A motion redirection height adaptive calibration method based on ground contact analysis
A multi-level height adaptive calibration method based on ground contact analysis solves the problem of ground inconsistency in motion retargeting, achieving stability of ground contact and realism of motion, and is suitable for ground calibration in robot motion retargeting.
Patent Information
- Application Number
- CN202610412162.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-26
AI Technical Summary
In existing technologies, ground inconsistency issues during motion redirection lead to ground penetration and floating phenomena, affecting visual effects and contact modeling in physical simulations, increasing training difficulty and reducing the generalization ability of the strategy, especially in movements with rapid gait switching and large altitude changes.
A motion redirection height adaptive calibration method based on ground contact analysis is adopted. By acquiring dynamic video, the height of the root node and the lowest point is calculated, and a multi-level height adaptive calibration strategy is formulated. Combining sequence-level global calibration and frame-level real-time pull-back calibration, a safety buffer distance and low-pass filtering mechanism are introduced to eliminate artifacts and maintain the stability and physical feasibility of motion.
It effectively eliminates ground penetration and floating phenomena, maintains the continuity and realism of motion, improves the stability and applicability of robot motion, reduces computational overhead, and provides a reliable reference for subsequent learning and simulation.
Smart Images
Figure CN122290209A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of robot motion, specifically relating to a motion redirection height adaptive calibration method based on ground contact analysis. Background Technology
[0002] Motion retargeting aims to map human or source character movements onto target robots or virtual characters with different shapes, scales, and joint topologies. It is a key technology in robot imitation learning, physical simulation training, and digital human animation generation. The cost of acquiring human motion data has been significantly reduced, providing a rich data source for large-scale robot motion learning. However, in practical applications, source movements are often obtained under geometric and dynamic assumptions inconsistent with the target system. Humans and robots differ significantly in joint length ratios, foot structure, center of mass distribution, and contact models. Even if the joint angle sequence corresponding to the target robot is obtained through inverse kinematics or optimization methods, the retargeted movements can still easily exhibit inconsistencies with the ground.
[0003] Specifically, ground consistency issues manifest in two main types of artifacts: one is ground penetration, where the feet or other body parts are below the ground plane in some frames; the other is levitation, where the root node is raised as a whole to avoid penetration, resulting in an unnatural gap between the feet and the ground for extended periods. These two issues not only affect visual effects but also directly interfere with contact modeling in physical simulations, especially during rapid gait transitions, jumps, or squats, where significant height changes occur.
[0004] It is important to emphasize that ground artifacts not only affect animation and visualization effects, but also incur direct costs in robot learning and simulation: penetration causes unrealistic collision constraints and energy injection, making the simulated contact force distribution deviate from physical laws; levitation leads to missing or misaligned contact events, disrupting the temporal structure of the support phase. These problems significantly interfere with downstream reinforcement learning or imitation learning's modeling of contact timing and mechanical allocation, increasing training difficulty and reducing the generalization ability of the policy; simultaneously, they weaken the stability of the simulation-to-real transfer process, leading to failures such as slippage, missteps, or falls during real-world execution. Therefore, a general method is needed that can explicitly characterize ground contact constraints and adaptively select the height calibration granularity under different action types, preserving the spatiotemporal characteristics of the original action as much as possible while ensuring physical feasibility, providing a reliable reference for downstream learning and real-world execution. Summary of the Invention
[0005] This invention provides a motion redirection height adaptive calibration method based on ground contact analysis to address the technical challenges of existing geometric calibration and kinematic detection methods, which, although easy to integrate and low in engineering cost, still struggle to simultaneously address penetration elimination, levitation control, motion fidelity, and inter-frame stability in scenarios with diverse motion types and time-varying contact patterns.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A motion redirection height adaptive calibration method based on ground contact analysis includes the following steps: Step 1: Acquire a dynamic video segment and calculate the root node height and lowest point height of the skeleton model corresponding to each frame. Step 2: Based on the changing trends of the root node height and the lowest point height, formulate a multi-level height adaptive calibration strategy: for stable motion trends, use sequence-level global calibration; for non-stationary motion trends, use frame-level real-time pull-back calibration. Step 3: Perform artifact removal and stabilization processing on the calibrated dynamic video to complete the adaptive calibration.
[0007] Furthermore, let the dynamic video be of a given length of action sequence , When performing sequence-level global calibration, first calculate the global minimum height of the entire motion sequence. , in, This represents the height of the lowest point in the t-th frame of the image; Then, a uniform offset is applied vertically to the root node of the entire action sequence to align the global lowest point with the ground buffer height: in, Indicates the safe buffer distance; Let the original root node height of each frame be... The height of the calibrated root node is : When performing frame-level real-time pull-back calibration, the height of the lowest point of each frame image is first calculated. The root node height of the corresponding frame image is calibrated in real time using the following formula. : .
[0008] Furthermore, if the variance of the root node height is less than the first threshold, and the peak and valley values of the root node height change are less than the second threshold, or if the variance of the lowest point height is less than the third threshold, and the peak and valley values of the lowest point height change are less than the fourth threshold, then it is determined to be a stationary sequence and sequence-level global calibration is required. If the peak or valley value of the change in root node height or the change in minimum node height exceeds the fifth threshold, it is determined to be a non-stationary sequence and frame-level real-time pull-back calibration is required.
[0009] Furthermore, forward kinematics can be used to obtain the first frame of each image in the action sequence. Position of a body part in the world coordinate system , in, Indicates the number of body parts considered. Indicates from the root node to the... The kinematic mapping of each body part, with the root node representing the base node or Pelvis node in the skeleton model, at the waist of the robot; The height of the lowest point of the entire body in this frame is: in, Indicates position The vertical coordinates.
[0010] Furthermore, in step three, it is determined whether the root node of each frame in the calibrated dynamic video meets the penetration trigger condition. If it does, correction is performed, and then low-pass filtering smoothing is performed on the corrected root node; if it does not meet the condition, low-pass filtering smoothing is performed directly.
[0011] Furthermore, for any frame of image, if the penetration triggering condition is met... Or, the floating trigger condition is met. Then, the correction process is performed using the following formula: in, and Let represent the three-dimensional position column vectors of the robot's root node before and after correction in the world coordinate system, respectively. Indicates the floating tolerance threshold; After the correction is performed, a low-pass filter is applied for smoothing according to the following formula, and the correction amount for the unsmoothed root node vertical height is denoted as . The vertical height correction of the smoothed root node is Then we have: in, For smoothing coefficients, This represents the vertical height of the root node after low-pass filtering and smoothing.
[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. To address the differences in height variation characteristics across various actions, this invention employs two strategies: sequence-level global calibration and frame-level real-time calibration. Experimental results show that in stable actions with minimal height variation, global calibration maintains good motion continuity while ensuring consistent ground contact; however, in actions with significant height variation, such as jumping and squatting, frame-level real-time calibration more effectively avoids ground penetration and better preserves the original height characteristics of the action.
[0013] 2. To further improve the stability of the calibration process, the method of this invention introduces a safety buffer distance and a low-pass filtering mechanism, thereby alleviating the high jitter problem caused by frame-by-frame correction to a certain extent. Experimental verification on the MuJoCo simulation environment and the Unitree G1 humanoid robot model shows that the proposed method can stably improve ground contact consistency in various types of actions, and has the advantages of simple implementation and low computational cost.
[0014] 3. Since the method of this invention is only used as a geometric calibration module after the relocation result and does not rely on additional dynamic optimization or force feedback information, it can be easily integrated into the existing motion relocation and imitation learning process, providing a more reliable reference motion for subsequent reinforcement learning training and simulation to real-world transfer. At the same time, complex multi-point contact requires more complex methods. Future work will combine more complex contact models and mechanical feedback information to further extend the height calibration strategy, so as to improve the applicability of the method in complex scenarios and real robot execution. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram comparing the root node height under different strategies of the present invention; Figure 3 This is a schematic diagram comparing the lowest point height under different strategies of the present invention; Figure 4 This is an example of the sequence-level global height calibration effect of the present invention. Icon (a) shows the redirection result without height calibration; icon (b) shows the result after applying sequence-level global height calibration. Figure 5The following is an example of the frame-level real-time height calibration effect of the present invention: Icon (a) shows the redirection result without height calibration; Icon (b) shows the result after frame-level real-time height calibration. Figure 6 This is a schematic diagram comparing the impact of the safety buffer distance δ of the present invention on the stability of the lowest point height. Detailed Implementation
[0016] To make the technical means, creative features, objectives and effects of the present invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the motion redirection height adaptive calibration method based on ground contact analysis of the present invention. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.
[0017] like Figure 1 As shown, this invention proposes a height adaptive calibration method based on ground contact analysis to eliminate ground penetration and levitation artifacts in motion retargeting results. The overall method flow is as follows: Given a retargeted robot motion sequence, firstly, forward kinematics is used to calculate the lowest point height of the entire body (or contact candidate parts) in each frame, constructing geometric and physical constraints that ensure the lowest point is not lower than the ground, and using this as a unified criterion for penetration and levitation artifacts; secondly, a multi-level height calibration strategy is designed based on the height change characteristics of the motion: sequence-level global calibration is used for stable motion to maintain overall height consistency and reduce computational overhead; frame-level real-time pullback is used for dynamic motions such as jumping and squatting to more accurately maintain local height characteristics and contact accuracy; finally, a safety buffer distance δ is added to the calibration closed loop to improve numerical and visual stability, and low-pass filtering is introduced to smooth the height abrupt changes caused by frame-by-frame correction, thereby obtaining stable, natural, and physically feasible retargeting results under different motion types. Experiments show that this method can effectively eliminate penetration and levitation artifacts while maintaining high motion fidelity.
[0018] Specifically as follows: I. Lowest Point Detection and Ground Constraints 1.1 Forward Kinematics and Definition of Minimum Point Given a length of action sequence That is, a video with a total number of frames T, where the target system (robot or character) is at time T. The posture is determined by the joint angle vector. This indicates that the forward kinematics can be used to obtain the first image frame of each frame. The position of a body part (or key point) in the world coordinate system: (1) in, The number of body parts to be considered. Indicates from the root node to the... The kinematic mapping of each body part is given, where the root node represents the base node or Pelvis node in the skeleton model, located at the robot's waist. This invention defines the height of the lowest point of the entire body in this frame as: (2) in, for The vertical coordinates. To improve robustness, in normal walking / standing scenarios, the minimum value can be calculated only within the contact candidate set (both feet); in rolling, kneeling, or multi-point contact scenarios, the candidate set can be expanded to cover parts such as knees and hands.
[0019] 1.2 Ground Model and Penetration Measurement This invention assumes the ground is a plane, and the ground height is denoted as . Ground penetration depth is defined as: (3) when When, it indicates the presence of a penetration artifact; when Furthermore, when the foot is in the contact phase, it corresponds to a tendency to float. The objective of this invention is to satisfy the following after calibration: (4) 1.3 Safety buffer distance Due to numerical errors, discretization, and collision detection accuracy issues in simulation, the lowest point coinciding with the ground could induce contact jitter or visual Z-fighting. Therefore, a safety buffer distance is introduced. Write the constraints as (5) II. Multi-level height adaptive calibration strategy Different actions exhibit significant differences in height distribution: the height changes during steady walking and standing are approximately stable, while actions such as jumping and squatting show obvious non-stationary height peaks and troughs. To achieve a balance between physical contact consistency, motion fidelity, and computational efficiency, this invention proposes a multi-level height calibration strategy: sequence-level global calibration is used for stable motion, while frame-level real-time pull-back is used for non-stationary motion.
[0020] 2.1 Sequence-level global calibration Given a length of action sequence For a video with a total of T frames, first calculate its global minimum height: (6) Then, a uniform offset is applied vertically to the root node of the entire action sequence to align the global lowest point with the ground buffer height: (7) Let the position of the root node be... The height of the calibrated root node is: (8) This strategy has two advantages: (1) the offset can be obtained by scanning the sequence once, with very low computational overhead; (2) a consistent translation is applied to the entire action sequence without destroying the relative height relationship between frames, making it suitable for smooth actions with small height changes. Its limitation is that when the sequence contains obvious non-stationary height changes (such as jump apex, squat trough), the global offset may compress the true height range of the action.
[0021] 2.2 Frame-level real-time pullback For non-stationary motion sequences, this invention employs frame-level calibration. The minimum point height is calculated independently in each frame, and the root node is corrected in real time. (9) (10) This guarantees that: (11) In other words, penetration is eliminated instantly in each frame. This strategy can preserve the local height features of actions such as jumping and crouching, and is more adaptable to time-varying height distributions. However, its frame-by-frame correction is easily sensitive to noise and may introduce height jitter, so it needs to be combined with a smoothing mechanism.
[0022] 2.3 Principles of Strategy Selection This multi-level framework can switch between different levels using a simple height change criterion: when the root node height or lowest point height variance of the sequence is small and there are no obvious height peaks or valleys, it is determined to be a stationary motion sequence, and global calibration is used; when a significant non-stationary height interval is detected (such as height change kurtosis or velocity peak exceeding a threshold due to jumping / crouching), it is determined to be a non-stationary motion sequence, requiring a switch to frame-level pullback. The effectiveness and applicability of the two strategies are reported in the experiments of this invention.
[0023] In other words, if the variance of the root node height is less than the first threshold and the peak and valley values of the root node height change are less than the second threshold, or if the variance of the lowest point height is less than the third threshold and the peak and valley values of the lowest point height change are less than the fourth threshold, then it is determined to be a stationary sequence and sequence-level global calibration is required. The first to fourth thresholds can be determined according to actual needs. If the peak or valley value of the change in root node height or the change in minimum node height exceeds the fifth threshold (which needs to be determined according to the actual situation), it is determined to be a non-stationary running sequence, and frame-level real-time pull-back calibration is required.
[0024] 3. Perform artifact removal and stabilization processing on the calibrated action sequence. 3.1 Penetration and levitation correction closed loop Combining the minimum point constraints and height calibrations in sections 2.1 and 2.2, this invention constructs a unified artifact correction method. For any frame, if the penetration triggering condition is met... Or, the floating trigger condition is met. ,(in If the floating tolerance threshold is 0.2m, then a correction will be performed: (12) in, and These are the three-dimensional position column vectors of the robot's root node before and after correction in the world coordinate system. The above formula means that the vector addition is the element-wise addition of the corresponding dimensions, and only the vertical height (z-axis) is corrected, without changing the horizontal position.
[0025] This correction can be viewed as applying the minimum necessary vertical compensation to the root node, thereby restoring the lowest point to the ground buffer height. For the floating problem, frame-level calibration naturally avoids overcompensation by raising the robot as a whole to eliminate penetration; global calibration maintains consistency in stable scenarios, avoiding false contact phenomena caused by prolonged ground lifting.
[0026] 3.2 High smoothness After performing the correction in step 3.1, to suppress jitter caused by noise in the lowest point estimation or switching of contact areas during frame-level pullback, this invention performs low-pass filtering to smooth the root node vertical offset. Let the unsmoothed root node vertical height correction be denoted as... The vertical height correction of the smoothed root node is Then we have: (13) (14) in, For smoothing coefficients, This represents the vertical height of the root node after low-pass filtering and smoothing. (Larger) Retains greater immediacy, smaller This provides a smoother height sequence. In the experiment, this invention used a fixed... And analyze its impact on stability and fidelity.
[0027] After low-pass filtering, the core changes in the vertical height of the root node are divided into two parts: the invariant core constraints and the significantly optimized characteristics: 1. Smoothing does not violate the core requirements of ground contact; 2. The inter-frame height continuity is greatly improved, and high-frequency jitter is completely suppressed.
[0028] In summary, the method of this invention constitutes a closed loop through minimum point detection, multi-level height calibration, buffering, and smoothing stabilization: the minimum point provides an interpretable geometric metric for ground consistency, multi-level calibration adaptively selects global or local compensation under different motion types, and safe buffer distance and low-pass filtering further enhance numerical and visual stability. This closed loop significantly improves the physical feasibility of redirected motion while maintaining the original spatiotemporal structure of the motion, laying the foundation for subsequent experimental verification.
[0029] To verify the feasibility of this invention, we conducted the following experiment: 1. Experimental setup The proposed height adaptive calibration method was evaluated in the simulation platform. The MuJoCo simulation environment was used as the physics engine, and the 29-DOF Unitree G1 was selected as the experimental robot model. The motion retargeting and height calibration pipeline was implemented based on the GMR (General Motion Retargeting) framework. The IK solution in GMR was completed jointly by the mink library and MuJoCo. Without changing the original IK matching table and joint constraint configuration, this invention only inserts the height calibration module after the retargeting result and uses the forward kinematics interface to calculate the position of each joint in the world coordinate system. The motion source of this experiment is the motion sequence based on SMPLX representation extracted from the video platform through the GVHMR project. In order to highlight the height behavior difference, this invention divides the motion sequence into two categories: (1) Stable type: such as standing, walking slowly, slight trunk swaying, etc., with small changes in the height of the root node and the lowest point; (2) Dynamic type: such as jumping, squatting, quick stepping up and down, kicking, etc., with obvious height peaks and valleys and contact reorganization. For each type of action, several representative templates were selected and redirected in the robot to verify the robustness of the method to scale differences. At the same time, this experiment compared the following ground calibration strategies: (1) the original redirection result without calibration; (2) sequence-level global height offset; (3) frame-level real-time pullback; (4) frame-level real-time pullback with low-pass smoothing.
[0030] 2. Results Comparison (1) Comparison of the overall effects of different height calibration strategies. First, the performance of three strategies—uncalibrated, sequence-level global height offset, and frame-level real-time pull-back—in redirection actions is compared. Experimental results show that uncalibrated actions exhibit varying degrees of ground penetration or levitation artifacts in all scenarios. Sequence-level global calibration can effectively eliminate penetration in stable actions, but it tends to compress the true height range in dynamic actions with significant height changes, introducing levitation or weakening key action features. In contrast, frame-level real-time pull-back ensures that the lowest point is not lower than the ground in each frame, significantly reducing penetration in dynamic actions and better preserving the height change characteristics of the original action. Comparison of root node heights of robots using different strategies in the same action is shown below. Figure 2 As shown, the X-axis represents the frame index, and the Y-axis represents the robot's root height. The blue line represents the uncalibrated root height change curve (Root (no calibration)), where a low point appears around frame 100, meaning the robot's feet sink below the ground. The orange line represents the global offset (Root (global offset)), which raises other points after the low point, causing the robot to float in the air. The green line represents frame-wise adjustment (Root (frame-wise)), showing that the root height does not fluctuate significantly, meaning the robot's feet are basically near the ground. The red line represents adding low-pass filtering to the frame-wise adjustment (Root (frame-wise + low-pass)), resulting in much smoother movement compared to without low-pass filtering. Based on these four curves, it can be seen that the root height is much better after adopting the height calibration strategy compared to the uncalibrated version, and some unexpected low points are brought back to their corresponding heights. Meanwhile, Figure 3The X-axis represents the frame index, and the Y-axis represents the robot's minimum height. The blue dashed line represents the ground height (Groud + δ). The blue curve represents the robot's minimum height without calibration, showing that the robot's minimum height is below the ground at frames 70 and 210. The orange curve represents the global offset, which pulls the minimum height above the ground, including around frame 210. The green curve represents frame-wise pullback, showing a significant reduction in the previous sinking and floating phenomena, with the robot's minimum height near the ground at frames 70 and 210. The red curve represents frame-wise pullback combined with low-pass filtering, resulting in a smoother curve and more fluid robot movement compared to not using low-pass filtering. The above clearly demonstrates the relationship between the robot's minimum height and the ground under different strategies, showing that the minimum height will not appear below the ground after calibration.
[0031] (2) Performance analysis under different action types Further analysis by action type revealed that sequence-level global calibration performed stably in stable movements with minimal height changes, such as standing and walking slowly, but its effectiveness was limited in dynamic movements such as jumping and squatting. Frame-level real-time pull-back showed a significant advantage in dynamic movements, such as... Figure 4 The robot performs a slow walking motion. Figure 4 In icon (a), the robot is floating at approximately a distance from the ground, with its front and back at roughly the same height. Figure 4 In icon (b), after sequence and global calibration, the distance between the robot and the ground is essentially eliminated; however, in complex contact scenarios, the switching of the lowest point between different body parts can easily introduce height jitter, such as in... Figure 5 The robot performs relatively vigorous dance moves. Figure 5 In icon (a), no calibration strategy was used; the robot sometimes had its feet sinking into the ground and sometimes floated into the air. Figure 5 The use of frame-level real-time pullback in icon (b) effectively avoids this problem and preserves the local height change characteristics of the motion well. This phenomenon indicates that different motion types have significantly different requirements for height calibration granularity, and also verifies the necessity of adopting a height adaptive calibration strategy.
[0032] (3) The impact of stabilization mechanisms Based on frame-level real-time pull-back, a safety buffer distance δ is introduced. After m-pass filtering, the continuity of the root node altitude sequence is significantly improved. This stabilization mechanism effectively suppresses altitude jitter caused by noise and contact switching without violating ground contact constraints, making the calibrated motion more natural and stable both visually and geometrically. Specifically, as follows... Figure 6 As shown, the X-axis represents the frame sequence, the Y-axis represents the lowest point height, and the blue dashed line represents the ground height. The blue curve represents the robot's lowest point position without using a safety buffer distance and low-pass filtering (δ=0), showing instances of sinking below the ground and high-frequency jitter. The orange dashed line represents the position with the introduction of a safety buffer distance and low-pass filtering (δ>0), increasing the safety redundancy in distance from the ground and significantly reducing high-frequency jitter. Based on the above, using a safety buffer distance and low-pass filtering significantly eliminates high-frequency jitter.
[0033] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples. Various changes or modifications can be made to these embodiments without departing from the principles and essence of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims.
[0034] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications or variations that can be made by those skilled in the art without creative effort within the scope of the appended claims are still within the scope of protection of this patent.
Claims
1. A motion redirection height adaptive calibration method based on ground contact analysis, characterized in that: Includes the following steps: Step 1: Acquire a dynamic video segment and calculate the root node height and lowest point height of the skeleton model corresponding to each frame. Step 2: Based on the changing trends of the root node height and the lowest point height, formulate a multi-level height adaptive calibration strategy: for stable motion trends, use sequence-level global calibration; for non-stationary motion trends, use frame-level real-time pull-back calibration. Step 3: Perform artifact removal and stabilization processing on the calibrated dynamic video to complete the adaptive calibration.
2. The motion redirection height adaptive calibration method based on ground contact analysis according to claim 1, characterized in that: Let the dynamic video be of a given length. action sequence , When performing sequence-level global calibration, first calculate the global minimum height of the entire motion sequence. , in, This represents the height of the lowest point in the t-th frame of the image; Then, a uniform offset is applied vertically to the root node of the entire action sequence to align the global lowest point with the ground buffer height: in, Indicates the safe buffer distance; Let the original root node height of each frame be... The height of the calibrated root node is : When performing frame-level real-time pull-back calibration, the height of the lowest point of each frame image is first calculated. The root node height of the corresponding frame image is calibrated in real time using the following formula. : 。 3. The motion redirection height adaptive calibration method based on ground contact analysis according to claim 2, characterized in that: If the variance of the root node height is less than the first threshold, and the peak and valley values of the root node height change are less than the second threshold, or if the variance of the lowest point height is less than the third threshold, and the peak and valley values of the lowest point height change are less than the fourth threshold, then it is determined to be a stationary sequence and sequence-level global calibration is required. If the peak or valley value of the change in root node height or the change in minimum node height exceeds the fifth threshold, it is determined to be a non-stationary sequence and frame-level real-time pull-back calibration is required.
4. The motion redirection height adaptive calibration method based on ground contact analysis according to claim 2, characterized in that: The forward kinematics can be used to obtain the first frame of each image in the action sequence. Position of a body part in the world coordinate system , in, Indicates the number of body parts considered. Indicates from the root node to the... The kinematic mapping of each body part, with the root node representing the base node or Pelvis node in the skeleton model, at the waist of the robot; The height of the lowest point of the entire body in this frame is: in, Indicates position The vertical coordinates.
5. The motion redirection height adaptive calibration method based on ground contact analysis according to claim 1, characterized in that: In step three, it is determined whether the root node of each frame in the calibrated dynamic video meets the penetration trigger condition. If it does, correction is performed, and then low-pass filtering smoothing is performed on the corrected root node. If the conditions are not met, low-pass filtering and smoothing are performed directly.
6. The motion redirection height adaptive calibration method based on ground contact analysis according to claim 5, characterized in that: For any frame of image, if the penetration trigger condition is met Or, the floating trigger condition is met. Then, the correction process is performed using the following formula: in, and Let represent the three-dimensional position column vectors of the robot's root node before and after correction in the world coordinate system, respectively. Indicates the floating tolerance threshold; After the correction is performed, a low-pass filter is applied for smoothing according to the following formula, and the correction amount for the unsmoothed root node vertical height is denoted as . The vertical height correction of the smoothed root node is Then we have: in, For smoothing coefficients, This represents the vertical height of the root node after low-pass filtering and smoothing.