Industrial robot precise path generation system based on three-dimensional visual reconstruction

Abstract

The invention belongs to the technical field of industrial robot motion control, and discloses an industrial robot precise path generation system based on three-dimensional visual reconstruction, comprising an environment sensing device which is composed of a multi-baseline structured light projector and a multi-view camera array and obtains point cloud data through light interference suppression and spectrum correction; the three-dimensional reconstruction processing unit generates a three-dimensional model through rigid transformation and deep learning correction; the geometric feature analysis unit performs curvature analysis and boundary recognition to form region description; the trajectory planning device adopts a weighted B spline to generate a trajectory under curvature constraint and energy optimization; the dynamic correction unit is switched between vision feedback and force feedback according to an error threshold value; the control conversion processor combines inverse kinematics and dynamics compensation and predicts deviation to output a control instruction; and the execution driving mechanism utilizes a servo motor and closed loop feedback to execute actions, so that high-precision generation and stable execution of a track under a complex working condition are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial robot motion control, and particularly relates to an industrial robot precise path generation system based on three-dimensional vision reconstruction. BACKGROUND

[0002] Industrial robots have become important equipment for realizing high-efficiency and high-precision operation in modern manufacturing industry, and are widely applied to welding, assembly, spraying and precision machining occasions. In order to enable an end effector of the robot to closely adhere to a workpiece surface for operation, an environment information acquisition system is usually relied on to acquire environment information, and a trajectory conforming to geometric features of the workpiece is generated on the basis of the environment information. However, there are still great deficiencies in point cloud acquisition, three-dimensional reconstruction and trajectory planning in the prior art.

[0003] At present, a common method adopts a monocular or binocular camera to acquire point cloud data, and combines an iterative closest point (ICP) algorithm or a projection matching method to perform three-dimensional reconstruction. However, in the case of complex illumination, reflection or partial occlusion, the point cloud data often has noise or is missing, which leads to insufficient accuracy of the reconstructed model. The trajectory generated based on such a model is difficult to accurately reflect the surface features of the workpiece, which affects the quality of subsequent operation. Meanwhile, traditional trajectory planning is mainly based on spline fitting, and only pays attention to geometric continuity, lacks curvature constraint and energy optimization, and is prone to speed fluctuation in the turning area, thereby reducing the smoothness and execution stability of the trajectory.

[0004] In the trajectory execution and correction link, the prior art usually relies on single visual feedback to perform error adjustment. When the working condition deviates greatly or mechanical interference occurs, there is a lack of multi-modal feedback switching mechanism, which leads to correction lag or even failure. In addition, the control layer mostly only performs conversion between geometric coordinates and joint angles, and does not introduce a dynamics compensation mechanism. Therefore, when running at high speed or bearing a large load, the execution accuracy is significantly reduced, and trajectory drift may occur. SUMMARY

[0005] The application aims to provide an industrial robot precise path generation system based on three-dimensional vision reconstruction to solve the problems in the background.

[0006] In order to achieve the above-mentioned purpose, the application provides the following technical scheme: an industrial robot precise path generation system based on three-dimensional vision reconstruction, which comprises: An environment perception device, which is composed of a multi-baseline structured light projector and a multi-camera array, is used to acquire point cloud data of a workpiece surface under a preset illumination condition, and a light interference suppression unit and a spectral correction unit are used to filter noise and compensate brightness of the point cloud data, so as to ensure that the data meets the geometric accuracy of three-dimensional reconstruction. A three-dimensional reconstruction processing unit is configured to receive the point cloud data and perform multi-view registration based on feature point matching and rigid transformation estimation, and fuse the registration result with a deep learning error correction network to generate a three-dimensional geometric model containing curvature distribution, normal vector and boundary point set; A geometric feature analysis unit is configured to input the three-dimensional geometric model, analyze the main direction of curvature, and jointly determine the second derivative change rate and curvature gradient to segment and fit the surface of the workpiece, and output a region description with surface constraint parameters; A trajectory planning device is configured to generate a continuous three-dimensional trajectory using a weighted B-spline interpolation method combining curvature upper limit control and energy minimization principle based on the region description, and introduce arc length regularization constraint and node density adaptive adjustment at trajectory nodes to ensure trajectory smoothness and reduce speed fluctuation at turning points. A dynamic correction unit is configured to receive visual feedback and force sensing sensor signals during trajectory execution, and select feedback mode based on error threshold, when the error is less than the preset threshold, the visual feedback is mainly used for correction, when the error is greater than the threshold, the force feedback is used for fine tuning, and the correction result is consistent with the curvature and energy constraints in the trajectory planning stage. A control conversion processor is configured to convert the corrected trajectory into robot joint control instructions, the processor performs inverse kinematics solving and dynamics compensation, and through a trajectory prediction unit, potential deviations are corrected in advance in high-speed sections to ensure the realizability of the trajectory in actual execution. An execution driving mechanism including a servo motor and a closed-loop feedback circuit is configured to receive the joint control instructions and drive the end effector, and in long-period operation, the feedback signal is smoothed by combining an error accumulation filter to avoid trajectory drift and keep the end effector motion consistent with the corrected trajectory.

[0007] Preferably, the environment perception device includes a multi-baseline structured light projector and a multi-view camera array, the multi-baseline structured light projector is provided with different pitch projection baselines for projecting stripe patterns to the workpiece surface to avoid measurement blind area generated under single baseline condition, and the multi-view camera array includes at least two cameras, preferably three cameras, which are respectively arranged at different angles to obtain multi-angle point cloud data of the same workpiece surface. In the point cloud acquisition process, the ambient light may interfere with the recognition effect of the structured light, so a light interference suppression unit is arranged in the environment perception device, a band-pass filtering and time sequence coding method is used to distinguish the projected light and the background light, so as to filter out the ambient light noise, further, a spectrum correction unit is introduced, the image data obtained by the multiple cameras is subjected to gray balance and brightness compensation, so as to correct the brightness inconsistency between different cameras caused by the difference in light conditions, the gray balance can be realized by using the histogram equalization method, the brightness compensation is completed by linear gain adjustment, and the compensation coefficient range is preferably 0.8-1.2, and the point cloud data after the processing is taken as the input of the three-dimensional reconstruction processing unit.

[0008] Preferably, the three-dimensional reconstruction processing unit receives the point cloud data collected by the environment perception device, performs multi-view registration and fusion, first extracts the feature points of the point cloud, and uses a rigid transformation model to perform coarse registration on the multi-view point cloud; On the basis of coarse registration, a pre-trained convolutional neural network model is called in the fusion process to correct the registration residual error, the neural network is trained on a point cloud alignment data set, the input is the coordinate difference value between the multi-view point pairs, and the output is the corresponding residual vector, in this way, the registration error caused by occlusion and light difference is reduced, and finally the generated three-dimensional geometric model not only contains point coordinates, but also includes curvature distribution, normal vector and boundary point set, which are taken as the input of the geometric feature analysis unit.

[0009] Preferably, the geometric feature analysis unit takes the model output by the three-dimensional reconstruction processing unit as the input, first analyzes the curvature principal direction of the model surface, calculates the curvature distribution of each region, and the curvature calculation can use the following formula: ; In the formula, κ represents the curvature, x and y are the coordinate functions of the points on the curve; x' and y' are the first-order derivatives of the curve coordinate pair with respect to the arc length parameter s, that is, 、 ; And are the second-order derivatives of the curve coordinate pair with respect to the arc length parameter, that is, 、 The above formula is used to calculate the curvature of the local curve in the two-dimensional plane, so as to describe the surface change trend of the local region of the three-dimensional geometric model. After the curvature calculation is completed, the rate of change of the second-order derivative and the curvature gradient value are further compared for identifying the boundary region and the curvature mutation region in the model surface, and the curvature gradient can be represented as: ; In the formula, represents the gradient vector of the curvature, 、 respectively, are partial derivatives of curvature value in x and y direction, the gradient is used to measure the rate of change of curvature in the spatial direction, so as to identify the mutation or boundary position in the geometric model; The identified boundary area and mutation area are processed by segmentation and piecewise fitting to form a local surface patch, the fitting can be based on polynomial fitting or spline fitting method, and a surface expression satisfying the constraint condition is generated, and finally a region description containing the surface constraint parameters is output as the input of the trajectory planning device.

[0010] Preferably, the trajectory planning device receives the region description of the geometric feature analysis unit, and sets a curvature upper limit constraint when generating the trajectory, and the constraint relationship can be represented as: ; In the formula, represents the curvature of the trajectory at the arc length s, represents the curvature upper limit value, which is determined according to the maximum turning angle and the arm length of the robot manipulator, and is preferably set in the range of 0.01-0.05 mm -1 Under the above curvature constraint, the trajectory optimization objective is established based on the energy minimization principle; Under the constraint of the optimization objective, a weighted B-spline interpolation method is used to generate a continuous three-dimensional trajectory; the node distribution of the trajectory is realized through arc length regularization constraint, and the node spacing can be adaptively adjusted in the range of 1-5 mm according to the geometric complexity of the workpiece surface, and finally a three-dimensional trajectory is generated as the input of the dynamic correction unit.

[0011] Preferably, the dynamic correction unit receives the trajectory generated by the trajectory planning device when the trajectory is executed, and simultaneously collects the feedback of the visual sensor and the force sensor; The system sets a preset error threshold, which is set according to the accuracy requirement of the end effector, for example, 0.5 mm for high-precision assembly and 2 mm for ordinary carrying, when the detected trajectory error is not greater than the threshold, the correction based on visual feedback is executed; When the error is greater than the threshold, local correction combined with visual feedback and force feedback is executed, and the corrected trajectory is used as the input of the control conversion processor, which is different from the scheme relying on single visual correction. The present application realizes multi-modal feedback correction through clear threshold switching standard, avoiding ambiguous description.

[0012] Preferably, the control conversion processor receives the corrected trajectory output by the dynamic correction unit, and first executes inverse kinematics solving to convert the trajectory point into joint angle instructions; In the generation process, the inertia moment, friction coefficient and load parameter of the joint are combined for dynamics compensation; ​In addition, the processor comprises a trajectory prediction unit, which predicts the position and velocity state of the end effector based on a Kalman filtering method, and outputs a compensation control instruction in advance when the predicted deviation exceeds the range of 1-3mm of the reference trajectory, and finally generates joint control instructions as the input of the execution driving mechanism.

[0013] Preferably, the execution driving mechanism comprises a servo motor for receiving the joint control instructions output by the control conversion processor and driving the end effector to move, and a closed-loop feedback circuit for collecting joint position feedback signals in real time and smoothing the signals through an error accumulation filter. The smoothed feedback signals are used to correct the driving output of the servo motor in real time, so that the end effector movement process can accurately follow the corrected trajectory, unlike the traditional servo closed-loop control which only performs instant feedback.

[0014] The beneficial effects of the present application are as follows: 1、The present application proposes a scheme combining multi-baseline structured light and multi-view camera in the perception and modeling link, and sets up a light interference suppression and spectral correction module at the acquisition end, effectively overcoming the problems of large noise and uneven brightness of existing point clouds under complex lighting or local occlusion, and the point clouds after the preprocessing enter the three-dimensional reconstruction unit, first complete preliminary alignment based on feature point matching, and then use a pre-trained convolutional neural network to correct the residual, which on the one hand ensures the geometric accuracy of the registration between point clouds, and on the other hand improves the stability of the model under different acquisition conditions, and the generated geometric model contains curvature distribution, normal vector and boundary point set, providing high-quality data input for subsequent feature analysis.

[0015] 2、The present application, through the cooperation of geometric feature analysis and trajectory planning in the trajectory generation link, makes up for the defects of the traditional method which only pays attention to geometric fitting and leads to insufficient smoothness and execution, the analysis unit analyzes the main direction of curvature of the three-dimensional model, and combines the second derivative change rate and curvature gradient to determine the boundary and mutation area, generates a regional description with constraint parameters, the trajectory planning device introduces curvature upper limit constraint in the trajectory construction process according to this, and takes the weighted sum of joint angular velocity and angular acceleration as the energy optimization target, generates a continuous trajectory under the constraint by using weighted B-spline method, and combines arc length regularization and node adaptive adjustment to ensure the continuity and dynamics feasibility of the trajectory in the transition area.

[0016] 3、The present application is through in track execution and control link, the present application is directed to the prior art in the correction lag, the execution is easy to drift's insufficient, the closed loop structure of dynamic correction, control conversion and execution drive is designed, the dynamic correction unit switches between vision and force feedback through error threshold, so that the correction strategy can adapt to different error scenarios;Control conversion processor introduces inertia moment, friction coefficient and load parameter for dynamics compensation based on inverse kinematics solution, and uses trajectory prediction unit to identify potential deviation in advance and output compensation instruction by Kalman filtering method;The execution drive mechanism adopts the closed loop feedback mode of servo motor combined with sliding window filtering, suppresses signal jitter and error accumulation in long period operation. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The present application is based on a three-dimensional visual reconstruction of an industrial robot precise path generation system flow chart. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0019] As shown in Figure 1 The present application provides an industrial robot precise path generation system based on three-dimensional visual reconstruction, which comprises: An environment perception device composed of a multi-baseline structured light projector and a multi-view camera array, used to collect point cloud data of the workpiece surface under a preset lighting condition, and to filter out noise and compensate brightness of the point cloud data through a light interference suppression unit and a spectrum correction unit, so as to ensure that the data meet the geometric accuracy of three-dimensional reconstruction; A three-dimensional reconstruction processing unit for receiving point cloud data and performing multi-view registration based on feature point matching and rigid transformation estimation, while combining a deep learning error correction network for fusion, thereby generating a three-dimensional geometric model containing curvature distribution, normal vector and boundary point set; A geometric feature analysis unit taking the three-dimensional geometric model as input, using curvature principal direction analysis, second derivative change rate and curvature gradient joint discrimination to perform surface segmentation and piecewise fitting on the workpiece surface, and outputting a region description with surface constraint parameters; A trajectory planning device, which generates a continuous three-dimensional trajectory by using a weighted B-spline interpolation method combining curvature upper limit control and energy minimization principle according to the region description, and introduces arc length regularization constraint and node density adaptive adjustment at trajectory nodes to ensure trajectory smoothness and reduce speed fluctuation at turning points. a dynamic correction unit, which receives visual feedback and force sensor signals simultaneously during trajectory execution, and selects a feedback mode based on an error threshold, when the error is less than a preset threshold, visual feedback is mainly used for correction, when the error is greater than the threshold, fine tuning is combined with force feedback, and the correction result remains consistent with the curvature and energy constraints in the trajectory planning stage; a control conversion processor, which converts the corrected trajectory into robot joint control instructions, the processor performs inverse kinematics solving and dynamics compensation, and through a trajectory prediction unit, potential deviations are corrected in advance in the high-speed section to ensure the realizability of the trajectory in actual execution; an execution driving mechanism, including a servo motor and a closed-loop feedback circuit, for receiving joint control instructions and driving the end effector, and in the long-period operation process, a error accumulation filter is used to smooth the feedback signal to avoid trajectory drift and keep the end effector motion consistent with the corrected trajectory.

[0020] The environment perception device includes a multi-baseline structured light projector and a multi-view camera array. The multi-baseline structured light projector is provided with different pitch projection baselines for projecting a stripe pattern onto the workpiece surface, thereby avoiding measurement blind areas caused by a single baseline condition. The multi-view camera array includes at least two cameras, preferably three cameras, which are arranged at different angles to obtain multi-angle point cloud data of the same workpiece surface. During point cloud acquisition, ambient light may interfere with the recognition effect of structured light. Therefore, a light interference suppression unit is provided in the environment perception device to distinguish between projected light and background light using a band-pass filter and a time sequence coding method to filter out ambient light noise. Further, a spectral correction unit is introduced to perform gray balance and brightness compensation on the image data obtained by the multi-camera to correct the brightness inconsistency between different cameras due to differences in lighting conditions. Gray balance can be achieved using a histogram equalization method, and brightness compensation is completed through linear gain adjustment. The compensation coefficient range is preferably 0.8-1.2. The point cloud data processed in this way is used as the input of the three-dimensional reconstruction processing unit. Compared with the existing scheme of directly collecting point cloud data relying on a single camera, the combination of multi-baseline structured light + multi-view camera + light interference suppression + spectral correction ensures the stability and consistency of the point cloud input data, thereby providing a high-quality data basis for subsequent three-dimensional reconstruction and trajectory planning.

[0021] The three-dimensional reconstruction processing unit receives the point cloud data collected by the environment perception device, performs multi-view registration and fusion, first extracts point cloud feature points, and uses a rigid transformation model to coarsely register multi-view point clouds. The registration relationship can be expressed as: ; In the formula, is the original point coordinate, for the registered point coordinates, for a rotation matrix, for a translation vector; On the basis of coarse registration, a pre-trained convolutional neural network model is called in the fusion process to correct the registration residual error, the neural network is trained on a point cloud alignment data set, the input is the coordinate difference value between the points in multiple views, and the output is the corresponding residual vector, in this way, the registration error caused by occlusion and light difference is reduced, and finally the generated three-dimensional geometric model not only includes point coordinates, but also includes curvature distribution, normal vector and boundary point set, as the input of the geometric feature analysis unit; Different from the existing ICP or two-dimensional projection, the application adds deep learning correction outside the traditional registration algorithm, and discloses the input, output and training source, so as to ensure the implementability of this link.

[0022] Among them, the geometric feature analysis unit takes the model output by the three-dimensional reconstruction processing unit as input, first analyzes the curvature main direction of the model surface, calculates the curvature distribution of each region, and the curvature calculation can adopt the following formula: ; In the formula, κ represents curvature, x and y are coordinate functions of points on a curve; x' and y' are first-order derivatives of the curve coordinate with respect to the arc length parameter s, that is, 、 ; and are second-order derivatives of the curve coordinate with respect to the arc length parameter, that is, 、 ; the above formula is used to calculate the curvature of the local curve in the two-dimensional plane, so as to describe the surface change trend of the local region of the three-dimensional geometric model; After completing the curvature calculation, further comparison is made between the change rate of the second-order derivative and the curvature gradient value, which is used to identify the boundary region and the curvature mutation region in the model surface, and the curvature gradient can be represented as: ; In the formula, represents the gradient vector of the curvature, 、 are partial derivatives of the curvature value in x and y directions respectively, and the gradient is used to measure the change rate of the curvature in the spatial direction, so as to identify the mutation or boundary position in the geometric model; The identified boundary region and mutation region are processed by segmentation and piecewise fitting to form a local surface patch, the fitting can be based on a polynomial fitting or a spline fitting method, and a surface expression meeting the constraint condition is generated, and finally a region description containing surface constraint parameters is output as the input of the trajectory planning device; Unlike the existing method of directly fitting a surface to the whole point cloud, the present application not only calculates the curvature distribution, but also accurately locates the boundary region through the joint determination of the second derivative change rate and the curvature gradient, thereby introducing a constraint condition when generating the region description, ensuring the integrity and pertinence of the trajectory planning input data.

[0023] wherein the trajectory planning device receives the region description of the geometric feature analysis unit and sets a curvature upper limit constraint when generating the trajectory, and the constraint relationship can be represented as: ; wherein, represents the curvature of the trajectory at the arc length s, represents the curvature upper limit value, which is determined according to the maximum turning angle and the arm length of the robot manipulator, and is preferably set in the range of 0.01-0.05 mm -1 ; Under the above curvature constraint, the trajectory optimization objective is established based on the energy minimization principle, and the objective function can be represented as: ; wherein, is the optimization objective function; is the angle of the th joint; and are the angular velocity and angular acceleration, respectively; is the number of joints; and are weight coefficients, which can be determined according to the robot load condition and the trajectory smoothness requirement; Under the constraint of this optimization objective, a weighted B-spline interpolation method is used to generate a continuous three-dimensional trajectory; the node distribution of the trajectory is realized through an arc length regularization constraint, and the node spacing can be adaptively adjusted in the range of 1-5 mm according to the geometric complexity of the workpiece surface, and the finally generated three-dimensional trajectory is input to the dynamic correction unit; Unlike the existing trajectory generation method which simply relies on B-spline interpolation or random sampling, the present application explicitly proposes the dual conditions of curvature constraint and energy optimization, which not only ensures the geometric feasibility of the trajectory, but also has smoothness and executability in dynamics, thereby improving the stability and precision of the trajectory planning.

[0024] wherein the dynamic correction unit receives the trajectory generated by the trajectory planning device when the trajectory is executed, and simultaneously collects the feedback of the visual sensor and the force sensor; The system sets a preset error threshold, which is set according to the end effector accuracy requirement, for example, 0.5 mm for high-precision assembly and 2 mm for ordinary handling, and when the detected trajectory error is not greater than the threshold, the correction based on the visual feedback is executed; When the error is greater than the threshold value, local correction combined with visual feedback and force feedback is performed, and the corrected trajectory is taken as the input of the control conversion processor. Unlike the scheme relying on single visual correction, the present application realizes multi-modal feedback correction through explicit threshold switching criteria, avoiding ambiguous expressions.

[0025] Wherein, the control conversion processor receives the corrected trajectory output by the dynamic correction unit, first performs inverse kinematics solving to convert the trajectory point into joint angle instructions, and the relationship can be expressed as: : In the formula, represents a joint angle vector, represents the desired pose of the end effector; During generation, dynamics compensation is performed in combination with the inertia moment, friction coefficient and load parameters of the joint, and its calculation model can be expressed as: : In the formula, is a joint torque vector; is a mass matrix related to the joint angle, is the Coriolis force and centrifugal force term; is the gravity term; and are the joint angular velocity and angular acceleration, respectively, and the above compensation can offset the deviation caused by inertia and external disturbance; In addition, the processor includes a trajectory prediction unit which, based on the Kalman filtering method, predicts the position and velocity state of the end effector, and when the predicted deviation exceeds the 1-3mm range of the reference trajectory, the compensation control instruction is output in advance, and the finally generated joint control instruction is taken as the input of the execution driving mechanism; Unlike the existing method which only focuses on geometric coordinate conversion, the present application combines dynamics compensation on the basis of inverse kinematics, and adds a prediction judgment threshold at the control layer, thereby ensuring the accuracy and stability of the trajectory during actual execution.

[0026] Wherein, the execution driving mechanism includes a servo motor and a closed-loop feedback circuit, the servo motor is used to receive the joint control instruction output by the control conversion processor and drive the end effector to move, and the closed-loop feedback circuit is used to collect the joint position feedback signal in real time and smooth the signal through an error accumulation filter; The filter adopts a sliding window weighted average method, and its calculation relationship can be expressed as: ; In the formula, represents the smoothed signal obtained at the th sampling time, represents the smoothed signal obtained at the a raw feedback signal of a sampling moment, represents a window length, under a sampling frequency of 100Hz, The value of is usually 10-50, so as to balance the signal smoothing effect and real-time performance; The smoothed feedback signal is used to correct the driving output of the servo motor in real time, so that the end effector movement process can accurately follow the corrected trajectory, which is different from the traditional servo closed-loop control which only performs instant feedback, The application introduces the error accumulation filtering method in the feedback link, and the window parameter range is defined, so as to ensure the stability of the trajectory execution and the reproducibility of the scheme.

[0027] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0028] Although the embodiments of the application have been shown and described, it is understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the application, and the scope of the application is defined by the appended claims and their equivalents.

Claims

1. An industrial robot precision path generation system based on three-dimensional vision reconstruction, characterized by: The system comprises: An environment perception device composed of a multi-baseline structured light projector and a multi-view camera array, used to collect point cloud data of the workpiece surface under preset lighting conditions, and to filter noise and compensate brightness of the point cloud data through a light interference suppression unit and a spectral correction unit; A three-dimensional reconstruction processing unit used to receive the point cloud data and perform multi-view registration based on feature point matching and rigid transformation estimation, while combining a deep learning error correction network for fusion, thereby generating a three-dimensional geometric model containing curvature distribution, normal vector and boundary point set; A geometric feature analysis unit taking the three-dimensional geometric model as input, using curvature principal direction analysis, second derivative change rate and curvature gradient joint discrimination to perform surface segmentation and piecewise fitting on the workpiece surface, and outputting a region description with surface constraint parameters; A trajectory planning device that generates a continuous three-dimensional trajectory using a weighted B-spline interpolation method combining curvature upper limit control and energy minimization principle based on the region description, and introduces arc length regularization constraint and node density adaptive adjustment at trajectory nodes; A dynamic correction unit that simultaneously receives visual feedback and force sensation sensor signals during trajectory execution, and selects feedback mode based on error threshold, with visual feedback as the main correction when error is less than the preset threshold, and fine-tuning combined with force feedback when error is greater than the threshold, with the correction result consistent with the curvature and energy constraints in the trajectory planning stage; A control conversion processor used to convert the corrected trajectory into robot joint control instructions, the processor performing inverse kinematics solving and dynamics compensation, and correcting potential deviations in advance through a trajectory prediction unit in the high-speed section; An execution driving mechanism including a servo motor and a closed-loop feedback circuit, used to receive the joint control instructions and drive the end effector, and to smooth the feedback signal by combining an error accumulation filter during long-period operation.

2. The system of claim 1, wherein: The environment perception device includes a multi-baseline structured light projector and a multi-view camera array, which collects point cloud data of the workpiece surface under preset lighting conditions, and filters environmental light noise in the point cloud through a light interference suppression unit, and then corrects the point cloud brightness through a spectral correction unit to obtain point cloud data that meets the input requirements of the three-dimensional reconstruction processing unit.

3. The system of claim 2, wherein: The three-dimensional reconstruction processing unit performs multi-view registration on the point cloud data collected by the environment perception device, the registration is based on feature point matching and rigid transformation estimation, and a trained convolutional neural network model is called during registration to correct registration errors, to fuse and generate a three-dimensional geometric model containing curvature distribution, normal vector and boundary point set, which serves as input data for the geometric feature analysis unit.

4. The system of claim 3, wherein: The geometric feature analysis unit receives a three-dimensional geometric model generated by the three-dimensional reconstruction processing unit, performs surface curvature calculation on the three-dimensional geometric model based on curvature principal direction analysis, determines boundary regions and curvature mutation regions in the model through comparison of second derivative change rate and curvature gradient value, performs segmentation and piecewise fitting on the regions, and outputs region descriptions containing surface constraint parameters, which are input data of the trajectory planning device.

5. The system for precise path generation for industrial robots based on 3D vision reconstruction as claimed in claim 4 wherein: The trajectory planning device receives the region descriptions output by the geometric feature analysis unit, sets an upper limit constraint on the curvature value during trajectory generation, establishes an optimization function based on the principle of energy minimization, generates a continuous three-dimensional trajectory using a weighted B-spline interpolation method, and introduces an arc length regularization constraint and a node density adaptive adjustment strategy at trajectory nodes, which are input data of the dynamic correction unit.

6. The system of claim 5, wherein: The dynamic correction unit receives the three-dimensional trajectory generated by the trajectory planning device during trajectory execution, collects visual sensor feedback and force sensor feedback, and determines a feedback mode based on a preset error threshold. When the detected trajectory error is not greater than the threshold, the trajectory correction based on visual feedback is performed. When the detected trajectory error is greater than the threshold, the local trajectory correction combining visual feedback and force feedback is performed. The corrected trajectory is input data of the control conversion processor.

7. The system of claim 6, wherein: The control conversion processor receives the corrected trajectory output by the dynamic correction unit, performs inverse kinematics solving on the corrected trajectory, converts the trajectory points into angle instructions of each joint of the robot, and performs dynamics compensation during instruction generation to offset joint motion inertia and load disturbance. The control conversion processor further includes a trajectory prediction unit. When the robot is running at high speed, the trajectory prediction unit predicts the trajectory based on historical trajectory data and real-time feedback, and outputs compensation control instructions in advance when the predicted trajectory deviation exceeds a set range. The joint control instructions are input of the execution driving mechanism.

8. The system of claim 7, wherein: The execution driving mechanism includes a servo motor and a closed-loop feedback circuit. The servo motor receives the joint control instructions output by the control conversion processor and drives the end effector to move. The closed-loop feedback circuit collects joint position feedback signals during operation, and smoothes the feedback signals through an error accumulation filter. The smoothed signals are used to adjust the driving output of the servo motor in real time, so that the end effector performs actions according to the corrected trajectory.

Images