Robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud

Abstract

The invention discloses a robot collision prediction and autonomous obstacle avoidance system based on a real-time point cloud, and the system comprises an omnidirectional sensing module which is used for collecting depth image data of a plurality of view fields, and obtaining an original three-dimensional point cloud; the dynamic obstacle segmentation module is used for eliminating static background point clouds based on the original three-dimensional point clouds and detecting obstacles based on dynamic obstacle point clouds; when an obstacle is detected, a three-dimensional bounding box is adopted to mark the obstacle, and the three-dimensional bounding box is optimized through gradient descent, so that the active search volume of the three-dimensional bounding box is minimum; the collision prediction module is used for generating a prediction track according to the obstacle marked by the three-dimensional bounding box and confirming a collision risk based on the prediction track; the autonomous obstacle avoidance module is used for triggering an avoidance action according to the collision risk; according to the invention, the determination of the three-dimensional target is further realized on the basis of target detection, the bounding box is further optimized, and the compactness of target identification and the accuracy of subsequent obstacle avoidance planning are improved.

Description

Technical Field

[0001] This invention relates to the field of robot navigation technology, and more specifically to a robot collision prediction and autonomous obstacle avoidance system based on real-time point clouds. Background Technology

[0002] With the widespread application of robots in industrial automation, logistics and warehousing, and other fields, their operating environments are becoming increasingly complex. Ensuring the safe and efficient operation of robots in dynamic or semi-dynamic environments, and avoiding collisions with people, equipment, and other obstacles, is one of the key challenges in robotics technology.

[0003] Existing robot obstacle avoidance technologies mainly rely on the following methods: Obstacle avoidance based on 2D LiDAR: This method is low in cost, but it can only acquire 2D planar information and cannot effectively detect suspended or low obstacles, resulting in a perception blind spot and easy collision.

[0004] Vision-based obstacle avoidance: This method can provide rich texture information, but it is greatly affected by factors such as changes in lighting, shadows, and reflections, resulting in insufficient stability. Furthermore, the calculation of depth information is complex, making it difficult to guarantee real-time performance.

[0005] Obstacle avoidance based on ultrasonic / infrared sensors: These sensors have short detection range, low accuracy, and narrow field of view, and are usually only used as an auxiliary obstacle avoidance method at close range.

[0006] Obstacle avoidance based on predefined paths: The robot strictly follows the preset path and simply stops when it encounters an obstacle. It lacks proactive and intelligent obstacle avoidance capabilities, which seriously affects work efficiency.

[0007] Therefore, how to provide omnidirectional, accurate, and real-time perception, and possess forward-looking collision prediction and smooth autonomous obstacle avoidance capabilities, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide a robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud that overcomes or at least partially solves the above problems.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A robot collision prediction and autonomous obstacle avoidance system based on real-time point clouds, comprising: The omnidirectional perception module is used to acquire depth image data from multiple fields of view to obtain the original 3D point cloud. The dynamic obstacle segmentation module is used to remove static background point clouds based on the original 3D point cloud and detect obstacles based on the dynamic obstacle point cloud. When detecting obstacles, a 3D bounding box is used to mark the obstacles, and the 3D bounding box is optimized by gradient descent to minimize its active search volume. The collision prediction module is used to generate local trajectories based on the obstacles marked by the three-dimensional bounding box, and to confirm the collision risk based on the local trajectories. The autonomous obstacle avoidance module is used to trigger avoidance actions based on the collision risk.

[0010] Preferably, the omnidirectional perception module includes a data fusion submodule and multiple 3D vision sensors. The multiple 3D vision sensors are arranged circumferentially in the robot's workspace to collect corresponding in-field depth image data and generate local point clouds. The fusion submodule is used to fuse the various local point clouds to obtain the original three-dimensional point cloud.

[0011] Preferably, the dynamic obstacle segmentation module includes a background culling submodule, a target detection submodule, and an optimized output submodule; The background removal submodule is used to remove static background point clouds by point cloud registration based on the original 3D point cloud and the prior map. The target detection submodule is used to identify dynamic obstacles based on dynamic point cloud data and output corresponding 2D detection boxes; and to project the 2D detection boxes into 3D space to obtain the target point cloud space. The optimized output submodule is used to generate a 3D bounding box based on the target point cloud space, and optimize it through gradient descent to output the optimal 3D bounding box.

[0012] Preferably, the optimized output submodule includes a coordinate system construction unit, a bounding box construction unit, and a bounding box optimization unit; The coordinate system calculation unit is used to determine the geometric direction based on each point in the target point cloud space through principal component analysis and to construct a local coordinate system. The bounding box construction unit is used to determine the local coordinates of the target point cloud space based on the local coordinate system. After obtaining the local bounding box, it is transformed to world coordinates based on the local coordinates. The bounding box optimization unit is used to calculate the volume based on the bounding box parameters in the world coordinate system and introduces a distance penalty term as the objective function; it is iteratively updated using the gradient descent method to obtain the optimal bounding box parameters.

[0013] Preferably, the distance penalty term is:

[0014] Where n is the total number of point clouds, , and Let i be the local coordinates of the i-th point. , and λ is the size of the bounding box, and λ is the regularization weight.

[0015] Preferably, the collision prediction module includes a motion modeling module, a trajectory generation module, and a risk quantification module; The motion modeling module is used to model the motion state of the robot and dynamic obstacles and confirm the motion state; The trajectory generation module is used to generate predefined trajectories; The risk quantification module is used to perform local sampling based on the predefined trajectory and calculate the collision risk according to the motion state corresponding to the sampling results.

[0016] Preferably, the risk quantification module includes a local sampling submodule, a risk quantification submodule, and a risk output submodule; The local sampling submodule is used to perform local sampling on the predefined trajectory, and to confirm the robot motion state and dynamic obstacle motion state corresponding to each path reference point within the sampling range, as local sampling information; The risk quantification submodule is used to quantify risk based on the local sampling information; The risk output submodule is used to output the risk quantification results to the autonomous obstacle avoidance submodule.

[0017] Preferably, the trajectory generation module includes a global trajectory generation submodule and a local path replanning submodule; The global trajectory generation submodule is used to generate predefined trajectories; The local path replanning submodule is used to monitor the obstacle avoidance signal of the autonomous obstacle avoidance module. When the obstacle avoidance signal is received, the local sampling information is used as input to minimize the motion error of the trajectory, the deviation from the predefined trajectory, and the obstacle distance cost, under the premise of satisfying the motion constraints and obstacle avoidance constraints, so as to obtain the optimal local trajectory.

[0018] Preferably, the autonomous obstacle avoidance module includes a strategy configuration submodule and a behavior decision submodule; The strategy configuration submodule configures corresponding decision-making behaviors for different collision risks; The behavior decision-making submodule is used to execute corresponding decision-making behaviors based on real-time collision risks.

[0019] Preferably, the decision-making behavior includes slowing down and following, yielding politely, or local replanning.

[0020] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud. It further realizes the determination of three-dimensional targets on the basis of target detection, and further optimizes the bounding box based on PCA, thereby improving the compactness of target recognition and the accuracy of subsequent obstacle avoidance planning. The present invention can consider dynamic obstacles to predict potential collision risks, thereby activating the obstacle avoidance strategy in advance before physical collision occurs. At the same time, it realizes the replanning of local paths based on dynamic obstacles during obstacle avoidance, resulting in diversified obstacle avoidance strategies. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of a robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of dynamic obstacle segmentation provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figure 1 and Figure 2 This invention discloses a robot collision prediction and autonomous obstacle avoidance system based on real-time point clouds, comprising: The omnidirectional perception module is used to acquire depth image data from multiple fields of view to obtain the original 3D point cloud. The dynamic obstacle segmentation module is used to remove static background point clouds based on the original 3D point cloud and detect obstacles based on the dynamic obstacle point cloud. When detecting obstacles, a 3D bounding box is used to mark the obstacles, and the 3D bounding box is optimized by gradient descent to minimize its active search volume. The collision prediction module is used to generate a predicted trajectory based on the obstacles marked by the three-dimensional bounding box, and to confirm the collision risk based on the predicted trajectory. The autonomous obstacle avoidance module is used to trigger avoidance actions based on the collision risk.

[0025] The following is a detailed explanation of each part of this embodiment: The omnidirectional perception module is used to detect the robot's working environment and generate raw point cloud data.

[0026] In one embodiment, the omnidirectional perception module includes a data fusion submodule and multiple 3D vision sensors. The multiple 3D vision sensors are circumferentially arranged in the robot's workspace to collect corresponding in-field depth image data and generate local point clouds. The fusion submodule is used to fuse the various local point clouds to obtain the original three-dimensional point cloud.

[0027] Specifically, four identical 3D vision sensors can be set up, which are fixedly arranged on the sensing module base in a circumferentially uniform manner with the central axis of the sensing module base as the center; specifically, the angle between the principal optical axes of two adjacent 3D vision sensors is 90 degrees, so that the field of view of the four 3D vision sensors are connected end to end in the horizontal direction, together forming a 360-degree horizontal panoramic field of view around the central axis. The vertical field of view of each of the 3D vision sensors is selected so that the vertical fields of view of the four sensors are superimposed in the vertical direction to form a complete three-dimensional coverage, thereby constructing a continuous and seamless three-dimensional point cloud data field around the robot and realizing omnidirectional perception without blind spots.

[0028] Furthermore, the 3D vision sensor can be selected from, but is not limited to, a binocular stereo vision camera, a structured light 3D camera, a time-of-flight 3D camera, or a lidar depth camera. Furthermore, the sensing module substrate is preferably a columnar or polyhedral structure, with mounting surfaces on its sides matching the number of 3D vision sensors.

[0029] In this embodiment, the fusion submodule is communicatively connected to the four 3D vision sensors to receive and synchronize the original depth images or point cloud data collected by each sensor. Through coordinate system transformation and point cloud stitching algorithm, the four local 3D point clouds are fused into a global 3D environment model with a consistent coordinate system, namely the original 3D point cloud.

[0030] The dynamic obstacle segmentation module is used to identify dynamic obstacles. In this embodiment, dynamic obstacles are given extra attention to improve the accuracy of robot navigation trajectory prediction.

[0031] In one embodiment, the dynamic obstacle segmentation module includes a background culling submodule, a target detection submodule, and an optimized output submodule.

[0032] The background removal submodule is used to remove static background point clouds based on the original 3D point cloud and the prior map through point cloud registration. Specifically, it uses a pre-built static high-precision map of a closed scene (prior map) to align the real-time point cloud with the prior map through point cloud registration algorithms (such as ICP and NDT) and remove point clouds belonging to the static background, thereby separating potential dynamic obstacle point cloud clusters.

[0033] The target detection submodule is used to identify dynamic obstacles based on dynamic point cloud data and output the corresponding 2D detection boxes; then, the 2D detection boxes are projected into 3D space to obtain the target point cloud space. Specifically, YOLOv12 is used for target detection to obtain the 2D detection box [x1,x2,y1,y2], where (x1,y1) is the coordinate of the upper left corner of the target and (x2,y2) is the coordinate of the lower right corner. Using the camera intrinsic parameter matrix K and the camera extrinsic parameter transformation matrix T, the 2D detection box is projected into 3D space to obtain the target's 3D point cloud space.

[0034] The optimized output submodule generates a 3D bounding box based on the target point cloud space and optimizes it using gradient descent to output the optimal 3D bounding box.

[0035] Furthermore, the optimized output submodule includes a coordinate system building unit, a bounding box building unit, and a bounding box optimization unit.

[0036] The coordinate system calculation unit is used to determine the geometric orientation of each point in the target point cloud space through principal component analysis and construct a local coordinate system. The bounding box construction unit is used to determine the local coordinates of the target point cloud space based on the local coordinate system, and after obtaining the local bounding box, it is transformed to world coordinates based on the local coordinates. The bounding box optimization unit is used to calculate the volume based on the bounding box parameters in the world coordinate system and introduce a distance penalty term as the objective function. The optimal bounding box parameters are obtained by iterative updating through gradient descent.

[0037] Specifically, the three-dimensional boundary is calculated based on the target's 3D point cloud space: First, the geometric orientation of the point cloud is calculated based on PCA. The spatial centroid is determined for each point in the 3D point cloud space, and the covariance is calculated based on the centroid. Then, eigenvalue decomposition is performed based on the covariance. j=1,2,3

[0038]

[0039] Where C is the covariance matrix, Let j be the feature obtained from the decomposition, and j be the feature variable. Here, is the eigenvalue, n is the number of points in the 3D point cloud space, and u is the centroid. Let be the coordinates of the i-th point. Thus, a local coordinate system is obtained with the centroid u as the origin and b1, b2, b3 as axes, which describes the geometric orientation of the 3D point cloud space.

[0040] Secondly, construct the initial 3D bounding box along the main direction.

[0041] Calculate the local coordinates q of each point in the local coordinate system. i : q i =(m i ,v i ,w i ) T =(b1,b2,b3) T (p i -u) Then find the extreme values ​​of the point cloud in the three principal directions: The maxima and minima in the V1 direction are respectively: m min = min(m i ), m max = max(m i ); The maxima and minima in the V2 direction are respectively: v min = min(v i ), v max = max(v i ); The maxima and minima in the V3 direction are respectively: w min = min(w i ), w max = max(w i ) The local coordinate system of the bounding box in 3D point cloud space is: Center: c local =( , , ) Dimension: s=(m) max -m min ,v max -v min ,w max -w min ) T Transform the center coordinates back to the world coordinate system. The center coordinates of the world coordinate system are C. world for: C world = u+[b1,b2,b3]·c local Finally, a gradient descent-based optimization algorithm is applied to minimize the bounding box volume while ensuring that it contains all point clouds. Let the optimized bounding box be centered at C in its world coordinate system. world The dimensions and rotation matrices are defined.

[0042] Where the dimension s = (Sx, Sy, Sz) T =(m max -m min ,v max -v min ,w max -w min ) T The rotation matrix R (defined by Euler angles θ=(θx,θy,θz)) is used to transform points in the global coordinate system to the local coordinate system. By optimizing the rotation matrix, the optimal rotation direction can be obtained so that the bounding box can both contain all points and minimize the volume V.

[0043] The goal is to minimize the bounding box volume while ensuring that all points are inside the box. This is a constrained optimization problem, which can be transformed into an unconstrained optimization problem:

[0044] Where L is the objective function, c, s, and θ are the center coordinates of the bounding box in the world coordinate system, the bounding box size parameters, and the Euler angles that determine the bounding box rotation matrix R, respectively; V(s) is the bounding box volume, and DistancePenalty is the distance penalty term for each point, used to ensure that the point is contained within the bounding box.

[0045] The distance penalty term can be defined as the sum of squares of the "hinge loss" beyond the boundary:

[0046] Here, ReLU(x) = max(0,x) indicates that the penalty function only applies when a point exceeds the boundary. λ is the regularization weight, used to balance volume minimization with the constraint of including all points (usually a large value, 10³). During gradient descent updates: the gradient is updated based on the objective function value, and the optimization parameters are iteratively updated in combination with the preset learning rate until the volume change is less than the threshold or the maximum number of iterations is reached. The optimized 3D bounding box is then output, including the center position, size, and orientation.

[0047] The orientation is represented by the final rotation matrix, or converted into the yaw angle ψ about the Z-axis (vertical direction): ψ=atan2(R 21 ,R 11 ) Where R 11 R21 It is the first column element of the rotation matrix.

[0048] The collision prediction module is used to plan a global model for the robot based on obstacle information. The collision prediction module includes a motion modeling module, a trajectory generation module, and a risk quantification module.

[0049] The motion modeling module is used to model the motion state of the robot and dynamic obstacles; the trajectory generation module is used to generate predefined trajectories; and the risk quantification module is used to perform local sampling based on the predefined trajectories and calculate the collision risk based on the motion state corresponding to the sampling results.

[0050] For example, the x and y velocities and angular velocities of the differential-driven robot are described by differential equations, and the maximum values ​​of its linear velocity, angular velocity, and corresponding acceleration are constrained to construct a robot motion model. The constant linear motion of the obstacle is described by multiplying the initial position acceleration by time, thus constructing a dynamic obstacle model.

[0051] The robot modeled with differential drive is as follows: ; ; ; in, and The rate of change of velocity in different directions, For robot speed, For the robot's heading angle, The rate of change of the robot's angular velocity. It is the robot's angular velocity command; The constraints of this model include: ; ;

[0052] in, , and The maximum speed, angular velocity, and rate of change of velocity supported by the robot.

[0053] The dynamic obstacle modeling for constant-velocity linear motion is as follows:

[0054]

[0055] in, Let J be the coordinates of obstacle J at time t. As the initial coordinates, For the speed of the obstacle, and This represents the velocity components in different directions.

[0056] In this embodiment, the trajectory generation module can predefine a trajectory under user operation as the global trajectory for robot operation. In order to quantify the collision risk, it is necessary to extract relevant trajectory portions from the predefined global trajectory.

[0057] Furthermore, the risk quantification module includes a local sampling submodule, a risk quantification submodule, and a risk output submodule.

[0058] The local sampling submodule is used to perform local sampling on the predefined trajectory, confirm the robot motion state and dynamic obstacle motion state corresponding to each path reference point within the sampling range, and use them as local sampling information; the risk quantification submodule is used to perform risk quantification based on the local sampling information; the risk output submodule is used to output the risk quantification results to the autonomous obstacle avoidance submodule.

[0059] To further implement the above technical solutions, risk quantification specifically includes: The distance between the robot and the obstacle at each moment is determined based on the locally sampled trajectory points, i.e., the first distance parameter; combined with the preset safety distance parameter, the probability of continuous collision at each moment is determined:

[0060] Where dj(tk) is the distance between the robot and obstacle j at time tk. For a safe distance, The relative velocity between the robot and the obstacle. Let be the unit direction vector between the obstacle and the robot. For the sigmoid function, The weighting is determined by the speed effect.

[0061] Furthermore, the total risk under the local trajectory is: P total = max(max(P coll,j (t k ))) In this embodiment, the total risk under the local trajectory is used as the reference risk for this obstacle avoidance action.

[0062] In one embodiment, the autonomous obstacle avoidance module can identify the risk quantification result. When the identified risk exceeds the preset condition, local path replanning is required.

[0063] To further implement the above technical solution, the trajectory generation module includes a global trajectory generation submodule and a local path replanning submodule; the global trajectory generation submodule is used to generate predefined trajectories.

[0064] The local path replanning submodule is used to listen to the obstacle avoidance signal of the autonomous obstacle avoidance module. When the obstacle avoidance signal is received, the local sampling information is used as input to minimize the motion error of the trajectory, the deviation from the predefined trajectory, and the obstacle distance cost under the premise of satisfying the motion constraints and obstacle avoidance constraints, so as to obtain the optimal local trajectory.

[0065] In one embodiment, the autonomous obstacle avoidance module includes a strategy configuration submodule and a behavior decision submodule; the strategy configuration submodule configures corresponding decision behaviors for different collision risks; wherein, the decision behaviors include deceleration and following, courteous yielding, or local replanning; the behavior decision submodule is used to execute corresponding decision behaviors according to the real-time collision risk.

[0066] Specifically, corresponding risk threshold ranges can be set for various decision-making behaviors to form a mapping relationship. When the actual collision risk falls within the threshold range, the corresponding mapping relationship is selected to execute the appropriate strategy.

[0067] To further implement the above technical solution, a control execution module is also included: connected to the autonomous obstacle avoidance module.

[0068] The discrete trajectory output by the autonomous obstacle avoidance module, i.e. the optimal local trajectory, is converted into low-level control commands; cubic spline interpolation is performed on the discrete trajectory to obtain a continuous function, and then Pure Pursuit tracking is performed. During the tracking process, the forward-looking distance is first set based on the robot's current moving speed (the faster the speed, the farther the forward-looking distance). Then, the target point corresponding to the forward-looking distance is found at the nearest point on the continuous path. The turning angle is calculated based on the relative relationship between the target point and the current position. Then, the moving speed command is determined according to the target speed, and the angular velocity command is calculated by combining the turning angle and the forward-looking distance. These two commands are sent to the underlying driver at a frequency of not less than 50Hz to control the robot's movement. At the same time, the data from the gyroscope and wheel speed odometry are integrated, and the robot's position and attitude are estimated in real time with the help of extended Kalman filtering. Once a deviation from the trajectory is detected, the motion command is corrected through a closed-loop mechanism to ensure that the robot accurately follows the planned path.

[0069] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0070] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A robot collision prediction and autonomous obstacle avoidance system based on real-time point clouds, characterized in that, include: The omnidirectional perception module is used to acquire depth image data from multiple fields of view to obtain the original 3D point cloud. The dynamic obstacle segmentation module is used to remove static background point clouds based on the original 3D point cloud and detect obstacles based on the dynamic obstacle point cloud. When detecting obstacles, a 3D bounding box is used to mark the obstacles, and the 3D bounding box is optimized by gradient descent to minimize its active search volume. The collision prediction module is used to generate local trajectories based on the obstacles marked by the three-dimensional bounding box, and to confirm the collision risk based on the local trajectories. The autonomous obstacle avoidance module is used to trigger avoidance actions based on the collision risk.

2. The robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud as described in claim 1, characterized in that, The omnidirectional perception module includes a data fusion submodule and multiple 3D vision sensors. The multiple 3D vision sensors are circumferentially arranged in the robot's workspace to collect corresponding in-field depth image data and generate local point clouds. The fusion submodule is used to fuse the various local point clouds to obtain the original three-dimensional point cloud.

3. The robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud as described in claim 1, characterized in that, The dynamic obstacle segmentation module includes a background removal submodule, a target detection submodule, and an optimized output submodule; The background removal submodule is used to remove static background point clouds by point cloud registration based on the original 3D point cloud and the prior map. The target detection submodule is used to identify dynamic obstacles based on dynamic point cloud data and output corresponding 2D detection boxes; and to project the 2D detection boxes into 3D space to obtain the target point cloud space. The optimized output submodule is used to generate a 3D bounding box based on the target point cloud space, and optimize it through gradient descent to output the optimal 3D bounding box.

4. The robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud as described in claim 3, characterized in that, The optimized output submodule includes a coordinate system construction unit, a bounding box construction unit, and a bounding box optimization unit; The coordinate system calculation unit is used to determine the geometric direction based on each point in the target point cloud space through principal component analysis and to construct a local coordinate system. The bounding box construction unit is used to determine the local coordinates of the target point cloud space based on the local coordinate system. After obtaining the local bounding box, it is transformed to world coordinates based on the local coordinates. The bounding box optimization unit is used to calculate the volume based on the bounding box parameters in the world coordinate system and introduces a distance penalty term as the objective function; it is iteratively updated using the gradient descent method to obtain the optimal bounding box parameters.

5. A robot collision prediction and autonomous obstacle avoidance system based on real-time point clouds according to claim 4, characterized in that, The distance penalty term is: Where n is the total number of point clouds, , and Let i be the local coordinates of the i-th point. , and λ is the size of the bounding box, and λ is the regularization weight.

6. The robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud as described in claim 1, characterized in that, The collision prediction module includes a motion modeling module, a trajectory generation module, and a risk quantification module. The motion modeling module is used to model the motion state of the robot and dynamic obstacles and confirm the motion state; The trajectory generation module is used to generate predefined trajectories; The risk quantification module is used to perform local sampling based on the predefined trajectory and calculate the collision risk according to the motion state corresponding to the sampling results.

7. A robot collision prediction and autonomous obstacle avoidance system based on real-time point clouds according to claim 6, characterized in that, The risk quantification module includes a local sampling submodule, a risk quantification submodule, and a risk output submodule; The local sampling submodule is used to perform local sampling on the predefined trajectory, and to confirm the robot motion state and dynamic obstacle motion state corresponding to each path reference point within the sampling range, as local sampling information; The risk quantification submodule is used to quantify risk based on the local sampling information; The risk output submodule is used to output the risk quantification results to the autonomous obstacle avoidance submodule.

8. The robot collision prediction and autonomous obstacle avoidance system based on real-time point cloud as described in claim 7, characterized in that, The trajectory generation module includes a global trajectory generation submodule and a local path replanning submodule; The global trajectory generation submodule is used to generate predefined trajectories; The local path replanning submodule is used to monitor the obstacle avoidance signal of the autonomous obstacle avoidance module. When the obstacle avoidance signal is received, the local sampling information is used as input to minimize the motion error of the trajectory, the deviation from the predefined trajectory, and the obstacle distance cost, under the premise of satisfying the motion constraints and obstacle avoidance constraints, so as to obtain the optimal local trajectory.

9. A robot collision prediction and autonomous obstacle avoidance system based on real-time point clouds according to claim 1, characterized in that, The autonomous obstacle avoidance module includes a strategy configuration submodule and a behavior decision submodule; The strategy configuration submodule configures corresponding decision-making behaviors for different collision risks; The behavior decision-making submodule is used to execute corresponding decision-making behaviors based on real-time collision risks.

10. A robot collision prediction and autonomous obstacle avoidance system based on real-time point clouds according to claim 9, characterized in that, The decision-making behaviors include slowing down and following, yielding politely, or local replanning.

Images