Oversized target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion

Abstract

The invention discloses an oversized target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion, particularly relates to the technical field of industrial automation and three-dimensional machine vision crossing, and comprises a multi-view 3D camera array module, a point cloud fusion processing module, a three-dimensional reconstruction module, an AI positioning and sorting decision module and a robot execution module. View angle shielding is eliminated through multi-sensor surrounding synchronous collection, and full coverage of an ultra-large target is achieved. And according to point cloud density and confidence coefficient dynamic fusion, the data integrity of the reflective weak texture region is improved. Segmented parts are rebuilt in parallel and then spliced, and the calculation complexity of an oversized model is reduced. Coarse registration and multi-modal attention fine correction are combined, and both speed and sub-centimeter-level positioning precision are considered. A time sequence model is used to compensate sensor drift on line, and the stability in a vibration temperature change environment is enhanced. And on the basis of a dense map real-time collision detection planning track, flexible grabbing is achieved by combining with feedback, and precise sorting of the special-shaped super-large workpieces is completed.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of industrial automation and 3D machine vision, and particularly to a system for the reconstruction of ultra-large targets and robot AI sorting based on multi-view 3D point cloud fusion. Background Technology

[0002] The field of 3D machine vision and industrial automation technology includes technologies that use optical sensors to acquire 3D information of objects and guide robots to complete tasks such as grasping and assembly. The core content of this field involves recovering the 3D geometric structure and spatial pose of a target object from 2D images or depth data, and converting this information into motion control commands that can be executed by the robot. Its systematic process covers multiple stages such as data acquisition, 3D reconstruction, target recognition and localization, and motion planning, aiming to achieve unmanned and precise operation of various workpieces in the production and logistics process.

[0003] Among them, the ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion refers to a technical solution for automated operation of irregular physical objects with a size exceeding one meter. The technical issues addressed by this topic cover the synchronization and integrity of multi-view 3D data acquisition, rapid registration and fusion of massive point cloud data, efficient construction and detail restoration of ultra-large 3D models, robust estimation of the accurate pose of targets in complex scenes, and generation of safe and collision-free operation trajectories for robots. The specific solution involves arranging a ring sensor array and using hardware triggering to achieve microsecond-level synchronous acquisition. The weights are dynamically calculated based on the local density of the point cloud and the confidence of the sensor ranging to achieve multi-source data fusion. The target is divided into sub-components for parallel mesh reconstruction and then stitched together through feature matching. The target position and pose are calculated by combining a two-stage process of geometric coarse registration and visual feature fine correction. The robot arm's motion path is calculated in a known environment map using random sampling search and real-time collision detection.

[0004] However, traditional monocular and binocular systems have limited fixed viewing angles, resulting in blind spots when observing very large targets. This leads to incomplete point cloud data and voids in the reconstructed model. Existing point cloud fusion methods mostly use static parameters, which are difficult to adapt to data quality fluctuations caused by changes in lighting and surface reflections, easily leading to fusion distortion or residual noise. For very large targets, the computational load of the overall one-time reconstruction method increases cubically with size, and the processing delay cannot meet the production line cycle time. Most localization methods rely on a single geometric or image modality, failing to effectively complement each other, thus limiting the accuracy of pose estimation. Systems often ignore the slow drift of sensors during long-term operation and lack online compensation, resulting in accuracy degradation over time. Robot path planning is often based on static environmental models, unable to respond to dynamic scenes, and grasping strategies lack consideration for object deformation and force feedback. Summary of the Invention

[0005] The main objective of this invention is to provide a system for the reconstruction of ultra-large targets and the sorting of robots based on multi-view 3D point cloud fusion, which can effectively solve the problems mentioned in the background technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The ultra-large target reconstruction and robotic AI sorting system based on multi-view 3D point cloud fusion is configured as follows: Multi-view 3D camera array module: arranged in a ring around the target, used to simultaneously acquire multi-view depth images and generate local point clouds; Point cloud fusion processing module: used to perform spatiotemporal alignment, weighted fusion and noise filtering on the local point cloud to generate a global high-density point cloud; 3D Reconstruction Module: Used to perform layered reconstruction and detail enhancement of ultra-large targets based on the high-density point cloud, and output a complete 3D model; AI positioning and sorting decision module: used for target detection, pose estimation and semantic guidance of grasping point prediction of the 3D model; Robot execution module: used for path planning and adaptive grasping control based on the grasping point and pose information.

[0007] Preferably, the point cloud fusion processing module includes: The spatiotemporal synchronization unit uses hardware trigger signals and the PTP protocol to align the timestamps of multi-sensor data; the adaptive weighted fusion unit dynamically adjusts the fusion weights based on the local density and confidence level of each sensor point cloud; and the noise filtering unit removes dynamic interference point clouds in real time based on the spatial trajectory fluctuation index and texture change rate.

[0008] Preferably, the three-dimensional reconstruction module employs a layered reconstruction strategy, including: The bottom-level reconstruction unit uses Poisson reconstruction or voxelization to generate the basic mesh model; the detail enhancement unit uses NeRF (Neural Radiation Field) for texture refinement and hole repair; the local-global optimization unit divides the target into several sub-modules for parallel reconstruction and then stitches them together through feature matching.

[0009] Preferably, the AI ​​positioning and sorting decision module includes: The multi-task 3D detection network, based on a sparse convolutional architecture, outputs the target's 3D bounding box and grasp point heatmap; the semantic guidance unit supports text or voice command input and dynamically adjusts sorting priority and strategy; the real-time pose correction unit uses a cross-modal attention mechanism to perform sub-centimeter-level correction of the initial pose.

[0010] Preferably, the robot execution module includes: Collision-aware path planning unit, based on Euclidean symbolic distance field (ESDF) and RRT The algorithm generates collision-free trajectories; the compliant control unit combines force feedback and visual servoing to achieve adaptive grasping; and the dynamic replanning unit updates the motion path in real time when the environment changes or the target moves.

[0011] Preferably, the point cloud fusion step includes: The Super4PCS algorithm is used for coarse registration; the ICP algorithm and normal vector constraints are combined for fine registration; and the fusion weights of RGB-D point cloud and laser point cloud are dynamically adjusted based on the local density of the point cloud.

[0012] Preferably, the three-dimensional reconstruction step includes: The target is segmented at the component level to generate multiple sub-modules; Poisson reconstruction is performed in parallel on each sub-module; and the sub-modules are stitched together into a complete model through feature matching and surface optimization.

[0013] Preferably, the AI ​​positioning steps include: adopting a hierarchical positioning architecture: first, coarse positioning is performed through point cloud registration, and then fine positioning is performed through a cross-modal attention mechanism; combining LSTM network to predict sensor drift and adapt to vibration and temperature change environments.

[0014] Preferably, the sorting system further includes: a substrate; multiple cones arranged in an array on the substrate, the cones being arranged in a regular grid layout; the horizontal spacing between adjacent cones being 200mm±1mm; the height-to-base-diameter ratio of the cones being a preset optimized value, enabling the 3D camera to stably identify their vertex features.

[0015] Compared with the prior art, the present invention has the following beneficial effects: By employing multi-sensor surround synchronous acquisition, single-view occlusion is eliminated, achieving full-coverage data acquisition for ultra-large targets. Fusion weights are dynamically allocated based on local point cloud density and sensor confidence levels to improve data integrity in reflective or weakly textured areas. A strategy of first segmenting components for parallel reconstruction, followed by feature matching and stitching, significantly reduces the overall computational complexity of ultra-large models. A two-stage localization process consisting of coarse registration and multimodal attention-based fine correction is designed, balancing processing speed and sub-centimeter-level positioning accuracy. A temporal model is introduced for online estimation and compensation of sensor system drift, enhancing long-term stability under vibration and temperature variations. Real-time collision detection combined with a dense environmental map is used to plan the robotic arm trajectory, and force feedback is utilized to achieve adaptive compliant grasping, enabling precise sorting of irregularly shaped ultra-large workpieces. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0017] 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.

[0018] Example 1, as Figure 1 As shown, the ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion is configured as follows: Multi-view 3D camera array module: arranged in a ring around the target, used to simultaneously acquire multi-view depth images and generate local point clouds; Point cloud fusion processing module: used to perform spatiotemporal alignment, weighted fusion and noise filtering on local point clouds to generate a global high-density point cloud; 3D Reconstruction Module: Used for layered reconstruction and detail enhancement of ultra-large targets based on high-density point clouds, outputting a complete 3D model; AI localization and sorting decision module: used for target detection, pose estimation and semantic guidance of grasping point prediction in 3D models; Robot execution module: used for path planning and adaptive grasping control based on grasping point and pose information.

[0019] Specifically, this is implemented in a car wheel hub sorting scenario. Six binocular structured light 3D cameras are equidistantly deployed in a 3-meter radius ring above the sorting area. Each camera simultaneously starts acquiring data at time t0 via the same hardware trigger. For a wheel hub target 2.5 meters from the camera base, each camera acquires a local point cloud P_i (i=1 to 6) containing approximately 500,000 points. Each point cloud data is accompanied by a timestamp t0. The six local point clouds P_i are transmitted to the fusion processing module. This module calls the pre-stored transformation matrix T_i of each camera relative to the origin of the global coordinate system, obtained through a calibration device, and performs a coordinate transformation operation P_gi=T_i on each point cloud. P_i imports all transformed point clouds {P_gi} into a common data buffer. A statistical outlier removal algorithm is used to calculate the average distance between each point in the buffer and its 50 nearest neighbors. If the average distance of a point exceeds 2.0 times the standard deviation of the global point cloud's average distance, that point is removed from the buffer. The denoised point cloud is then downsampled using a voxel grid, with a cube grid side length of 5 mm, retaining the centroids of all points within each grid cell. This generates a global high-density point cloud P_global containing approximately 2 million points. P_global is then input into the 3D reconstruction module, which first calculates... The normal vectors of the points in P_global are used to segment the point cloud into multiple connected regions based on a region growing algorithm. For each connected region, the Poisson surface reconstruction function is called, setting the reconstruction depth to 9 to generate a basic triangular mesh model M_base. Next, texture information for the corresponding regions is extracted from the RGB images synchronously acquired from each original camera. By establishing a mapping relationship between 3D points and 2D pixels, color information is attached to the corresponding triangular faces of M_base. For texture-deficient regions caused by occlusion, the interpolation of the average color of its adjacent triangular faces is used for filling, outputting a complete textured 3D wheel hub model M. The AI ​​localization and sorting decision module loads M_final and calls a pre-trained multi-task 3D detection network model. This model takes a voxelized point cloud as input, with a voxel size of 0.1 meters. The network outputs a 3D bounding box parameter (center coordinates [x=1.2, y=0.5, z=2.55], dimensions [length=0.6, width=0.6, height=0.2], yaw angle θ=30°) and a gripping point heatmap. In the heatmap, the point with coordinates (1.18, 0.52, 2.56) has the highest heat value of 0.92 and is selected as the gripping point G. The robot then executes the module... The block receives the coordinates of the gripping point G and the bounding box attitude angle θ. Combined with the known hub weight (20 kg) and gripper friction coefficient (0.4), the required gripping force is calculated to be 80 N. The planning module takes the robotic arm base as the origin and point G as the target point. On the known obstacle map (including a shelf with a height of 1 meter), through random sampling and collision detection, a collision-free joint space trajectory consisting of 20 path points is generated. The trajectory and force control parameters are sent to the robot controller, which drives the end effector of the six-axis robotic arm to move at a speed of 0.5 m / s and perform gripping. This control system completes the entire process from vision acquisition to gripping execution.

[0020] The point cloud fusion processing module includes: The spatiotemporal synchronization unit uses hardware trigger signals and the PTP protocol to align the timestamps of multi-sensor data; the adaptive weighted fusion unit dynamically adjusts the fusion weights based on the local density and confidence level of each sensor point cloud; and the noise filtering unit removes dynamic interference point clouds in real time based on the spatial trajectory fluctuation index and texture change rate.

[0021] Specifically, when handling large cargo boxes in a logistics warehouse, the spatiotemporal synchronization unit uses a switch supporting the PTP protocol. The master clock sends synchronization messages to ensure that the timestamp deviation of each sensor is less than 1 millisecond. When the adaptive weighted fusion unit processes the point cloud P_k of sensor k, it calculates the average number of points ρ_k within a sphere with a radius of 0.1 meters, queries the sensor confidence level c_k under the current ranging, and calculates the weight α_k=(ρ_k / 100). For example, if the densities of two sensors are 80 and 120, and the confidence levels are 0.9 and 0.98, then α_1 = 0.72 and α_2 = 1.176, which are normalized to 0.38 and 0.62. The coordinates of the fused point in this area are obtained by multiplying the coordinates of the two sensor points by their normalized weights and then adding them together. The noise filtering unit analyzes the X-coordinate sequence of a specific point (1,0,2) in 5 consecutive frames on the point cloud of the cargo box on the conveyor belt and calculates the standard deviation σ_x. If σ_x > 0.02 meters threshold, it is determined to be a dynamic point and deleted. At the same time, the gray-scale change rate of the point area is calculated, and it is also deleted if it exceeds 30%.

[0022] The 3D reconstruction module employs a layered reconstruction strategy, including: The bottom-level reconstruction unit uses Poisson reconstruction or voxelization to generate the basic mesh model; the detail enhancement unit uses NeRF (Neural Radiation Field) for texture refinement and hole repair; the local-global optimization unit divides the target into several sub-modules for parallel reconstruction and then stitches them together through feature matching.

[0023] When reconstructing the robotic arm base, the bottom reconstruction unit sets the Poisson reconstruction depth to 10 for the fused point cloud, solves the Poisson equation to generate the main mesh M_base, and the detail enhancement unit loads a pre-trained NeRF model for the hole region on M_base. Taking the coordinates (x,y,z) and view direction as input, the neural network predicts the color and density of the point. For the point (1.1,2.2,0.5) inside the hole, the color (128,128,128) and density are obtained from multi-view integration. Points with a density higher than 0.5 are added to the surface point cloud. The local-global optimization unit first divides the base point cloud into 4 component sub-modules based on the normal vector and curvature using the region growing algorithm. Poisson reconstruction (depth 8) is called independently for each sub-module to generate a sub-mesh. Point sets are extracted at the boundaries of adjacent sub-modules, and the ICP algorithm is used for registration until the average distance is less than 0.001 meters. The transformation matrix is ​​obtained and the sub-mesh is aligned. Finally, Laplacian smoothing is performed on the splicing area.

[0024] The AI-based localization and sorting decision-making module includes: The multi-task 3D detection network, based on a sparse convolutional architecture, outputs the target's 3D bounding box and grasp point heatmap; the semantic guidance unit supports text or voice command input and dynamically adjusts sorting priority and strategy; the real-time pose correction unit uses a cross-modal attention mechanism to perform sub-centimeter-level correction of the initial pose.

[0025] Furthermore, during the sorting of aircraft blades, the multi-task 3D detection network takes the voxelized representation of the foreground point cloud (0.05 m voxels) as input. After sparse convolution and upsampling, the network outputs the bounding box parameters of the blades and the heatmap of the grasping points. For example, it outputs the center (0.5, 0.2, 1.0) and marks the root point (0.48, 0.21, 1.02) with a heatmap of 0.88. The semantic guidance unit receives the text instruction "prioritize grasping blades with wear marks", encodes the instruction into a feature vector, and simultaneously extracts the HSV color statistics of each cluster of the point cloud. The system calculates the cosine similarity between the instruction vector and each color feature. For clusters with high similarity (associated with dark and low saturation), the corresponding capture point heat value is multiplied by the excitation coefficient 1.5. The real-time pose correction unit extracts 10 SIFT feature points on the blade from the RGB image and establishes a 3D-2D correspondence set with the corresponding points in the 3D model. The PnP algorithm is used to solve the pose P_pnp, and the translation difference between it and the initial pose P_initial is calculated. If it is greater than 0.005 meters, P_initial is linearly interpolated with a weight of 0.7.

[0026] The robot execution module includes: Collision-aware path planning unit, based on Euclidean symbolic distance field (ESDF) and RRT The algorithm generates collision-free trajectories; the compliant control unit combines force feedback and visual servoing to achieve adaptive grasping; and the dynamic replanning unit updates the motion path in real time when the environment changes or the target moves.

[0027] When gripping the gearbox housing between obstacles, the collision-aware path planning unit converts the environment into an ESDF (Electronic Stability Path Definition). During planning, it randomly samples the state q_rand in the joint space, finds the nearest node q_near in the tree, and extends q_rand by 0.1 meters to obtain q_new. When calculating q_new, the robotic arm link envelope is calculated, and the distance values ​​in the ESDF are queried. If all distances are greater than the 0.1-meter safety threshold, q_new is accepted. This process is repeated until the target state q_goal is reached. The compliance control unit reads the readings of the six-dimensional force sensor after the gripper contacts the surface. When the axial force Fz reaches 10 N, it switches to force control mode and inputs the difference between the actual force and the target force into the PI controller (Kp=0.5, Ki=0.1) to calculate the joint torque adjustment. After detecting a new obstacle, the dynamic replanning unit updates the ESDF map and immediately reruns the planning algorithm to generate an obstacle avoidance trajectory starting from the current state.

[0028] Point cloud fusion steps include: The Super4PCS algorithm is used for coarse registration; the ICP algorithm and normal vector constraints are combined for fine registration; and the fusion weights of RGB-D point cloud and laser point cloud are dynamically adjusted based on the local density of the point cloud.

[0029] When reconstructing the car bumper, Super4PCS is used for coarse registration of the two-view point clouds P_A and P_B. Four points are selected from P_A to form a group, and six internal distances are calculated. In P_B, a search is conducted for four points whose distances to the group distances are all less than a tolerance of 0.06 meters. Based on this, the initial transformation T_coarse is calculated. Fine registration uses ICP combined with normal vector constraints. For the T_coarse-transformed P_A' and P_B, the nearest neighbor pair in P_B is found for the midpoint of P_A'. The cosine of the angle between the normal vectors of point pairs is calculated, and point pairs with a cosine value less than 0.95 are removed. The remaining point pairs are then used to solve for the optimal transformation T_icp through SVD. The iteration continues until the average point pair distance change is less than 0.001 meters. The fusion weights are dynamically adjusted according to the local density. For example, if the RGB-D density ρ_rgbd = 150 and the laser density ρ_lidar = 400 in a certain area, then the weight ratio α_rgbd:α_lidar = 0.375:1, which is approximately 0.27:0.73 after normalization.

[0030] The three-dimensional reconstruction steps include: The target is segmented at the component level to generate multiple sub-modules; Poisson reconstruction is performed in parallel on each sub-module; and the sub-modules are stitched together into a complete model through feature matching and surface optimization.

[0031] When reconstructing the engine block, component-level segmentation calculates the normal vector and curvature of each point in the point cloud. Using unvisited points with low curvature as seeds, region growth is performed, incorporating neighboring points with a normal vector angle of less than 15 degrees, until no further growth is possible. This process segments six sub-module point clouds, such as cylinder bores and water passages. The Poisson reconstruction function (depth 8) is called in parallel on each sub-module point cloud, outputting six independent mesh models. During feature matching and stitching, at least three pairs of corresponding feature points are selected at the boundaries of adjacent sub-modules, and the alignment transformation matrix is ​​solved using the least squares method. Subsequently, the boundary point set is fine-tuned and registered using ICP until the average distance is less than 0.0005 meters. The mesh data is then merged, duplicate vertices are deleted, and the triangular faces at the connection points are re-subdivided.

[0032] The AI ​​positioning steps include: adopting a hierarchical positioning architecture: first, coarse positioning is performed through point cloud registration, and then fine positioning is performed through a cross-modal attention mechanism; combining LSTM network to predict sensor drift and adapt to vibration and temperature change environments.

[0033] When locating a large gear, coarse localization involves extracting the FPFH features of the gear CAD model point cloud P_model and the scene point cloud P_scene, performing RANSAC matching, and estimating the initial transformation T_coarse to achieve a 60% overlap. Fine localization projects the coarse localization point cloud onto an RGB image to obtain a prediction mask, calculates the IoU with the mask obtained from the instance segmentation network, and if the IoU < 0.85, concatenates the global features of the point cloud with the RGB region features, inputs them into the MLP, and outputs 6-DOF pose corrections (Δx, Δy, Δz, ...). Δroll, Δpitch, Δyaw), such as (0.003, -0.002, 0.001), are applied to the coarse localization results. The sensor drift prediction records the changes in camera extrinsic parameters every 5 minutes to form a sequence, which is input into a 2-layer 64-unit LSTM network to predict the drift amount in the next 5 minutes (δx_pred, δy_pred, δz_pred, δθ_pred), such as (0.0002, 0.0001, -0.0003, 0.01°). The original visual pose is subtracted from this predicted value in advance.

[0034] The sorting system also includes: Substrate; multiple cones are arrayed on the substrate in a regular grid layout; the horizontal spacing between adjacent cones is 200mm±1mm; the ratio of the height to the base diameter of the cones is a preset optimized value, which enables the 3D camera to stably recognize their vertex features.

[0035] During deployment in the assembly workshop, a 2m x 2m matte white aluminum substrate was used, on which 49 ABS cones (7 x 7) were arranged in a regular pattern. Each cone had a base diameter of 40mm and a height of 60mm (height-to-diameter ratio 1.5), with a horizontal spacing of 200mm ± 1mm ​​(actual measurements such as 199.8mm and 200.1mm). The device was placed at the center of the common field of view of the camera array, with 8 cameras simultaneously capturing images. The pixel coordinates of each cone's vertices were identified, and the 3D coordinate set {Q_j} in each cone's coordinate system was calculated based on depth. This set was compared with the world coordinates {W} on the design drawings. The optimal transformation matrix T_j for each camera was then obtained using least squares, ensuring that T_j... The average distance error between {Q_j} and {W} is the smallest. After calibration, the translation error between each camera is less than 2 mm and the rotation error is less than 0.5 degrees.

[0036] 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 this invention is defined by the appended claims and their equivalents.

Claims

1. A system for ultra-large target reconstruction and robotic AI sorting based on multi-view 3D point cloud fusion, characterized in that: Configured as: Multi-view 3D camera array module: arranged in a ring around the target, used to simultaneously acquire multi-view depth images and generate local point clouds; Point cloud fusion processing module: used to perform spatiotemporal alignment, weighted fusion and noise filtering on the local point cloud to generate a global high-density point cloud; 3D Reconstruction Module: Used to perform layered reconstruction and detail enhancement of ultra-large targets based on the high-density point cloud, and output a complete 3D model; AI positioning and sorting decision module: used for target detection, pose estimation and semantic guidance of grasping point prediction of the 3D model; Robot execution module: used for path planning and adaptive grasping control based on the grasping point and pose information.

2. The ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion according to claim 1, characterized in that: The point cloud fusion processing module includes: The spatiotemporal synchronization unit uses hardware trigger signals and the PTP protocol to align the timestamps of multi-sensor data; the adaptive weighted fusion unit dynamically adjusts the fusion weights based on the local density and confidence level of each sensor point cloud; and the noise filtering unit removes dynamic interference point clouds in real time based on the spatial trajectory fluctuation index and texture change rate.

3. The ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion according to claim 1, characterized in that, The 3D reconstruction module employs a layered reconstruction strategy, including: The bottom-level reconstruction unit uses Poisson reconstruction or voxelization to generate the basic mesh model; the detail enhancement unit uses NeRF (Neural Radiation Field) for texture refinement and hole repair; the local-global optimization unit divides the target into several sub-modules for parallel reconstruction and then stitches them together through feature matching.

4. The ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion according to claim 1, characterized in that, The AI-based localization and sorting decision-making module includes: The multi-task 3D detection network, based on a sparse convolutional architecture, outputs the target's 3D bounding box and grasp point heatmap; the semantic guidance unit supports text or voice command input and dynamically adjusts sorting priority and strategy; the real-time pose correction unit uses a cross-modal attention mechanism to perform sub-centimeter-level correction of the initial pose.

5. The ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion according to claim 1, characterized in that, The robot execution module includes: Collision-aware path planning unit, based on Euclidean symbolic distance field (ESDF) and RRT The algorithm generates collision-free trajectories; the compliant control unit combines force feedback and visual servoing to achieve adaptive grasping; and the dynamic replanning unit updates the motion path in real time when the environment changes or the target moves.

6. The ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion according to claim 1, characterized in that, The point cloud fusion step includes: The Super4PCS algorithm is used for coarse registration; the ICP algorithm and normal vector constraints are combined for fine registration; and the fusion weights of RGB-D point cloud and laser point cloud are dynamically adjusted based on the local density of the point cloud.

7. The ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion according to claim 1, characterized in that, The three-dimensional reconstruction steps include: The target is segmented at the component level to generate multiple sub-modules; Poisson reconstruction is performed in parallel on each sub-module; and the sub-modules are stitched together into a complete model through feature matching and surface optimization.

8. The ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion according to claim 1, characterized in that, The AI ​​positioning steps include: adopting a hierarchical positioning architecture: first, coarse positioning is performed through point cloud registration, and then fine positioning is performed through a cross-modal attention mechanism; combining LSTM network to predict sensor drift and adapt to vibration and temperature change environments.

9. The ultra-large target reconstruction and robot AI sorting system based on multi-view 3D point cloud fusion according to claim 1, characterized in that, The sorting system also includes: A substrate; multiple cones arranged in an array on the substrate, the cones being arranged in a regular grid layout; the horizontal spacing between adjacent cones is 200mm±1mm; the height-to-base-diameter ratio of the cones is a preset optimized value, enabling the 3D camera to stably recognize their vertex features.

Images