A ton bag cargo identification method for port tallying

Abstract

This invention belongs to the field of intelligent warehouse automation detection, specifically a method for identifying ton bag cargo in port cargo handling. The system includes a multimodal acquisition array, a texture noise suppression module, a multi-cluster spatial fusion engine, and a physical posture correction module. It suppresses texture interference through dual spatial and frequency domain filtering, segments adhered targets using normal vector mutation points, and estimates the physical centroid based on point cloud volume integral to optimize the grasping position. This improves the stability and grasping safety of ton bag detection under complex working conditions. By constructing a high-frequency texture noise suppression module and using dual spatial and frequency domain filtering technology, the interference of the woven texture on the surface of the ton bag on edge extraction is eliminated from a physical perspective. This technique effectively solves the problem of recognition frame breakage caused by non-rigid wrinkles, significantly improving the continuity and stability of target recognition.

Description

Technical Field

[0001] This invention belongs to the field of intelligent warehouse automation detection, specifically a method for identifying ton bag cargo used in port cargo handling. Background Technology

[0002] With the deep evolution of industrial and smart logistics systems, automated transfer and warehousing of bulk cargo has become the core driving force for improving operational efficiency. Ton bags, with their advantages of strong load-bearing capacity, light weight, and low turnover cost, have become the core carrier for bulk cargo transportation. Target detection and attitude estimation algorithms based on computer vision are the key to unmanned loading and unloading of ton bags. Their accuracy and real-time performance directly determine the success or failure and safety of the grabbing operation.

[0003] Existing ton bag detection solutions rely on deep learning object detection or traditional edge detection algorithms. They extract edge and texture features of the ton bags using visible light images, combine geometric clustering to lock the contour, and infer the geometric centroid as the grasping benchmark. Under ideal working conditions, they can achieve effective detection and provide basic support for automated statistics and sorting. However, as application scenarios extend to complex industrial sites, existing technologies have deep limitations: high-frequency pseudo-noise generated by the flexible deformation of ton bags can easily lead to recognition errors; deviations between the geometric centroid and the actual physical center of gravity can cause safety hazards; and visual adhesion under high-density stacking makes it impossible to separate the target. The limitations of its single perception dimension and lack of flexible modeling make it difficult to balance accuracy and safety. Therefore, this invention provides a ton bag cargo recognition method for port cargo handling. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows: A method for identifying ton bag cargo in port tallying, comprising a detection system, wherein the detection system is built on the end-sensing platform of an automated loading and unloading mechanism, and the detection system includes:

[0006] A multimodal data acquisition array is used to synchronously acquire the original multidimensional information of the target bag to be detected. The original multidimensional information includes at least a two-dimensional image stream, a three-dimensional spatial point cloud stream, and the instantaneous motion attitude vector of the sensing platform.

[0007] The spatiotemporal reference synchronization module is connected to the multimodal data acquisition array and is used to assign a unified timestamp to the original multidimensional information through a hardware-level triggering mechanism, and to align the multimodal data in a spatial coordinate system based on the instantaneous motion attitude vector.

[0008] The high-frequency texture noise suppression module is used to receive the raw two-dimensional image output by the industrial camera. It adopts a dual architecture that combines spatial domain guided filtering and frequency domain notch filtering. By identifying and filtering out the high-frequency energy peaks that represent the weaving density distribution of the ton bag, it suppresses texture pseudo-noise on non-rigid surfaces.

[0009] The multi-cluster spatial fusion engine is used to map noise-suppressed image features and point cloud data acquired by LiDAR into a unified voxel space, perform multi-dimensional clustering based on semantics, deep connectivity and normal vector consistency, and use normal vector mutation points as physical segmentation benchmarks to cut off the visual adhesion between adjacent ton bags.

[0010] The dynamic physical attitude correction module is used to extract the point cloud clusters of individual ton bags after segmentation, estimate the mass density distribution of the material inside the bag by the point cloud volume integral and echo intensity distribution, calculate the physical center of gravity of the ton bag in three-dimensional space, and project the physical center of gravity onto the surface of the bag along the gravity vector direction to determine the optimized gripping center point and operation normal vector.

[0011] The central control and task scheduling system is used to integrate the detection reports output by each module, generate comprehensive instructions including target UID, three-dimensional spatial pose, physical diameter and grasping confidence, and issue them to the automated loading and unloading execution mechanism.

[0012] Preferably, the multimodal data acquisition array is fixedly installed on the end flange or sensing beam of the loading and unloading actuator, and includes no less than two sets of industrial cameras, one set of multi-line pulse lidar, and one set of embedded inertial measurement modules; the industrial cameras adopt global shutter CMOS sensors with an effective pixel count of no less than 20 million, and are equipped with low-distortion industrial lenses with shockproof structures, for capturing original images of ton bags and acquiring high-definition color texture information and boundary contrast features of the ton bag surface; the multi-line pulse lidar is set on the geometric center axis of the camera array, and its ranging principle is based on the time-of-flight method, with a horizontal field of view of no less than 120 degrees, a vertical resolution of no less than 0.1 degrees, and a ranging accuracy of ±1 cm, for constructing a three-dimensional spatial point cloud map of the ton bag stack in real time; the embedded inertial measurement module includes a three-axis gyroscope and a three-axis accelerometer, for sensing the instantaneous attitude angle and vibration vector of the platform in real time; the protective cover of the industrial cameras is made of tempered glass, and the surface is coated with a nano-level superhydrophobic self-cleaning coating.

[0013] Preferably, the spatiotemporal reference synchronization module includes a clock source and a hardware trigger distributor. The hardware trigger distributor is electrically connected to the industrial camera and LiDAR via a dedicated synchronization cable. This module is driven by a field-programmable gate array (FPGA) to generate synchronization trigger pulses with microsecond-level precision, ensuring that the exposure start time of the industrial camera is aligned with the scan start time of each revolution of the multi-line pulse LiDAR on the same time scale. At the same time, this module timestamps the frequency information output by the inertial measurement module with the image and point cloud data streams, achieving strict overlap of sensor data on the time axis and eliminating motion distortion caused by the movement of the actuator.

[0014] Preferably, the high-frequency texture noise suppression module is located in the preprocessing pipeline of the central control system. This module receives raw image data from an industrial camera. For high-frequency pseudo-noise generated by the woven texture on the surface of the ton bag, the module adopts a dual suppression architecture combining spatial domain low-pass filtering and frequency domain notch filtering. In the spatial domain, the module uses an adaptive guided filter to smooth the image. This process suppresses texture noise while preserving the true physical edge features of the ton bag by analyzing the local gradient direction. In the frequency domain, the module performs a fast Fourier transform on the image to identify and filter out high-frequency energy peaks representing the woven density distribution, thereby eliminating local contrast abrupt changes caused by wrinkles and preventing logical breaks in the target recognition box during subsequent processing. Dynamic exposure compensation logic: monitors the brightness distribution of the shadow area of ​​the ton bag wrinkles in the image in real time, and enhances the extraction of regional features by adjusting the exposure time and gain parameters of the industrial camera.

[0015] Preferably, the multi-cluster spatial fusion engine is the core recognition logic of this system. The engine first aligns the spatial coordinates of the noise-suppressed 2D image pixels with the 3D point cloud data using a pre-stored calibration parameter matrix. The engine executes a multi-dimensional spatial clustering logic: the first dimension is semantic clustering based on color and contrast, used to initially divide the potential regions of the ton bag targets; the second dimension is spatial connectivity clustering based on depth gradient, used to identify entities with continuous geometric characteristics in the point cloud; the third dimension is curvature clustering based on surface normal vector consistency. The engine fuses the clustering results of the above three dimensions in voxel space through a weighted voting mechanism. For the visual adhesion caused by adjacent ton bags, the engine uses the normal vector mutation points in the curvature clustering results as the segmentation benchmark to forcibly cut off the pixel connections that appear to be adhered in the 2D image, thereby achieving accurate individual segmentation of tightly stacked ton bags.

[0016] Preferably, the dynamic physical attitude correction module is used to extract the real working position from the segmented ton bag target. This module not only obtains the geometric center of the target, but also uses the point cloud distribution density captured by the lidar to estimate the mass distribution of the material inside the bag. This module projects the point cloud cluster of the target onto the gravity coordinate system, calculates the volume weight of each region of the point cloud and combines it with the geometric envelope features of the target to deduce the physical center of gravity of the ton bag in three-dimensional space.

[0017] This module combines the instantaneous platform attitude feedback from the inertial measurement module to perform vector correction on the center of gravity shift caused by bag deformation;

[0018] This module outputs an optimized grab center point and its normal vector direction. The center point is located in the vertical projection area of ​​the physical center of gravity on the top surface of the bag, ensuring the balance of force at the moment of lifting and preventing the ton bag from shaking or slipping.

[0019] Preferably, the central control and task scheduling system is responsible for the logical control and data flow management of the entire system. This system adopts an industrial-grade multi-core processor architecture, runs a real-time operating system, and is responsible for receiving the processing results of each module, generating a comprehensive detection report including the target UID, three-dimensional spatial pose, physical diameter, estimated weight, and grasping confidence, and issuing instructions to the underlying lifting actuators through an industrial Ethernet interface. The central control and task scheduling system uses an industrial-grade multi-core processor architecture to execute tasks in parallel, with the perception data synchronization task having real-time priority, and the target detection and clustering fusion tasks being executed in parallel through a pipelined manner of the computing core, and the detection latency of a single target being controlled within 100 milliseconds. The detection system also integrates a watchdog module to monitor the heartbeat status of the algorithm process in real time and execute a logical restart when an anomaly is detected.

[0020] Preferably, the detection system further includes meteorological compensation logic and a clustered collaborative interface, wherein:

[0021] The meteorological compensation logic dynamically adjusts the reflection intensity gain of the lidar based on the visibility parameters fed back by the environmental sensors, and calls the long-wave infrared enhancement algorithm to perform fog-penetrating processing on the images acquired by the industrial camera.

[0022] The clustered collaborative interface supports real-time synchronization of cargo handling data, center of gravity correction vector diagrams, and capture execution parameters to the cloud management platform via 5G network, so as to realize pose data sharing and collaborative operation among multiple loading and unloading equipment.

[0023] Preferably, in the multimodal data acquisition array, the mounting bracket of the industrial camera is made of aerospace-grade aluminum alloy and the surface is treated with hard anodizing. The bracket has an integrated active cooling channel, which ensures the thermal stability of the sensor under high-intensity operation through circulating coolant. The protective cover of the industrial camera is made of special tempered glass with self-cleaning function and coated with a superhydrophobic coating to prevent dust and oil stains in the industrial site from affecting the imaging quality.

[0024] Preferably, the high-frequency texture noise suppression module also includes a dynamic exposure compensation logic. This logic monitors the brightness distribution of the shadow area of ​​the ton bag wrinkles in the image in real time. By controlling the exposure time and gain parameters of the industrial camera, it enhances the extraction of shadow area features and ensures that the boundary features of the ton bag still have sufficient contrast under strong direct light or shadow coverage.

[0025] Preferably, the multi-cluster spatial fusion engine also integrates a non-rigid model matching operator. This operator pre-stores a set of flexible deformation topology templates of standard ton bags under different load conditions. During the clustering process, the operator matches the detected edge features with the templates and corrects the missing geometric boundaries caused by local collapse by calculating the deformation tensor, thereby improving the detection robustness of the system under extreme deformation conditions.

[0026] Preferably, the dynamic physical attitude correction module also includes a safety verification logic. After determining the gripping center point, the module analyzes the flatness of the point cloud and the dispersion of the surface normal vector around the gripping point. If a serious risk of surface tearing or extreme unevenness of the normal vector is detected in the gripping area, the system will automatically search for a suboptimal robust gripping point in the global range and send a safety warning signal to the central control system.

[0027] The beneficial effects of this invention are as follows:

[0028] 1. The present invention provides a method for identifying ton bag cargo in port cargo handling. By constructing a high-frequency texture noise suppression module and utilizing dual filtering technology in the spatial and frequency domains, the interference of the woven texture on the surface of the ton bag on edge extraction is eliminated from a physical perspective. This technique effectively solves the problem of recognition frame breakage caused by non-rigid wrinkles and significantly improves the continuity and stability of target recognition.

[0029] 2. The ton bag cargo identification method for port tallying described in this invention overcomes the visual adhesion between ton bags in complex environments through three-dimensional collaborative clustering of depth, color and normal vectors. It uses the mutation of normal vectors as a physical segmentation benchmark, so that the system can still maintain extremely high target separation accuracy in high-density stacking scenarios, providing clear operational targets for automated loading and unloading.

[0030] 3. The ton bag cargo identification method for port tallying described in this invention marks a leap from geometric perception to physical perception through the introduction of a dynamic physical posture correction module. By estimating the physical center of gravity of the flexible target rather than relying solely on the geometric center, this invention ensures the mechanical balance of the grasping operation, reduces the risk of swaying, slipping, and tearing of the ton bag during loading and unloading, and greatly improves the operational safety of the automated system.

[0031] 4. The ton bag cargo identification method for port tallying described in this invention achieves full-process spatiotemporal reference synchronization through this system. Through hardware-level triggering mechanisms and inertial motion compensation, it ensures data consistency during dynamic movement. This spatiotemporal alignment provides a high-quality raw data foundation for subsequent complex algorithm fusion, ensuring the system's high robustness under all weather and operating conditions.

[0032] 5. The ton bag cargo identification method for port tallying described in this invention, through the adoption of an industrial-grade ruggedized architecture and redundant design, has extremely strong environmental adaptability and anti-interference capability, from sensor protection design to software-level error prevention mechanism. Attached Figure Description

[0033] The invention will now be further described with reference to the accompanying drawings.

[0034] Figure 1 This is a structural block diagram of a method for identifying ton bag cargo used in port tallying according to the present invention. Detailed Implementation

[0035] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0036] like Figure 1 As shown in the embodiment of the present invention, a method for identifying ton bag cargo for port tallying is presented. The entire system is physically manifested as a highly reinforced sensing terminal, which is installed on the end flange of the lifting equipment or a dedicated sensing beam. Its structural strength and vibration resistance meet the highest industrial-grade standards.

[0037] In terms of hardware configuration of the perception layer, the multimodal data acquisition array used in this invention includes two sets of industrial cameras with global shutters. The pixel scale of their photosensitive elements reaches more than 20 million, ensuring the integrity and edge sharpness of image acquisition during high-speed dynamic operations. The two sets of industrial cameras are installed in an asymmetrical convergence angle layout, and the optical axis intersection point is set in the preset core area of ​​the operation. This design can improve the feature resolution accuracy of the central area through parallax enhancement while ensuring a wide search field of view.

[0038] The accompanying lens is a low-distortion industrial lens with a shockproof structure. Its internal lens group uses a high-hardness calcium fluoride coating, which can effectively resist temperature fluctuations and wear from tiny particles in harsh working environments.

[0039] In the selection of multi-line pulse lidar, the system adopts an industrial-grade lidar based on the time-of-flight ranging principle, with a vertical resolution of up to 0.1 degrees. It can complete the dense reconstruction of the three-dimensional point cloud on the surface of the ton bag in a very short time. The multi-line pulse lidar is installed on the geometric center axis of the camera array to ensure that its scanning line distribution coincides with the image field of view to the maximum extent.

[0040] The embedded inertial measurement module includes a high-performance three-axis gyroscope and accelerometer. Through microelectromechanical systems (MEMS) technology, it can capture the instantaneous dynamic response of the sensing platform in six degrees of freedom in real time, providing underlying data support for subsequent motion compensation.

[0041] To ensure the physical protection and environmental adaptability of the sensing array, the mounting bracket of the industrial camera is made of aerospace-grade aluminum alloy. The surface corrosion resistance is improved through hard anodizing treatment. The bracket is equipped with active cooling microchannels through precision machining. The circulating coolant removes the high-frequency heat generated by the photosensitive element, ensuring the thermal stability and signal-to-noise ratio of the sensor under high load operation.

[0042] The protective cover of the sensing platform is made of special tempered glass with self-cleaning function and coated with a nano-level superhydrophobic coating, which can spontaneously use the gravity of droplets to remove attached dust and oil.

[0043] In low visibility or nighttime conditions, the near-infrared supplementary light group integrated in the array will automatically activate. The 850-nanometer near-infrared light emitted by the light can penetrate light smoke and dust, and the strong reflective properties of the ton bag material on specific wavelengths will further enhance the edge contrast of the target.

[0044] The spatiotemporal reference synchronization module, driven by a high-performance field-programmable gate array, generates microsecond-level pulse sequences through an internally built hardware timer. The hardware trigger distributor distributes the trigger signals to each sensor through a dedicated impedance-matched cable. The exposure start time of the industrial camera and the scanning start period of the multi-line pulse lidar are strictly locked at the hardware level, ensuring that the texture information in each frame of the image and its corresponding point cloud features have complete overlap on the time axis. The spatiotemporal reference synchronization module performs high-frequency sampling on the output of the inertial measurement module and uses differential timestamp logic to inject the platform attitude parameters into the data frame header of the image and point cloud in real time. This mechanism eliminates the geometric mapping distortion caused by asynchronous sampling of sensors during the translation or rotation of the crane mechanism.

[0045] Before the data stream enters the core processing flow, the high-frequency texture noise suppression module performs dual filtering on the original image. Due to the polypropylene weaving process, the surface of the ton bag has highly repetitive micro-textures, which appear as high-frequency periodic noise in the image and can easily interfere with the convergence of the edge detection algorithm. This module first performs adaptive guided filtering in the spatial domain. This algorithm is based on a local linear model and automatically adjusts the smoothing coefficient according to the image gradient within the calculation window. In areas with drastic gradient changes, such as the ton bag boundary, the filter retains the true physical contour by locking the linear slope; while in areas with woven texture distribution, the filter smooths the noise by reducing the weight.

[0046] The system transforms the image to the frequency domain and accurately identifies the characteristic frequency peaks corresponding to the weaving density by analyzing the two-dimensional power spectral density.

[0047] The notch filter operator automatically masks these peaks in the frequency domain, thereby eliminating local pseudo-features caused by surface wrinkles and ensuring the logical continuity of the bag outline in subsequent processing. The dynamic exposure compensation logic integrated in this module adjusts the gain parameters of the photosensitive element in real time according to the image histogram distribution, and enhances the detail of the wrinkle shadow area through multi-frame fusion technology in high contrast environments.

[0048] The multi-cluster spatial fusion engine first uses pre-stored camera intrinsic and extrinsic parameter matrices to spatially register the processed image pixels with the 3D point cloud of the multi-line pulse lidar, constructing an enhanced dataset in which each voxel contains color components, reflection intensity, depth value and surface curvature. The engine uses a weighted voting mechanism to execute the three-dimensional clustering task in parallel.

[0049] In the semantic clustering dimension, the system uses the distance metric of the Lab color space to perform region growing on the ton bag body and divide the target color consistency region. In the spatial connectivity clustering dimension, the system aggregates the point cloud based on the Euclidean distance threshold to identify physical entities with geometric continuity.

[0050] In the normal vector consistency clustering dimension, the engine extracts the principal normal vector direction of the surface by calculating the eigenvalues ​​of the covariance matrix of the local neighborhood point cloud.

[0051] To address the visual adhesion problem that easily occurs in scenarios where ton bags are tightly stacked, the multi-cluster spatial fusion engine performs forced segmentation by analyzing the dispersion of the normal vector at the cluster boundaries;

[0052] When two adjacent ton bags appear as color continuity in a two-dimensional image, the engine will detect the curvature change at their interface in three-dimensional space. If the rate of change of the angle between the normal vectors exceeds the preset threshold, the algorithm will automatically insert a physical segmentation surface at that point to forcibly truncate the connected pixels. This segmentation logic based on physical geometric topology can accurately distinguish between two non-rigid targets that are close to each other.

[0053] Furthermore, the non-rigid model matching operator integrated within the engine calls the pre-stored flexible topology template and automatically repairs the contour loss caused by local collapse or occlusion by calculating the deformation tensor of the target, significantly improving the system's ability to capture ton bags with extreme deformation.

[0054] The dynamic physical attitude correction module realizes the leap from simple geometric recognition to deep physical perception. As a flexible container for loading bulk materials, the physical center of gravity of the ton bag often shifts significantly with the flow and accumulation of materials inside the bag. Traditional geometric center gripping schemes are prone to causing the bag to become unstable at the moment of lifting.

[0055] This module extracts the segmented point cloud clusters of the ton bag and performs mass distribution estimation based on volume integral. The system divides the envelope space of the ton bag into several tiny modules and uses the spatial distribution of the echo intensity of the multi-line pulse lidar to estimate the local density of the material. Areas with higher echo intensity usually correspond to parts with dense material packing and high surface reflectivity. Based on this, the system assigns different weight coefficients to each volume module. By summing the weighted centroids of all volume modules, the system calculates the true physical center of gravity of the ton bag in its current posture.

[0056] The correction module projects the physical center of gravity along the current gravity vector direction onto the working plane at the top of the ton bag to determine the theoretically optimal gripping point.

[0057] During this process, the system will combine the instantaneous attitude of the sensing platform fed back by the inertial measurement module to compensate for the dynamic components caused by equipment vibration. In order to ensure the safety of the grasping operation, a layer of safety verification logic is set in the module. This logic will conduct a fine evaluation of the surface flatness around the pre-selected grasping point.

[0058] If severe wrinkles or tilts or extremely discrete normal vectors are detected in the grasping area, which may cause the lifting device to slip or the bag to tear, the system will automatically search for a suboptimal robust area with a gentler slope and a denser point cloud distribution along the normal. The module outputs an optimized grasping vector containing three-dimensional coordinates and the direction of its normal vector, guiding the actuator to cut in at the most stable angle.

[0059] The central control and task scheduling system runs on an industrial-grade real-time operating system and is responsible for the closed-loop management of sensing, processing, and execution instructions. The system adopts a multi-core processor redundancy architecture, which allocates different tasks to independent processing cores. Sensing data acquisition and synchronization tasks are given the highest priority, and real-time data throughput is guaranteed through a hardware interrupt mechanism. The core algorithm module is executed in a pipelined parallel manner on an independent computing core to ensure that the detection latency is stable at the level of hundreds of milliseconds. The system's built-in watchdog module monitors the heartbeat of each algorithm process at the second level. Once an abnormal state is detected, it immediately executes on-site protection and logic restart. The sorting results are distributed through a high-speed industrial Ethernet interface. The data packets contain the target's globally unique identifier, the corrected three-dimensional pose, the estimated weight distribution, and the grasping confidence score.

[0060] The central system supports clustered operation mode through 5G communication modules. Environmental features sensed by multiple devices can be integrated in real time in the cloud to build a dynamic digital twin of the entire work site.

[0061] Furthermore, to cope with complex weather conditions around the clock, the system integrates weather compensation logic at the algorithm layer. When the environmental sensor detects that rainfall or dense fog has reduced atmospheric transmittance, the system will automatically increase the reflection energy threshold of the multi-line pulse lidar and activate the image fog penetration enhancement operator. Through the dynamic redistribution of multimodal feature weights, the system maintains the consistency of detection accuracy.

[0062] Furthermore, the system's sensing platform layout employs a precisely designed asymmetrical perspective. The baseline distance between the two sets of industrial cameras is dynamically set according to the operating range, typically around 500 mm. This layout provides extremely high parallax resolution in the near field, which helps to accurately model the lifting lugs or folding details on the top of the ton bag.

[0063] In the far field, the long-range detection capability of the multi-line pulse lidar is used for large-scale coarse positioning. This asymmetry of perspective and the complementarity of sensor characteristics constitute the system's high-performance performance across the entire range.

[0064] In addition, the system's adaptive learning framework can record the physical feedback data of each operation. By analyzing the deviation between the real-time torque feedback from the crane mechanism and the system's predicted center of gravity, the system will automatically correct the internal material density distribution model. This closed-loop self-evolution mechanism enables the system to more and more accurately grasp the stacking characteristics of different types of materials in the ton bag during long-term operation, thereby continuously optimizing the success rate of center of gravity prediction.

[0065] In the engineering details of the system, the structural components of the sensing platform are equipped with specially designed damping pads at the connection points, which can filter out specific frequency harmonics generated by the operation of the lifting equipment. All signal transmission cables adopt a highly flexible double shielding mesh structure, and are equipped with IP68-level industrial connectors with threaded locking function, ensuring that the electrical performance of the system does not degrade in port environments with severe salt spray corrosion or in dusty chemical plant environments.

[0066] In terms of internal logic processing, the voxelization fusion stage adopts a data structure based on sparse octrees, allocating computing resources only to voxel nodes with property weight distribution. This optimization significantly reduces the computational complexity of spatial clustering, enabling complex geometric operations to run smoothly on embedded computing platforms.

[0067] In actual loading and unloading operations, the central control system automatically adjusts the sampling frame rate of the sensors according to the current task priority. During the cruise phase of searching for the target, the system operates in a low frame rate and wide field of view mode to save power consumption. During the precise grabbing and alignment phase, the system instantly switches to full frequency mode to provide the actuator with millisecond-level closed-loop feedback. This dynamic resource scheduling mechanism greatly improves the overall energy efficiency ratio and response speed of the system.

[0068] Based on the above, the implementation logic of this system is as follows:

[0069] In terms of hardware physical layer implementation, the system adopts a ruggedized embedded computing platform as the computing power hub. This platform connects to external sensors through redundant M12 industrial interfaces to ensure that signal transmission is not interrupted in environments with strong mechanical vibration, such as ports and chemical plants. The system's power supply module is designed with a wide voltage input range and multi-level electromagnetic compatibility filtering, which can effectively isolate surge interference from high-power hoisting motors.

[0070] In terms of the logical flow of data stream processing, after the system starts up, it first enters the device self-test and calibration alignment state. The hardware trigger distributor sends the starting pulse to the industrial camera and multi-line pulse lidar. Multimodal data begins to flow into the memory buffer synchronously at a fixed frame rate, such as 20Hz. Each frame of data carries a nanosecond-level timestamp allocated by the time-space reference synchronization module and the initial spatial pose value provided by the inertial measurement module.

[0071] In the specific execution of high-frequency texture noise suppression, the system converts the original RGB image into a luminance component map. The low-pass filter operator slides on the image at a set spatial scale to extract low-frequency information representing the main outline of the ton bag. At the same time, the frequency domain notch operator accurately locks and eliminates the periodic interference signal generated by the texture based on the inherent spatial frequency characteristics of the ton bag woven fabric. The processed image feature map and point cloud data are mapped in a unified spatial coordinate system to form an enhanced point cloud in which each voxel contains color, intensity and depth information.

[0072] In the operation phase of the multi-cluster space fusion engine, the system first uses a clustering algorithm based on normal vector consistency to aggregate point cloud points with similar surface orientations into different face elements. For targets with curved surface features, such as ton bags, the engine will identify the connectivity between face elements and combine the ton bag class probability map output by image semantic segmentation to merge discrete face elements into complete ton bag individuals.

[0073] During this process, if two ton bags appear to stick together, the engine will identify abrupt changes in depth and abnormal curvature at the sticking point. By setting a spatial segmentation surface at the sticking interface, the two entities will be forcibly separated. This segmentation method based on physical geometric constraints is far more accurate than single two-dimensional pixel segmentation.

[0074] During the calculation of dynamic physical attitude correction, the system extracts the complete point cloud cluster of each individual ton bag. The correction engine analyzes the distribution of the point cloud cluster in three-dimensional space and calculates its second moment about the center of mass to determine the direction of the object's principal axis. Considering that the ton bag is a flexible carrier, its center of gravity will change with the material accumulation state. The system uses the point cloud volume integration method to divide the bag into multiple small modules, estimates the mass density based on the reflection intensity of each module, and then synthesizes the physical center of gravity position of the target. The correction module then projects this center of gravity position onto the bearing plane on the top surface of the bag in the direction of gravity to calculate the optimal working position.

[0075] If the location deviates from the geometric center by more than a preset threshold, the system will automatically adjust the approach vector of the gripping mechanism to make the gripping axis coincide with the physical center of gravity axis.

[0076] The central control and task scheduling system performs closed-loop monitoring throughout the entire operation cycle. After each target detection is completed, the system will evaluate the current ambient light stability, sensor noise level, and confidence level of the detection results.

[0077] If the confidence level is lower than the safety threshold, the system will instruct the actuator to make a small-scale pose adjustment and obtain data from multiple perspectives for redundancy verification. All cargo handling data, pose correction process and operation images are encrypted and stored in the onboard solid-state disk and synchronized to the background database in real time through an encrypted link.

[0078] Furthermore, when the rain and fog sensor detects a decrease in visibility, the system automatically increases the gain of the multi-line pulse lidar reflection intensity and calls the long-wave infrared enhancement algorithm to process the image through fog. Through the dynamic weight allocation of multi-modal data, the system maintains the consistency of detection accuracy.

[0079] Furthermore, the multi-cluster spatial fusion engine also possesses self-learning capabilities. During long-term operation, the system collects capture site samples after manual intervention and continuously corrects the clustering weight allocation strategy through an internal parameter optimization mechanism, enabling it to better adapt to ton bags with different materials, coatings, and deformation degrees.

[0080] In terms of the physical layout of the sensing layer, the baseline distance between the two sets of industrial cameras is dynamically configured according to the working height. If the working height is above 5 meters, the baseline distance is set to 400 mm to 600 mm to ensure sufficient parallax range for assisting depth verification.

[0081] The camera lens uses a glass lens with a hard coating, which can resist chemical corrosion and mechanical scratches in industrial environments. The installation position of the multi-line pulse lidar is precisely calculated to ensure that its scanning lines can cover the entire field of view of the camera, achieving feature fusion without blind spots.

[0082] In the preprocessing stage of the software logic, the spatial domain guided filtering executed by the system adopts a local linear model. This model assumes that there is a linear relationship between the filtered output and the guided image within a local window. By solving the linear coefficients within the local window, the output image can accurately lock the edge position in areas with drastic gradient changes, such as the ton bag boundary, while maintaining global smoothness, without being disturbed by small-scale fluctuations in the woven texture. The frequency domain suppression processing automatically identifies the characteristic frequencies representing the repeating pattern of the woven mesh by analyzing the power spectral density of the image, and removes them by setting a notch filter. This process does not rely on manual intervention at all and has strong adaptability.

[0083] In the implementation logic of the clustering fusion engine, the system first projects the point cloud data onto a two-dimensional grid map. Each grid cell stores the maximum height, minimum height, average reflection intensity, and surface roughness of the point cloud in that area. The clustering algorithm first searches for jump lines with height changes in the grid map as preliminary candidate dividing lines.

[0084] The engine calls the 3D surface extraction program and uses the eigenvalues ​​of the covariance matrix in the local neighborhood to evaluate the surface flatness. For regions with continuously changing curvature, the algorithm marks them as components of the same target surface. When encountering interfaces with discontinuous curvature or normal vector angles exceeding a preset threshold, such as 30 degrees, the algorithm automatically truncates the clusters at these points, thereby achieving physical separation of adjacent ton bags.

[0085] The center of gravity estimation logic of the physical attitude correction module is based on the principles of mass conservation and volume integration. The system reconstructs the closed surface of the segmented ton bag point cloud and calculates the total volume within its closed envelope.

[0086] Since the materials inside the ton bag often exhibit non-uniform stacking during loading, the system analyzes the spatial distribution of the echo intensity of the multi-line pulse lidar to identify differences in material density. High-intensity echoes typically correspond to areas with dense material stacking and strong surface reflection. Based on this, the system assigns different mass weight coefficients to the volume modules of different areas. By summing the weighted centroids of all volume modules in the bag, the system obtains the physical center of gravity coordinates of the target.

[0087] To ensure that the corrected gripping point is actually operable, the correction module executes a virtual contact simulation logic. This logic constructs a circular evaluation area with a diameter equal to the opening degree of the gripping mechanism around the pre-selected gripping point. The system statistically analyzes the consistency of the point cloud normal vectors and the tilt angle relative to the horizontal plane within the evaluation area.

[0088] If the tilt angle is too large, such as exceeding 45 degrees, it means that the gripping mechanism may slip. The system will search for a flat area with a gentle slope in the neighborhood along the gravity projection line as the final location.

[0089] Under the management of the central control system, all processing tasks are assigned to different computational priority queues. Sensor driving and data synchronization tasks are given real-time priority to ensure that data acquisition is not lost; target detection and clustering fusion tasks are given high priority to meet the real-time feedback requirements of the job, with detection latency controlled within 100 milliseconds; while log storage and cloud communication tasks are given low priority and are processed during computational gaps.

[0090] The system also includes a watchdog module that monitors the running status of each algorithm process in real time. Once an infinite loop or abnormal crash is detected, the system can restart the process and restore its state within 10 milliseconds.

[0091] Furthermore, the system also integrates a set of near-infrared light sources in the data acquisition array, with a wavelength set at 850 nanometers. In nighttime or extremely low light environments, this light source, in conjunction with the infrared light-sensing capability of the industrial camera, can provide more stable ton bag boundary features than visible light, because the polypropylene material of most ton bags has good reflective properties for near-infrared light, which can produce clear edge contrast.

[0092] Furthermore, the multi-cluster spatial fusion engine maintains a short-term temporal memory during processing. When the ton bag is momentarily occluded during movement, the system extracts the target parameters from the previous frame from the memory and combines them with the current motion vector to predict the pose, thereby ensuring the continuity of target tracking and preventing detection loss due to momentary occlusion.

[0093] Furthermore, the attitude correction module of this system also includes a load dynamic balancing algorithm. When the gripping mechanism contacts the ton bag, the system will use the torque sensor to provide real-time feedback on the actual force situation and compare it with the estimated physical center of gravity.

[0094] If there is a deviation between the actual center of force and the estimated center, the system will fine-tune the lifting speed and angle, and make up for the posture through dynamic force feedback, so as to achieve the most stable lifting control for flexible goods.

[0095] The technical solution proposed in this invention is based on solving the sensing problem of flexible ton bags in complex industrial environments. Through in-depth analysis of non-rigid features and rigorous integration of spatial multi-dimensional features, a detection system with high robustness, accuracy and safety is constructed. The implementation of this system can significantly improve the automation rate of bulk material transfer, reduce the occurrence of safety accidents, and provide core technical empowerment for the digital transformation of modern smart ports and large manufacturing enterprises.

[0096] In the engineering details of the system, the sensor layout of the perception platform adopts an asymmetrical layout scheme. Two sets of industrial cameras are installed at a small convergence angle, and their optical axes intersect at a distance of 5 meters. This layout not only can acquire operation images with a large field of view, but also provide a higher overlap rate in the central operation area, thereby improving the accuracy of depth calculation. The multi-line pulse lidar adopts a downward tilted installation angle to ensure that its scanning line can obliquely scan the top surface and side walls of the ton bag, obtaining richer lateral contour information, which is crucial for evaluating the stacking level and lateral gap of the ton bags.

[0097] The system's cable management adopts a drag chain protection design. All signal and power lines are made of highly flexible shielded cables that are resistant to bending and oil contamination. The industrial-grade connectors at the cable ends are equipped with threaded locking mechanisms and have an IP68 waterproof rating to prevent signal attenuation or short circuits in port environments with high humidity and severe salt spray corrosion.

[0098] Inside the computing core, in order to improve the efficiency of voxel fusion, the system adopts a data structure based on sparse voxels and allocates memory space only for spatial regions with point cloud distribution. This optimization measure reduces the system's memory usage by more than 70%, enabling complex spatial clustering algorithms to achieve millisecond-level fast response on embedded platforms.

[0099] The system utilizes the parallel computing capabilities of multi-core CPUs to distribute clustering tasks of different dimensions to different processing cores for parallel execution, maximizing the use of hardware performance.

[0100] In the system's adaptive learning framework, the system records the success or failure of each capture operation and the corresponding environmental parameters, such as light intensity, wind force, and the degree of deformation of the ton bag. Through the built-in data mining logic, the system will automatically analyze the common characteristics of failed cases and adjust the confidence judgment threshold in the clustering fusion engine accordingly.

[0101] For example, in open-air storage yards with strong winds, the system will automatically increase the sensitivity of the center of gravity correction and increase compensation for the sway of the bag, thereby ensuring that the system can achieve self-evolution and precision optimization in constantly changing environments.

[0102] This invention provides a method for identifying ton bag cargo in port cargo handling. Through dual physical and digital modeling, it achieves high-dimensional perception and control of non-rigid flexible targets. Every technical indicator of the system has been rigorously verified through engineering experiments and can withstand long-term testing in harsh industrial environments.

[0103] To enhance the system's anti-interference capability, this invention introduces a dynamic weighting mechanism based on sensor confidence in the multi-source feature deep fusion module. In areas where the camera sensor is overexposed due to direct strong light, the system automatically reduces the weight of visible light features, relying mainly on the geometric features of the multi-line pulse lidar for target locking. In areas where the multi-line pulse lidar encounters sparse point clouds due to multipath reflection or strong absorption, such as black bags or water surfaces, the system automatically increases the weight of image features under infrared auxiliary light. This cross-modal intelligent compensation ensures that the system can still output continuous and reliable detection data under extreme physical conditions.

[0104] At the output end of the attitude correction module, the system also designed a virtual operation envelope verification layer. Before outputting the final operation position, the verification layer will simulate the motion trajectory of the grasping mechanism in the current attitude and detect whether it will spatially interfere with the surrounding obstacle bags or other equipment.

[0105] If a collision risk is detected, the system will automatically plan an obstacle avoidance path or correct the operating angle. This technical architecture, which deeply integrates perception, decision-making and safety verification, is an important guarantee for realizing automated loading and unmanned operation of ton bags.

[0106] Ultimately, through a central control and task scheduling system, this system achieves a digital closed loop for cargo handling data. Each ton bag's cargo handling record includes a real-time multimodal panoramic view, a center of gravity correction vector map, and the final capture execution parameters. These multi-dimensional data records not only provide accurate real-time data for the company's logistics management but also provide massive amounts of high-value raw data support for subsequent process analysis and operation optimization.

[0107] This system demonstrates strong adaptability when dealing with ton bags of different sizes, such as 500kg, 1000kg, and 2000kg. By acquiring global contour information through multi-line pulse lidar, the system can automatically identify the size and scale of the current target and call corresponding physical constants from the parameter library, such as the reference value of the elastic modulus of the bag material and the standard distribution model of the material density, to participate in the center of gravity correction calculation. This intelligent parameter calling mechanism based on target perception enables the system to maintain extremely high detection accuracy when dealing with mixed stacking scenarios.

[0108] In addition, this invention also designs a redundant hot standby architecture to meet the continuous operation requirements in industrial production. The central control and task scheduling system in the system consists of two symmetrical computing nodes, master and slave. The master node outputs the detection results in real time, and the slave node calculates synchronously in the background and performs consistency verification on the results of the master node.

[0109] If the master node experiences a hardware failure or a computation timeout, the system can switch to a slave node within milliseconds, ensuring the continuity of the workflow and preventing huge economic losses caused by downtime for repairs.

[0110] In addition to supporting standard industrial Ethernet protocols such as Profinet and EtherCAT, the system's external communication interface also integrates a 5G communication module.

[0111] In large yards, multiple tallying and inspection systems can work collaboratively through low-latency, high-bandwidth 5G networks, sharing positional data and environmental information at each point, thereby constructing a complete clustered intelligent tallying and sensing network. This networked deployment capability further expands the application prospects of this invention to the field of large-scale, group-coordinated intelligent logistics scheduling.

[0112] Example 1: Automated sorting and picking operation in a chemical plant's ton bag warehouse

[0113] In this embodiment, the ton bags are standard 1000kg polypropylene woven bags, filled with highly fluid industrial granular materials. The ton bags are stored in a semi-open warehouse in a 3-layer stacked manner. The sensing platform is installed under the automated bridge crane, about 4.5 meters away from the top surface of the ton bags.

[0114] During the perception phase, due to uneven lighting in the industrial site, some ton bags were overexposed due to direct sunlight, while the gaps between stacked bags were in deep shadow. The high-frequency texture noise suppression module successfully extracted the overexposed edges through guided filtering and enhanced the dark features using dynamic exposure compensation, eliminating interference fringes generated by the woven texture. The multi-source feature deep fusion module projected the depth point cloud onto the enhanced feature map, constructing a local spatial model containing 500,000 voxels.

[0115] During the clustering and segmentation stage, in response to the collapse and mutual compression of the bags caused by material filling, the multi-clustering spatial fusion engine identified that the normal vectors between adjacent bags have obvious gradient jumps at the contact interface, with an angle of 42 degrees. This accurately sets up a virtual cutting surface at the adhesion point, completely separating the three groups of close-to-each-bags, with an identification accuracy of 99.8%.

[0116] During the attitude correction phase, the system detected that one of the target ton bags was stacked unevenly, and its geometric center was 15 centimeters away from its physical center of gravity. The dynamic physical attitude correction module identified the higher material density on the left side of the bag by the laser echo intensity distribution and calculated the true physical center of gravity coordinates.

[0117] The system will then automatically correct the center point to the left and output the corrected operation vector.

[0118] Comparative Example 1: Traditional 2D Vision-Based Ton Bag Inspection System

[0119] Using the same operating environment, traditional systems rely solely on a single industrial camera and a target detection algorithm based on convolutional neural networks;

[0120] In cargo handling statistics, due to the interference of wrinkles and textures on the surface of ton bags, traditional systems often experience detection frame breakage when facing dark bags or strong shadows, resulting in the error of identifying a ton bag as multiple fragments, with an accuracy rate of only 86%.

[0121] More importantly, during the grabbing process, the traditional system can only output the geometric center of the ton bag outline. Since it does not take into account the center of gravity shift caused by material accumulation, significant bag shaking occurred in all three lifting tests. In one of them, due to excessive center of gravity deviation, the lifting device slipped during the ascent, posing a serious operational safety hazard.

[0122] Experimental data comparison table:

[0123]

[0124] The data comparison between the above embodiments and comparative examples clearly shows that the ton bag cargo identification method for port tallying involved in this invention has overwhelming technical advantages in core dimensions such as identification accuracy, operational safety, and processing efficiency through deep complementarity of multimodal data, physical modeling of three-dimensional geometric posture, and low-level suppression of high-frequency interference.

[0125] In the implementation of this project, the system also features adaptive reflectivity adjustment specifically designed for the diverse materials of ton bags, such as antistatic coated bags and aluminum foil composite bags.

[0126] When encountering a highly reflective metal film ton bag, the system automatically reduces the laser pulse width to prevent the photosensitive element from saturating, thereby ensuring that the point cloud distribution remains uniform even on special material surfaces.

[0127] In terms of software upgrades, the system has reserved an online model update interface based on industrial Ethernet, which can continuously optimize the deformation tensor correction model in the clustering engine based on new deformation samples collected on site, giving it self-optimization capabilities throughout its entire lifecycle.

[0128] Furthermore, the hardware connection structure of the sensing platform of this invention adopts a quick-locking mechanical device, which allows maintenance personnel to complete the replacement and hot-swap restart of the entire sensing module within 5 minutes. The calibration parameters inside the system are protected by non-volatile memory. After replacing the hardware, zero-point calibration can be completed by a single automated target scan, which lowers the technical threshold for operation and maintenance in industrial sites.

[0129] At the communication layer of the central control system, this invention supports standard industrial IoT protocols, enabling direct connection of inventory data to an enterprise's ERP or WMS system. The inventory reports not only include real-time inventory quantities, but also the volume distribution characteristics of each bag of goods and snapshots of their receiving posture. These digital assets provide valuable underlying data for subsequent goods stacking optimization, logistics quality traceability, and operational process analysis.

[0130] In a further expansion scheme of the system, the sensing platform can also adopt an array-style layout according to the actual operating radius. That is, multiple multimodal sensing nodes described in this invention are arranged at intervals on the crossbeam of a large crane. Each node achieves nanosecond-level alignment through a synchronous clock line, thereby constructing a collaborative sensing network covering an ultra-large field of view. The data collected by each node is processed through a distributed computing framework to achieve seamless stitching and global consistency optimization of point clouds in the background. This clustered deployment method can complete the digital modeling and sorting of the entire ship's hold or the entire warehouse stack in one go, and its operational efficiency will increase exponentially.

[0131] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying ton bag cargo in port tallying, comprising a detection system, characterized in that, The detection system is built on the end-sensing platform of the automated loading and unloading mechanism. The detection system includes: a multimodal data acquisition array, which synchronously acquires the original multidimensional information of the target ton bag to be detected. The original multidimensional information includes at least a two-dimensional image stream, a three-dimensional spatial point cloud stream, and the instantaneous motion attitude vector of the sensing platform. The spatiotemporal reference synchronization module is connected to the multimodal data acquisition array. It is used to assign a unified timestamp to the original multidimensional information through a hardware-level triggering mechanism, and to align the multimodal data in spatial coordinate system based on the instantaneous motion attitude vector. The high-frequency texture noise suppression module adopts a dual architecture combining spatial domain guided filtering and frequency domain notch filtering. It suppresses texture pseudo-noise on non-rigid surfaces by identifying and filtering out high-frequency energy peaks that represent the weaving density distribution of ton bags. The multi-clustering spatial fusion engine maps noise-suppressed image features and point cloud data into a unified voxel space, performs multi-dimensional clustering based on semantics, deep connectivity and normal vector consistency, and uses normal vector mutation points as physical segmentation benchmarks to sever the visual adhesion between adjacent ton bags. The dynamic physical attitude correction module extracts the point cloud clusters of individual ton bags after segmentation, estimates the mass density distribution of the material inside the bag by the point cloud volume integral and echo intensity distribution, calculates the physical center of gravity of the ton bag in three-dimensional space, and projects the physical center of gravity onto the surface of the bag along the gravity vector direction to determine the optimized gripping center point and operation normal vector. The central control and task scheduling system integrates the detection reports output by each module, generates a comprehensive instruction including the target UID, three-dimensional spatial pose, physical diameter and grasping confidence, and issues it to the automated loading and unloading execution mechanism.

2. The method for identifying ton bag cargo in port tallying according to claim 1, characterized in that, The multimodal data acquisition array includes: Industrial cameras are used to capture raw images of ton bags and acquire high-resolution color texture information and boundary contrast features on the surface of the ton bags. A set of multi-line pulse lidar, positioned on the geometric center axis of the camera array, is used to construct a high-density 3D point cloud of ton bag stacks; A set of embedded inertial measurement modules is used to sense the instantaneous attitude angle and high-frequency vibration vector of the sensing platform in real time.

3. The method for identifying ton bag cargo in port tallying according to claim 2, characterized in that, The spatiotemporal reference synchronization module drives a hardware trigger distributor through a field-programmable gate array to generate synchronization trigger pulses, so that the exposure start time of the industrial camera and the scanning start time of the multi-line pulse lidar are aligned on the same time scale. The spatiotemporal reference synchronization module is also used to inject the platform pose data output by the embedded inertial measurement module into the data frame header of the image and point cloud in real time, so as to eliminate geometric mapping motion distortion caused by the movement of the actuator.

4. The method for identifying ton bag cargo in port tallying according to claim 2, characterized in that, The high-frequency texture noise suppression module specifically includes the following when performing noise suppression: Spatial processing link: The original images captured by the industrial camera are smoothed by using an adaptive guided filter based on a local linear model. By solving the linear coefficients within a local window, the gradient change characteristics at the physical boundary of the ton bag are preserved while suppressing small-scale fluctuations in the woven texture. Frequency domain processing link: The original image is transformed to the frequency domain through fast Fourier transform, the two-dimensional power spectral density is analyzed to identify the characteristic frequencies representing the repeating pattern of the woven mesh, and a notch filter is set to remove local contrast abrupt changes caused by wrinkles. Dynamic exposure compensation logic: Real-time monitoring of the brightness distribution of the shadow area of ​​the ton bag wrinkles in the original image, and enhanced extraction of regional features by adjusting the exposure time and gain parameters of the industrial camera.

5. The method for identifying ton bag cargo in port tallying according to claim 1, characterized in that, The multi-cluster spatial fusion engine executes a three-dimensional collaborative clustering logic, specifically including: First dimension: Semantic clustering based on Lab color space is used to divide the surface color consistency region of the ton bag target; The second dimension: spatial connectivity clustering based on deep gradients, used to identify physical entities with continuous geometric properties in point clouds; The third dimension: Curvature clustering based on surface normal vector consistency, using the eigenvalues ​​of the covariance matrix in the local neighborhood to evaluate surface smoothness; The multi-clustering space fusion engine uses a weighted voting mechanism to fuse the clustering results of the first, second, and third dimensions in the voxel space; When two targets are detected to be visually stuck together, the rate of change of the angle between the normal vectors at the sticking interface is identified. If the angle between the normal vectors exceeds a preset threshold, a spatial physical segmentation surface is set at the sticking interface.

6. The method for identifying ton bag cargo in port tallying according to claim 5, characterized in that, The multi-cluster spatial fusion engine also integrates a matching operator for a non-rigid model, which pre-stores a set of flexible deformation topology templates for standard ton bags under different load states. During the clustering process, the matching operator matches the detected edge features with the topological template and corrects the missing geometric boundaries caused by local collapse or occlusion by calculating the deformation tensor.

7. A method for identifying ton bag cargo in port tallying according to claim 2, characterized in that, The dynamic physical attitude correction module executes mass distribution estimation logic based on volume integral, and the specific steps include: The spatial distribution of the laser radar echo intensity is used to identify the density difference of the material inside the bag, and the echo intensity region corresponds to the mass density volume module. The segmented ton bag point cloud cluster is divided into multiple small volume modules, and different mass weight coefficients are assigned to each module according to the density difference. The physical center of gravity coordinates of the target are calculated by weighted summation of the centroids of the entire bag volume modules. Based on the instantaneous platform attitude feedback from the embedded inertial measurement module, vector correction is performed on the center of gravity shift caused by bag deformation, and the optimized gripping center point is output.

8. A method for identifying ton bag cargo in port tallying according to claim 7, characterized in that, The dynamic physical attitude correction module also includes a safety verification logic, which is used to analyze the point cloud flatness and surface normal vector dispersion around the pre-selected point before outputting the final work position. If the slope of the pre-selected grasping area is detected to exceed the preset tilt angle threshold, the dynamic physical attitude correction module searches for a flat area with consistent normal vectors in the neighborhood along the gravity projection line as the suboptimal working position and issues a safety warning to the central control and task scheduling system.

9. A method for identifying ton bag cargo in port tallying according to claim 1, characterized in that, The central control and task scheduling system uses a multi-core processor architecture to execute tasks in parallel. The sensing data synchronization task is given real-time priority, while the target detection and clustering fusion tasks are executed in parallel through a pipelined manner of the computing cores.

10. A method for identifying ton bag cargo in port tallying according to claim 2, characterized in that, The detection system also includes meteorological compensation logic and a clustered collaborative interface, wherein: the meteorological compensation logic dynamically adjusts the reflection intensity gain of the multi-line pulse lidar based on the visibility parameters fed back by the environmental sensors, and calls a long-wave infrared enhancement algorithm to perform fog-penetrating processing on the images acquired by the industrial camera; The clustered collaborative interface supports real-time synchronization of cargo handling data, center of gravity correction vector diagrams, and capture execution parameters to the cloud management platform via 5G network, so as to realize pose data sharing and collaborative operation among multiple loading and unloading equipment.

Images