, the target vehicle inherits a unique target identifier corresponding to its trajectory; the target identifier update rule is expressed as follows:

[0062] For a target vehicle that is not successfully associated with any existing vehicle trajectory in the current frame, the system creates a new vehicle trajectory for it and assigns a new unique target identifier. The generation rule can be expressed as follows:

[0063] For vehicle trajectories that have not been updated by detection results within several consecutive frames, the system determines that the corresponding target vehicle has left the monitoring area and terminates the lifecycle of the vehicle trajectory.

[0064] In another preferred embodiment, in step S103, based on the target identifier and its corresponding image target region obtained in step S102, a vehicle local image region corresponding one-to-one with the target identifier is cropped from the original vehicle image. Let the detection bounding box corresponding to the target identifier k in frame t be:

[0065] in, This indicates the center position of the target vehicle in the image. This represents the width and height of the target detection bounding box. Therefore, the local vehicle image region can be represented from the original vehicle image. Sub-images obtained by cropping from the middle:

[0066] Crop(·) represents a bounding box-based image cropping operation. The local image region of the vehicle contains the wheel distribution information of the corresponding target vehicle and the spatial relationship between the target vehicle and the weighing area. By binding the local image region of the vehicle to a unique TruckID, the system can ensure that subsequent wheel detection and structural analysis processes only apply to the currently weighed vehicle, even in situations with multiple vehicles operating in parallel or interference from non-truck targets. This avoids erroneous associations between different target vehicles and ensures the uniqueness and continuity of the state analysis object.

[0067] As another preferred implementation, refer to Figure 4 Step S104 specifically includes the following operations: S401. Identify the wheels in the local image area of ​​the vehicle to obtain the positional distribution information of the wheels, and further organize them to obtain the wheel target set.

[0068] This implementation method uses the vehicle local image region obtained in step S103. The data is input into the wheel detection model to identify the wheels of the target vehicle and obtain their positional distribution information in the image space. Let the set of wheel targets detected in a local image region of the vehicle be:

[0069] Each wheel target It can be represented as:

[0070] This represents the position of the i-th wheel in the image coordinate system. This indicates the width and height of the bounding box of the wheel target. This indicates the confidence level of the corresponding wheel detection result.

[0071] S402. Based on the position distribution information of the wheels in the wheel target set, perform parametric modeling of the wheel spatial distribution, and map the two-dimensional position distribution information in the wheel target set into a one-dimensional ordered representation along the direction of travel of the target vehicle, so as to determine the projected position of the wheel in the direction of travel.

[0072] Considering that side-view or oblique-view cameras are commonly used in port weighing scenarios, this implementation method does not rely on the paired detection relationship of left and right wheels. Instead, it introduces the vehicle's travel direction as a unified reference direction to parametrically model the spatial distribution of wheels. Let the travel direction vector of the target vehicle in the image plane be... Then the one-dimensional projection position of the wheel in that direction can be expressed as:

[0073]

[0074] Here, <·,·> represents the vector inner product operation. Through the above projection operation, the two-dimensional wheel position is mapped to a one-dimensional ordered representation along the vehicle's direction of travel.

[0075] S403. Based on the projection position, sort and aggregate the one-dimensional ordered representation, group the wheels belonging to the same axle into a group, thereby dividing the wheel target set into several non-overlapping wheel subsets and constructing an axle-level structure sequence.

[0076] In this step, the system sorts the set of wheel targets according to the projected position of the wheel in the direction of travel:

[0077] The wheels are then grouped together based on the projected distance between adjacent wheels, grouping wheels belonging to the same axle into the same group. Let the distance between the projected positions of two adjacent wheels be:

[0078] when When it is determined that the wheels belong to the same axle, wherein A preset axle spacing threshold is used to distinguish the structural spacing between left and right (or front and rear) wheels within the same axle and between different axles. Based on the above judgment rules, the system divides the sorted set of wheel targets into several non-overlapping wheel subsets, each subset corresponding to one axle, thus constructing an axle-level structural sequence with TruckID k:

[0079] in, This represents the set of wheel targets corresponding to the j-th axle. Each... Let represent a group of wheels corresponding to the same axle, and m represent the total number of axles obtained through aggregation. Therefore, the number of axles of the target vehicle can be directly expressed as:

[0080] S404. Determine the vehicle model based on the axle-level structure sequence.

[0081] Once we obtain the number of axles of the target vehicle, it can help us determine the truck model. The model determination function is defined as follows:

[0082] in The vehicle type label for the target vehicle k: Indicates a two-axle truck model. This indicates a three-axle truck or medium-sized cargo vehicle. This indicates whether the vehicle is a heavy-duty or multi-axle truck. The determined vehicle type information is then automatically output and displayed on the human-machine interface of the weighbridge management system to achieve automatic vehicle type recognition and visual display.

[0083] S405. Generate a parameterized representation of the axle-level structure describing the structural integrity of the target vehicle based on the axle-level structure sequence, as shown below:

[0084] Where m represents the number of axles of the target vehicle. This represents the average projected position of the j-th axle in the direction of travel. The axle-level structural parameters... It is used to characterize the structural integrity of the target vehicle during the weighing process and serves as a structured input for subsequent vehicle status determination, stable weighing status analysis, and weighing rule indexing and invocation.

[0085] As another preferred embodiment, refer to Figure 5 Step S105 specifically includes the following operations: S501. The visual state of the target vehicle during the weighing process is abstracted into discrete state variables. The discrete state variables include the state of the vehicle not entering the weighing area, the state of the vehicle partially entering the weighing area, and the state of stable weighing, represented as:

[0086] in, This indicates that the vehicle has not entered the weighing area. This indicates that the vehicle has partially entered the weighing area, meaning the truck has partially entered the weighing area but has not yet met the stability requirements. This indicates a stable weighing state. The visual state is not determined by a single frame image, but rather by a comprehensive judgment based on the vehicle's structural state and spatial relationship over a continuous period of time.

[0087] S502. Based on the integrity of the axle-level structure, spatial consistency, motion stability, and temporal continuity, the stable vehicle weighing state is defined as a jointly valid state under multiple conditions and temporal constraints. Its decision function can be formally expressed as:

[0088] Among them represents the axle-level structure integrity condition, which is based on the axle-level structure parameters obtained in step S104 It is obtained and defined as follows:

[0089] Among them represents the number of axles detected at the current moment is the preset minimum effective axle number threshold. This condition is used to exclude non-complete structural states caused by occlusion, misdetection, or target absence

[0090] is the spatial consistency condition between the axle-level structure and the weighing area. Assume that the area of the weighing area in the image projection is , then the spatial consistency condition between the axle-level structure and the weighing area is:

[0091] Among them represents the average projection position of the j-th axle in the vehicle traveling direction. This condition ensures that all effective axle-level structures of the truck have completely entered the weighing area, avoiding abnormal situations such as pressing the edge and not fully on the scale

[0092] represents the vehicle motion stability condition, which is obtained based on the target vehicle motion state estimation result in step S102 and can be defined as:

[0093] Among them represents the estimated value of the motion speed of the target vehicle in the image plane is the preset low speed threshold. This condition is used to exclude misjudgments in weighing when the target vehicle is traveling slowly or has not come to a complete stop

[0094]

[0095] Among them represents the preset stable time length represents the continuous establishment of the logical "AND" in the time dimension. This condition is used to describe the physical meaning of the concept of "stable" in the time dimension, making the stable weighing state a continuously established state rather than an instantaneous judgment result

[0096] S503. Determine whether the target vehicle is in a stable weighing state. If not, prohibit the adoption of the weighbridge weighing data

[0097] In this implementation, the system determines that the target vehicle has entered a stable weighing state if and only if all of the above conditions are met at the current time t.

[0098]

[0099] Before a stable weighing state is established, the system continuously classifies the target vehicle as being in an unstable weighing state and prohibits the acceptance of weighbridge weighing data. By modeling the stable weighing state as a joint state constrained by axle-level structure, spatial relationships, motion characteristics, and temporal consistency, this embodiment can effectively distinguish between "short-term fulfillment of weighing conditions" and "true weighing capability," significantly reducing the risk of false triggering caused by vehicles not being fully on the weighbridge, moving slowly, parking close to the edge, or external interference in unattended weighing scenarios in port areas.

[0100] After determining the stable weighing status of the target vehicle in step S105, the system further controls the collection, locking, and legality of the weighbridge weighing data based on the stable weighing status, thereby achieving automatic triggering and reliable constraint of weighing behavior under unattended conditions.

[0101] As another preferred embodiment, refer to Figure 6 Step S106 specifically includes the following operations: S601. When the target vehicle is in a stable weighing state, collect the original weighing data of the weighbridge within a continuous time window to obtain the weighing data sequence.

[0102] In this embodiment, the weighing behavior of the weighbridge is not triggered by manual commands or a single sensor signal, but is uniformly gated based on the stable state of the vehicle on the weighbridge. The weighing trigger condition is defined as follows:

[0103] Only when the vehicle status meets Only when the system is in a certain condition will it allow entry into the weighing and data acquisition process, thus fundamentally avoiding abnormal situations such as incomplete weighing or weighing with the edges pressed down.

[0104] After the weighing trigger condition is met, the system collects the raw weighing data of the weighbridge within a continuous time window, and defines the weighing data sequence as follows:

[0105] in This represents the instantaneous weight value collected by the weighbridge at time i. This indicates the length of the weighing data acquisition window.

[0106] For example, to ensure the stability of the weighing results, this implementation introduces a weight variation constraint:

[0107] in, This is the maximum allowable weight fluctuation threshold. This constraint ensures that the weighing results are derived from a state range where the vehicle is stationary and under stable forces.

[0108] Currently, only when the weighing condition is triggered. With weight stability conditions When both conditions are met, the system locks the duplicated data, and the locking rule is defined as follows:

[0109] in This indicates the final, valid weighing value. This represents a statistical aggregation function used to calculate representative weight values ​​within the weighing window. Before the locking is complete, the system prohibits any weighing data from being output as valid business data.

[0110] S602. Based on the target vehicle model, obtain the maximum permissible weight threshold for the corresponding model from the rule base.

[0111] Based on the vehicle type determined in step S4 The system can obtain the corresponding maximum allowed weight threshold from the updated rule base:

[0112] in This represents the rule base mapping function, which supports dynamic updates based on policy requirements, vehicle type, and port area management rules.

[0113] S603. Based on the maximum permissible weight threshold obtained in the previous step and the weighing data sequence, determine whether the target vehicle is overweight:

[0114] S604. Implement weighing control decisions based on the overweight determination results.

[0115] Finally, based on the overweight determination result, the system executes a weighing control decision, the control logic of which is defined as follows:

[0116] in A value of 1 indicates that the vehicle is allowed to pass through the weighbridge; conversely, a value of 1 indicates that weighing is prohibited and triggers the corresponding alarm or manual intervention process. By unifying weighing triggering, data locking, and overweight determination into a stable weighbridge state gating system, this embodiment achieves closed-loop control of the entire process from "vehicle visual state determination" to "weighing data acceptance" and then to "weighbridge control decision," effectively improving the reliability and anti-cheating capabilities of the unmanned weighbridge system in the port area under complex traffic conditions.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for judging truck weighing in port areas based on deep learning, characterized in that, The method includes: S101. An image acquisition device is set up in the weighing area of ​​the port area to continuously acquire video of vehicles entering the weighing area and obtain a video image sequence. The video image sequence includes multiple original vehicle images. The video image sequence is preprocessed. S102. Identify the target vehicle in the preprocessed video image sequence, perform continuous tracking modeling, and obtain the target region of the image; S103. Based on the target identifier, crop the local image region of the vehicle from the target region of the image; S104. Detect the wheels in the local image region of the vehicle and project them in order along the direction of travel to obtain a sorted wheel set. Based on the sorted wheel set, determine the number of axles and the vehicle type, and generate a parameterized representation of the axle-level structure. S105. Based on the conditions of axle-level structural integrity, spatial consistency, motion stability and temporal continuity, jointly determine whether the target vehicle has entered a stable weighing state. S106. Collect vehicle weighing data under stable weighing conditions, combine vehicle type information and rule base to determine overweight vehicles, and execute weighing control decisions.

2. The method for judging truck weighing in port areas based on deep learning according to claim 1, characterized in that, Step S102 specifically includes the following operations: S201. Input the preprocessed original vehicle image into the deep learning-based target detection model to detect the target vehicle and output the detection result, which includes the candidate image target region in the current frame and the corresponding detection confidence. S202. Assign a unique target identifier to each detected target vehicle based on the detection results; S203. Continuously track and model the spatial position and motion state of the target vehicle in consecutive frames; S204. Based on the continuous tracking results, crop the image target region from the original vehicle image to obtain the target region that corresponds one-to-one with the target identifier.

3. The method for judging truck weighing in port areas based on deep learning according to claim 2, characterized in that, Step S203 specifically includes the following operations: S301. Establish a target vehicle trajectory set for the (t-1)th frame. The target vehicle trajectory set contains multiple vehicle trajectories, and each vehicle trajectory corresponds to a unique target identifier. S302. For any vehicle trajectory, define its state vector at frame t-1 based on the target vehicle center position, target scale parameters, and estimated motion speed. S303. Based on the state vector of the target vehicle in the previous frame, predict the state of the target vehicle in the t-th frame.

4. The method for judging truck weighing in port areas based on deep learning according to claim 3, characterized in that, In step S203, a matching cost function based on spatial overlap is constructed based on the intersection-union function. The matching cost function is used to determine the association probability between the target vehicle and the existing vehicle trajectories. If the association probability is greater than a preset threshold, the state of the corresponding vehicle trajectory is updated, and the target vehicle inherits the target identifier of the associated vehicle trajectory. If the target vehicle is not associated with any of the existing vehicle trajectories, a new vehicle trajectory is created for the target vehicle and a target identifier is assigned. For vehicle trajectories whose state has not been updated for several consecutive frames, the life cycle of the corresponding vehicle trajectory is terminated.

5. The method for judging truck weighing in port areas based on deep learning according to claim 1, characterized in that, In step S103, an image cropping operation is performed based on the target detection bounding box of the target vehicle in the target area of ​​the image to obtain a local image area of ​​the vehicle. The local image area of ​​the vehicle includes the wheel distribution information of the target vehicle and the spatial positional relationship between the target vehicle and the weighing area. The local image area of ​​the vehicle is then bound to the target identifier.

6. The method for judging truck weighing in port areas based on deep learning according to claim 5, characterized in that, Step S104 specifically includes the following operations: S401. Identify the wheels in the local image region of the vehicle to obtain the positional distribution information of the wheels, and further organize them to obtain a set of wheel targets. S402. Based on the position distribution information of the wheels in the wheel target set, perform parametric modeling of the wheel spatial distribution, and map the two-dimensional position distribution information in the wheel target set into a one-dimensional ordered representation along the direction of travel of the target vehicle, so as to determine the projected position of the wheels in the direction of travel. S403. Based on the projection position, sort and aggregate the one-dimensional ordered representation, group the wheels belonging to the same axle into a group, thereby dividing the wheel target set into several non-overlapping wheel subsets and constructing an axle-level structure sequence. S404. Determine the vehicle model based on the axle-level structure sequence; S405. Generate a parameterized representation of the axle-level structure describing the structural integrity of the truck based on the axle-level structure sequence.

7. The method for judging truck weighing in port areas based on deep learning according to claim 1, characterized in that, Step S105 specifically includes the following operations: S501. The visual state of the target vehicle during the weighing process is abstracted into discrete state variables, including the state of the vehicle not entering the weighing area, the state of the vehicle partially entering the weighing area, and the state of stable weighing. S502. Based on the integrity of the axle-level structure, spatial consistency, motion stability and temporal continuity, the stable weighing state of a vehicle is defined as a jointly established state under multiple conditions and temporal constraints. S503. Determine whether the target vehicle is in a stable weighing state; otherwise, the weighing data of the weighbridge shall not be accepted.

8. A method for judging truck weighing in port areas based on deep learning according to claim 6, characterized in that, Step S106 specifically includes the following operations: S601. When the target vehicle is in a stable weighing state, collect the original weighing data of the weighbridge within a continuous time window to obtain the weighing data sequence. S602. Based on the target vehicle model, obtain the maximum permissible weight threshold for the corresponding model from the rule base; S603. Based on the maximum permissible weight threshold obtained in the previous step and the weighing data sequence, determine whether the target vehicle is overweight. S604. Implement weighing control decisions based on the overweight determination results.

9. A method for judging truck weighing in port areas based on deep learning according to claim 8, characterized in that, When collecting the original weighing data of the weighbridge, a weight change constraint is introduced. If the weight change constraint is not met, the weighbridge weighing data will not be accepted. The weighing data sequence obtained by S601 is locked only when the target vehicle is in a stable weighing state and the weight change constraint is met. The weighing data sequence is only output after the weighing data sequence is locked.

Images