Laser beam-based agv environment perception method and system

Abstract

The application belongs to the technical field of AGV, and provides an AGV environment perception method and system based on a laser beam. The method comprises: continuously scanning the surrounding environment by a vehicle-mounted laser radar to obtain a time-sequentially continuous laser point cloud data frame set; obtaining a plurality of dynamic targets based on the laser point cloud data frame set, extracting dynamic characteristic parameters of each dynamic target, and identifying a light dynamic target from each dynamic target; performing motion center analysis on the corresponding light dynamic target based on the motion trajectory feature to obtain a motion center analysis result; determining a plurality of obstacle targets based on the motion center analysis result, planning a driving trajectory of each obstacle target in the surrounding environment, and executing the driving trajectory. The application can reduce the false stopping rate of AGV caused by flying lint interference, effectively avoid the safety risk of pulling between AGV and falling yarn, improve the perception reliability and operation intelligent level in a complex textile environment, and further improve the accuracy of trajectory planning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of AGV, in particular to an AGV environment perception method and system based on laser beams. BACKGROUND

[0002] The automated guided vehicle (AGV) undertakes the important task of automatically transporting spindles between different stations such as spinning machines, bobbin winder machines, and warehouses, which is of great significance to improving production efficiency and reducing labor costs.

[0003] The autonomous navigation and obstacle avoidance capability of AGV is highly dependent on its environment perception system. At present, the navigation and obstacle avoidance technology based on Lidar has become the mainstream perception scheme of AGV due to its high ranging accuracy, wide scanning range, and no need to modify the ground. This technology emits laser beams to the surrounding environment and receives the reflected signals to generate point cloud data of the surrounding environment, and then realizes AGV's self-positioning, map construction (SLAM), obstacle detection through algorithms, and finally realizes path planning.

[0004] However, there are a large number of cotton and chemical fiber flying fibers in the air of the spinning workshop. These flying fibers will reflect the laser beams emitted by AGV during falling, forming a large number of discrete and instantaneous noise points in the point cloud data. The existing conventional obstacle recognition algorithm cannot effectively distinguish these flying fiber noise points from real and stable obstacles (such as personnel, equipment, walls, etc.), often misjudging flying fibers as obstacles. This leads AGV to frequently perform unnecessary emergency braking or detour actions, not only seriously affecting the efficiency and continuity of material transfer, breaking the production rhythm, but also exacerbating the wear and tear of AGV mechanical parts and the unnecessary consumption of energy.

[0005] Therefore, how to accurately filter the interference of flying fibers in the environment and accurately identify real obstacles to improve the accuracy and rationality of AGV path planning is a technical problem that needs to be solved at present. SUMMARY

[0006] To solve the above technical problems, the present application provides an AGV environment perception method and system based on laser beams.

[0007] The application provides an AGV environment perception method based on a laser beam, comprising the following steps: an AGV in a spinning workshop controls a vehicle-mounted laser radar to continuously scan the surrounding environment to obtain a time-sequentially continuous laser point cloud data frame set; a plurality of dynamic targets are derived based on the laser point cloud data frame set, and dynamic characteristic parameters of each dynamic target are extracted, at least including a dynamic target existing duration and a motion trajectory feature, and a lightweight dynamic target is identified from each dynamic target based on the dynamic characteristic parameters; motion center analysis is performed on the corresponding lightweight dynamic target based on the motion trajectory feature to obtain a motion center analysis result, including whether the lightweight dynamic target is a motion state based on a constraint center and the number of constraint centers; a plurality of obstacle targets are determined based on the motion center analysis result, a driving trajectory in the surrounding environment is planned based on each obstacle target, and execution is performed.

[0008] The application also provides an AGV environment perception system based on a laser beam, comprising: a vehicle-mounted laser radar configured to continuously scan the surrounding environment of an AGV in a spinning workshop to obtain a time-sequentially continuous laser point cloud data frame set; a target preliminary identification unit configured to derive a plurality of dynamic targets based on the laser point cloud data frame set, and extract dynamic characteristic parameters of each dynamic target, at least including a dynamic target existing duration and a motion trajectory feature, and identify a lightweight dynamic target from each dynamic target based on the dynamic characteristic parameters; a target distinguishing unit configured to perform motion center analysis on the corresponding lightweight dynamic target based on the motion trajectory feature to obtain a motion center analysis result, including whether the lightweight dynamic target is a motion state based on a constraint center and the number of constraint centers; and a trajectory planning unit configured to determine a plurality of obstacle targets based on the motion center analysis result, plan a driving trajectory in the surrounding environment based on each obstacle target, and perform execution.

[0009] The application can effectively distinguish lightweight flying fluff and falling yarn in the spinning workshop by performing time sequence analysis on the laser point cloud data and extracting the motion features of the dynamic targets, and can accurately determine the physical motion constraint state of the target by introducing motion center analysis, filter out the unconstrained free flying fluff, and identify single / double constraint yarns that may constitute entanglement risk as real obstacles. The application can significantly reduce the false stop rate of the AGV caused by flying fluff interference, ensure the continuity of material transfer and production efficiency, effectively avoid the safety risk of pulling between the AGV and the falling yarn, improve the perception reliability and operation intelligent level in a complex textile environment, and further improve the accuracy of trajectory planning. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 is a flowchart of an AGV environment perception method based on a laser beam disclosed by an embodiment of the application.

[0011] Figure 2 is a structural schematic view of an AGV environment perception system based on a laser beam disclosed by an embodiment of the present application.

[0012] Figure 3 is another structural schematic view of an AGV environment perception system based on a laser beam disclosed by an embodiment of the present application. DETAILED DESCRIPTION

[0013] The technical solutions of the present application will be further described in detail below with the aid of drawings and embodiments.

[0014] To make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with the aid of drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present application.

[0015] To facilitate the understanding of the embodiments of the present application, further explanation and description will be made below with specific embodiments with the aid of drawings. The embodiments do not constitute a limitation on the embodiments of the present application.

[0016] As shown in Figure 1 , the embodiments of the present application disclose an AGV environment perception method based on a laser beam, comprising the following steps: step 100, the AGV in a spinning mill controls the vehicle-mounted laser radar to continuously scan the surrounding environment, and obtains a time-sequentially continuous laser point cloud data frame set.

[0017] In this step, in the spinning mill, the AGV controls its vehicle-mounted laser radar sensor to periodically scan the surrounding environment at a fixed scanning frequency (for example, 10 Hz). The environment distance and profile information obtained by each scanning operation constitute a frame of laser point cloud data. The multiple frames of point cloud data collected in time sequence are collected, which constitutes the above-mentioned time-sequentially continuous laser point cloud data frame set.

[0018] Step 200, based on the laser point cloud data frame set, a plurality of dynamic targets are derived, and dynamic characteristic parameters of each dynamic target are extracted, at least including the existence duration and motion trajectory characteristics of the dynamic target, and a light dynamic target is identified from each dynamic target based on the dynamic characteristic parameters.

[0019] In this step, each frame of laser point cloud data is pre-processed (e.g. filtering and noise reduction, etc.); and a Euclidean clustering algorithm is used to aggregate adjacent points in space to form several point cloud clusters of targets. Subsequently, a multi-target tracking algorithm based on Kalman filtering is used to associate and match these point cloud clusters between consecutive frames, thereby generating a continuous motion trajectory for each target, i.e. obtaining several dynamic targets.

[0020] At the same time, the dynamic characteristic parameters of each dynamically tracked dynamic target are calculated, including at least: (1) existence duration: the total length of time that the dynamic target has been stably tracked in consecutive frames from the first detection to the current time; (2) motion trajectory characteristics: parameters describing the target motion mode, such as instantaneous speed, acceleration, rate of change of motion direction, and trajectory curvature, etc.

[0021] Based on the above dynamic characteristic parameters, a screening rule is set. For example, if a dynamic target has a short existence duration (e.g. less than 0.5 seconds) and its motion trajectory exhibits high randomness and discontinuity (e.g. irregular floating), it is determined that the target is a light-weight dynamic target that is easily affected by air flow, mainly including flying fluff and accidentally falling yarn in the air.

[0022] In step 300, motion center analysis is performed on the corresponding light-weight dynamic target based on the motion trajectory characteristics, and a motion center analysis result is obtained, including whether the light-weight dynamic target is in a motion state based on a constraint center, and the number of constraint centers.

[0023] In this step, a more in-depth physical motion nature analysis is performed on the above light-weight dynamic targets that have been preliminarily screened out, in order to distinguish between harmless flying fluff and yarn that may constitute an obstacle.

[0024] Firstly, motion center analysis is performed, i.e. to determine whether the motion of a light-weight dynamic target is constrained by one or more fixed points. Flying fluff that is freely floating has random motion without a fixed constraint center. A piece of yarn that accidentally falls from the equipment usually has at least one end still attached to the spinning machine or bobbin, and its motion is actually a swing or pull around this or these fixed points.

[0025] By analyzing the historical motion trajectory characteristics (such as periodicity of the trajectory, center of rotation) of the light-weight dynamic target, a motion center analysis result can be obtained: (1) motion state judgment: to determine whether the motion is in a motion state based on a constraint center or in a free motion state; (2) constraint center number: to further analyze the number of constraint centers, mainly divided into single constraint center (such as one end of the yarn being fixed) and double constraint center (such as both ends of the yarn being accidentally caught, forming a taut line segment).

[0026] At step 400, a plurality of obstacle targets are determined based on the motion center analysis result, a driving trajectory in the surrounding environment is planned based on each obstacle target, and the driving trajectory is executed.

[0027] In this step, based on the analysis result of step 300, the following decision rule is executed, for example: the light dynamic target determined to be in a free motion state, i.e. flying dust, is filtered out and not regarded as an obstacle because it will not cause substantial obstruction to the AGV. The light dynamic target determined to be in a motion state based on the constraint center, i.e. yarn, especially the target identified as a double constraint center, means that a section of yarn may cross the path of the AGV, and there is a risk of pulling the equipment or itself when the AGV drives through, so it is determined as an obstacle target.

[0028] Further, the AGV takes into account all the finally determined real obstacle targets (including traditional obstacles and new obstacles composed of yarn) to plan a safe and efficient driving trajectory. The AGV then executes the planned trajectory, thereby effectively avoiding the potential risks brought by the yarn while avoiding misjudging the flying dust, ensuring the continuity and safety of the transfer process. The avoidance value corresponding to the single constraint center and the double constraint center obstacle target can be configured, and the avoidance value is used to determine the distance between the driving trajectory and the obstacle target. Obviously, the distance between the driving trajectory and the single constraint center obstacle target is less than the distance between the driving trajectory and the double constraint center obstacle target (because the double constraint center obstacle target is more dangerous to the AGV).

[0029] It can be understood that the planning of the driving trajectory can be the responsibility of the processor of the AGV itself or the server side, and the server side can be a network server or an edge server, and no specific limitation is made in this regard. When the server side is responsible, the AGV transmits the dynamic obstacle target obtained above and other static obstacle targets to the server side, the server side establishes a map of the surrounding environment where the AGV is located, and then plans the driving trajectory by using a suitable path planning algorithm.

[0030] The present application can effectively distinguish between light flying dust and falling yarn in a spinning workshop by performing time series analysis on laser point cloud data and extracting the motion characteristics of dynamic targets. In addition, by introducing motion center analysis, the physical motion constraint state of the target can be accurately determined, the free flying dust without constraint is filtered out, and the single / double constraint yarn that may cause entanglement risk is identified as a real obstacle. The present application can significantly reduce the false stop rate of the AGV caused by flying dust interference, ensure the continuity and production efficiency of material transfer, effectively avoid the safety risk of pulling the AGV and the falling yarn, improve the perception reliability and operation intelligence level in complex textile environments, and further improve the accuracy of trajectory planning.

[0031] As an example, after obtaining a set of laser point cloud data frames in time sequence, a step of data preprocessing is further included, specifically comprising: step 101, based on the structural prior information of fixed equipment in the spinning workshop, static background filtering is performed on the point cloud data.

[0032] In this step, the internal layout of the spinning workshop is relatively fixed, which contains a large number of fixed equipment with known positions and forms such as spinning machines, bobbin winder, support columns, and walls. These fixed equipment will form stable background point cloud when scanned by laser radar. If not handled, it will continue to be misjudged as an obstacle, which will seriously increase the subsequent calculation burden and affect the perception accuracy.

[0033] The structural prior information refers to the digital map containing the precise three-dimensional coordinates and geometric profiles of all fixed equipment, which is obtained by pre-laser SLAM mapping or importing workshop design drawings (CAD model). In actual operation, the AGV obtains its pose in the current global coordinate system through its positioning system, and then registers each frame of real-time collected laser point cloud data with the high-precision prior map. Through spatial coordinate transformation and comparison, point cloud data points falling within the range of known fixed equipment models can be accurately identified. The above identified point cloud data belonging to static background will be removed from the current frame. This operation can significantly purify the point cloud data, so that the subsequent processing flow can focus on the real dynamic or unknown objects, and the real-time performance and accuracy of environmental perception can be improved.

[0034] Step 102, for the point cloud data after filtering the static background, a statistical feature-based outlier filtering algorithm is used to remove discrete noise points caused by small-diameter suspended fly ash in the air that do not meet the spatial distribution continuity.

[0035] In this step, after static background filtering, the point cloud data still contains two main components: one is real dynamic targets (such as personnel, other AGVs, and falling yarn), and the other is instantaneous, discrete noise formed by the reflection of laser light from lightweight flying fluff such as cotton and chemical fibers diffused in the workshop. Refined filtering is needed for the latter, achieved using a statistical outlier filtering algorithm (e.g., a statistical outlier removal algorithm). The principle is as follows: the algorithm statistically analyzes the local spatial distribution characteristics of each point in the point cloud. Specifically, for any point P in the point cloud, its k nearest neighbors are searched, and the average distance from point P to these k neighbors is calculated. Subsequently, the average distance between all points in the entire point cloud or a local area and their neighbors is calculated, along with the mean (μ) and standard deviation (σ) of these distances. According to the characteristics of a normal (Gaussian) distribution, the average distance of the vast majority of points should fall within the interval [μ-ασ, μ+ασ] (where α is a scale parameter, typically taken as 1.0 to 3.0). Any point whose average distance is far beyond the upper limit of the interval (i.e., greater than μ+ασ) is considered an outlier that is statistically incompatible with the surrounding point cloud.

[0036] In a spinning workshop environment, small-diameter suspended fluff particles are small in size and sparsely distributed. The reflection points generated by the laser beam on them are often isolated in space, and their average distance to neighboring points is much greater than the typical distance between points within a continuous solid surface point cloud (such as a person's leg or the surface of a feed bin). Therefore, these fluff reflection points are efficiently and accurately identified as statistical outliers by the aforementioned algorithm and filtered out. Understandably, long chains or clumps of fluff will still be identified as lightweight dynamic targets and require further differentiation and identification in step 300.

[0037] This embodiment can effectively remove a large amount of invalid information caused by environmental interference, greatly reducing the probability of misjudging small-path flying catkins as obstacles in subsequent steps, and improving the reliability of dynamic target recognition and classification.

[0038] As an example, the method of identifying lightweight dynamic targets from each dynamic target based on dynamic feature parameters includes: step 201, calculating the duration of existence and the randomness index of the motion trajectory of each dynamic target; wherein, the randomness index of the motion trajectory is quantified by calculating the standard deviation of the change in the motion direction of the dynamic target in consecutive frames.

[0039] In this step, in order to realize accurate and objective identification of the light dynamic target and avoid subjective judgment, the following two quantifiable key dynamic characteristic parameters are used as criteria: existence duration: which represents the stability of the dynamic target in the perception field of view. For a tracked dynamic target, from the time when it is first clustered and identified in the point cloud data and successfully associated to the current frame, the total time length of being continuously tracked in the continuous frames is the existence duration of the target. It can be understood that the existence duration can be directly calculated by the frame rate and the number of continuous tracking frames.

[0040] Randomness index of motion trajectory: which is used to accurately quantify the degree of disorder of the motion trajectory of the dynamic target. The specific calculation method is as follows: in the continuous frame sequence of the tracked dynamic target, the motion direction angle is calculated frame by frame, for example, the angle value in the polar coordinate system with the AGV as the origin. The standard deviation of the change amount of the motion direction angle between adjacent two frames, that is, the difference between the direction angles of adjacent two frames, is calculated, and the standard deviation is defined as the randomness index. It can be understood that the flying dust which freely drifts has frequent and irregular changes in the motion direction, resulting in a very large standard deviation of the direction change amount; and the motion direction of the intentional moving body such as personnel and vehicle is usually continuous and smooth, and the standard deviation of the direction change amount is relatively small.

[0041] In step 202, if the existence duration of the dynamic target is less than the first preset threshold and the randomness index of the motion trajectory is greater than the second preset threshold, the dynamic target is determined to be a light dynamic target.

[0042] In this step, only when a dynamic target meets the following two conditions at the same time, it is determined to be a light dynamic target: (1) existence duration < first preset threshold: this condition is used to filter transient and short-lived targets. Light objects such as flying dust may quickly drift through the laser scanning plane, and the duration of being stably tracked is short. The first preset threshold can be determined by statistics or experiments according to the actual scene, for example, set to 0.3-0.8 seconds. It is worth noting that a yarn that has just left the device at the initial falling stage may also exhibit a short and unstable motion state, thereby meeting this condition.

[0043] (2) randomness index > second preset threshold: this condition is used to filter targets with irregular motion patterns. As described above, high randomness index is a typical feature of objects without active motion ability such as flying dust. The second preset threshold is also set based on scene measurement data to ensure that regular and irregular motion can be effectively distinguished. Similarly, a yarn that freely falls, is fixed at one end or even at both ends, may also exhibit a high randomness in its initial motion stage.

[0044] The dynamic target satisfying the above conditions simultaneously has behavior characteristics consistent with the physical characteristics of light-weight objects (including flying dust and yarn in a specific motion state) (short appearance time, chaotic motion trajectory), and is therefore preliminarily determined to be a light-weight dynamic target.

[0045] Based on the quantified existence duration and randomness index, this embodiment can efficiently and accurately preliminarily screen all targets with the floating feature (including flying dust and yarn) from a complex dynamic environment, to ensure the quality and pertinence of the input data in the subsequent motion center analysis stage.

[0046] As an example, the motion center analysis on the corresponding light-weight dynamic target based on the motion trajectory feature includes: step 301, for each identified light-weight dynamic target, extracting a historical motion trajectory point set within a time window, and performing circular or elliptical fitting on the historical motion trajectory point set.

[0047] In this step, in order to accurately analyze the motion mode of the light-weight dynamic target, the overall motion trend of the light-weight dynamic target within a period of time needs to be investigated. Specifically, for each light-weight dynamic target identified in step 200, all continuous position points within a preset time window length (for example, the last 1-2 seconds) are extracted from the tracked trajectory sequence of the light-weight dynamic target to form a historical motion trajectory point set of the light-weight dynamic target. It can be understood that the length of the above-mentioned time window needs to be reasonably set to ensure that enough trajectory points are included to reflect the motion mode, while not introducing too much noise or mode change due to too long time.

[0048] In the spinning workshop scene, the motion mode of a segment of constrained yarn is usually geometrically represented as swinging (approximating a circular arc or a sector) around a fixed point (constrained center) or vibrating between two fixed points (approximating a line segment, which may be elliptical when the motion trajectory is projected on a plane). Therefore, the circular fitting algorithm (such as the least squares circle fitting) and the elliptical fitting algorithm are used to perform curve fitting on the historical motion trajectory point set in this embodiment. It can be understood that the fitting process aims to find an optimal circle or ellipse, so that the sum of the geometric boundary distances of all points in the trajectory point set to the circle or ellipse is minimized.

[0049] Step 302, based on the fitting goodness, it is determined whether the motion of the light-weight dynamic target is around a stable geometric center, specifically: if the fitting goodness is higher than a preset threshold, it is determined that the light-weight dynamic target is in a motion state based on a constrained center, and the fitted center of the circle or ellipse is determined as the constrained center, and then the number of constrained centers is determined according to the fitted shape.

[0050] The goodness of fit in this step is a statistical measure for the matching degree between the historical motion trajectory point set and the fitted circle or ellipse, for example, the coefficient of determination R2 or other indicators can be used. If the goodness of fit is higher than the preset threshold (which can be calibrated by experiment, for example, set to 0.7), it indicates that the actual motion trajectory of the lightweight dynamic target is highly consistent with the theoretical motion model constrained by the geometric center.

[0051] When the goodness of fit is higher than the preset threshold, it is determined that the lightweight dynamic target is in a motion state based on the constraint center. At this time, the coordinates of the center of the fitted circle or the center of the ellipse are determined as the spatial position of the geometric center (i.e. the constraint center) surrounded by the motion constraint. Further, according to the fitted model used, the number of constraint centers can be inferred. Specifically, if the trajectory point set is highly fitted with a circular model, it usually indicates that the target mainly moves around a fixed constraint center point (single constraint center), for example, the swing of a single-end fixed yarn; if the trajectory point set is highly fitted with an elliptical model, it indicates that it is affected by two main constraint points (double constraint center), for example, the vibration of a tight state formed by a section of yarn hanging at both ends. The final motion center analysis result includes the determination of whether the lightweight dynamic target is in a constrained state, as well as the position and number of constraint centers.

[0052] This embodiment converts the abstract physical motion state judgment into precise mathematical model evaluation by fitting the motion trajectory with a geometric model, thereby realizing effective differentiation between free-floating flying dust and constrained motion yarn. At the same time, through the goodness of fit threshold judgment, the objectivity and accuracy of the state judgment are ensured. The number of constraint centers inferred according to the fitted shape can be used for subsequent accurate identification of high-risk tight yarn (double constraint center).

[0053] As an example, the circular or elliptical fitting of the historical motion trajectory point set includes: step 3011, after completing the preliminary fitting of the circle or ellipse, extracting the multi-dimensional features in the fitting process, including at least the distribution density of the trajectory points, the distribution uniformity of the fitting residual, and the projection point distribution continuity of the trajectory point set on the fitted curve.

[0054] In this step, in the prior art, when performing motion center analysis on the motion trajectory, only the goodness of fit (such as R 2The value) as the only basis for judgment. However, in the actual complex environment of the spinning workshop, this single index judgment method has obvious defects: for example, a short, sparse and noise-containing free flying flock trajectory may have a higher preliminary fitting goodness of fit with a certain geometric model due to chance, resulting in a false judgment as a constrained motion; on the contrary, a real constrained yarn trajectory may be underestimated due to uneven distribution of sampling points or a small number of outliers, resulting in a missed detection.

[0055] To solve the above problems, the embodiment further extracts multi-dimensional features that can more deeply reflect the fitting quality after completing the preliminary fitting, at least including: the distribution density of the trajectory points: used to evaluate the uniformity of the coverage of the trajectory points on the entire fitting curve, avoiding false judgment caused by local fitting. The calculation method is: (1) arc segment division: the complete circular or elliptical curve fitted out is equally divided into N arc segments (for example, N = 36, i.e. every 10 degrees an arc segment); (2) projection and counting: each point in the set of historical motion trajectory points is projected vertically onto the nearest fitting curve, and the arc segment number to which the projection point belongs is determined; (3) density statistics: the number of projection points contained in each arc segment is counted; (4) uniformity calculation: the standard deviation or coefficient of variation (standard deviation / average value) of the number of projection points in the N arc segments is calculated. The standard deviation or coefficient of variation is the quantitative value of the distribution density feature. The smaller the value, the more uniform the distribution; the larger the value, the more concentrated the distribution, and there is a risk of local fitting.

[0056] The distribution uniformity of the fitting residual: used to detect whether there is a systematic deviation, so as to judge whether the fitting model itself is applicable. The calculation method is: (1) obtain the residual: record the vertical distance of each trajectory point to the fitting curve as the fitting residual of the point, and record the azimuth angle of each trajectory point relative to the center of the fitted circle or ellipse; (2) azimuth partition: divide the 0-360 degree azimuth space into M sectors (for example, M = 8, i.e. every 45 degrees a sector); (3) sector residual statistics: calculate the average value of the fitting residuals of all trajectory points in each azimuth sector; (4) uniformity calculation: calculate the standard deviation or range (maximum value minus minimum value) of the average values of the M sector residuals. The standard deviation or range is the quantitative value of the residual distribution uniformity feature. The smaller the value, the more uniform the distribution of the residual in different directions, and there is no obvious systematic deviation; the larger the value, the residual in some directions is continuously large or small, there is a systematic deviation, and the model may not be applicable.

[0057] The continuity of the distribution of the projection points of the trajectory point set on the fitted curve: used to evaluate the coherence of the motion in time and identify unreliable trajectories caused by tracking jitter or mismatch. The calculation method is as follows: (1) projection in time sequence: according to the time sequence in which the trajectory points are collected, each point is projected onto the fitted curve to obtain a series of projection points arranged in time sequence; (2) calculation of the arc length distance between adjacent projection points: on the fitted curve, the shortest arc length distance between the i th projection point and the i+1 th projection point along the curve is calculated; (3) continuity evaluation: the mean μ and the standard deviation σ of the arc length distances between all consecutive projection point pairs are counted, and the proportion of the arc length distances that exceed the range (μ+k*σ) (k is a constant, usually 2 or 3) or the coefficient of variation of the arc length distances is calculated. The proportion or the coefficient of variation that exceeds the range is the quantitative value of the continuity of the distribution of the projection points. The smaller the value, the more coherent and stable the motion is; the larger the value, the more abnormal jumps or reversals of the trajectory on the curve, the poorer the coherence and the lower the reliability.

[0058] In step 3012, the goodness of fit obtained by preliminary fitting and the multi-dimensional features are jointly input into a pre-trained prediction model, the input features are comprehensively analyzed by the prediction model, and a corrected goodness of fit is output.

[0059] In this step, the prediction model is pre-constructed and trained in the embodiment. The model is trained by a large number of labeled trajectory samples (whose real motion state is known), and can learn the complex and nonlinear combination relationship between the preliminary goodness of fit and each multi-dimensional feature. In actual application, the preliminary goodness of fit and the extracted multi-dimensional features are jointly input into the prediction model for comprehensive analysis. For example, an unreliable fitting with high preliminary goodness of fit but uneven residual distribution is identified, and a lower corrected goodness of fit is output; a reliable fitting with medium preliminary goodness of fit but good features is also identified, and a higher corrected goodness of fit is output.

[0060] Finally, the corrected goodness of fit output by the prediction model integrates more dimensional quality information, and is more true than the single preliminary goodness of fit to reflect the degree of fit between the trajectory and the constraint motion model, which can significantly improve the accuracy and robustness of the motion center state determination and effectively reduce the misjudgment rate in complex environments.

[0061] It can be understood that the prediction model can be constructed by using a suitable existing algorithm, for example, a gradient boosting decision tree model such as XGBoost or LightGBM, which is composed of multiple decision trees. In the training stage, the prediction model generates new decision trees iteratively, and each tree is committed to correcting the residual error predicted by the previous tree. Finally, the output of the prediction model is the weighted sum of the prediction results of multiple trees. In the present scheme, the training target of the prediction model is to make the weighted sum as close as possible to the true correction goodness of fit (which can be labeled by an expert according to the true physical state of the trajectory).

[0062] The following is described by taking LightGBM as an example: input layer: input feature vector: [preliminary goodness of fit (R²), distribution density feature value (such as coefficient of variation), residual error distribution uniformity feature value (such as standard deviation), projection point continuity feature value (such as coefficient of variation)].

[0063] Model core parameters: n_estimators (number of trees): 100. max_depth (maximum depth of a single tree): 6 (used to limit tree depth to prevent overfitting and ensure model generalization ability). learning_rate (learning rate): 0.1. subsample (data subsampling ratio used when training each tree): 0.8 (enhances robustness).

[0064] Output layer: output a continuous numerical value between 0 and 1, i.e. the final goodness of fit after intelligent correction by the model.

[0065] The specific processing flow is as follows: the model outputs a corrected goodness of fit value through the comprehensive judgment of its 100 decision trees. For example, even if the preliminary R² is 0.75, but if the distribution density is uneven and the residual error has systematic bias, the model outputs a correction value of only 0.4; on the contrary, if the preliminary R² is 0.7 and all auxiliary features perform well, the model outputs a correction value as high as 0.85.

[0066] As shown in Figure 2 The embodiment of the present application also provides an AGV environment perception system 200 based on a laser beam, which comprises: a vehicle-mounted laser radar 201 configured to continuously scan the surrounding environment of the AGV in the spinning mill and obtain a set of time-sequentially continuous laser point cloud data frames.

[0067] A target preliminary identification unit 202 is configured to derive a plurality of dynamic targets based on the set of laser point cloud data frames, extract dynamic feature parameters of each dynamic target, at least including existence duration and motion trajectory features of the dynamic target, and identify a light dynamic target from each dynamic target based on the dynamic feature parameters.

[0068] The target differentiation unit 203 is configured to perform motion center analysis on the corresponding lightweight dynamic target based on the motion trajectory features, and obtain the motion center analysis results, including whether the lightweight dynamic target is in a motion state based on the constraint center, and the number of constraint centers.

[0069] The trajectory planning unit 204 is configured to: determine a number of obstacle targets based on the motion center analysis results, plan a driving trajectory in the surrounding environment based on each obstacle target, and execute the plan.

[0070] As an example, such as Figure 3 As shown, the system also includes a preprocessing unit 205, which is configured to: perform static background filtering on point cloud data based on prior structural information of fixed equipment in the spinning workshop; and use a statistical feature-based outlier filtering algorithm on the point cloud data after static background filtering to remove discrete noise points that do not meet the spatial distribution continuity caused by small-diameter suspended fluff in the air.

[0071] As an example, the target identification unit 202 is configured to: calculate the duration of existence and the randomness index of the motion trajectory of each dynamic target; wherein the randomness index of the motion trajectory is quantified by calculating the standard deviation of the change in the motion direction of the dynamic target in consecutive frames; if the duration of existence of the dynamic target is less than a first preset threshold and the randomness index of its motion trajectory is greater than a second preset threshold, then the dynamic target is determined to be a lightweight dynamic target.

[0072] As an example, the target differentiation unit 203 is configured to: for each identified lightweight dynamic target, extract its historical motion trajectory point set within a time window, and fit the historical motion trajectory point set into a circle or ellipse; determine whether the motion of the lightweight dynamic target revolves around a stable geometric center based on the goodness of fit. Specifically, the target differentiation unit is configured to: if the goodness of fit is higher than a preset threshold, determine that the lightweight dynamic target is in a motion state based on a constraint center, and determine the fitted circle center or ellipse center as the constraint center, and then determine the number of constraint centers based on the fitted shape.

[0073] As an example, the target discrimination unit 203 is further configured to: after completing the initial fitting of a circle or ellipse, extract multi-dimensional features from the fitting process, including at least the distribution density of trajectory points, the uniformity of the fitting residual distribution, and the continuity of the distribution of projection points of the trajectory point set on the fitting curve; input the initial goodness of fit and the multi-dimensional features into a pre-trained prediction model, and perform a comprehensive analysis of the input features through the prediction model to output the corrected goodness of fit.

[0074] While the application has been particularly shown and described with reference to preferred embodiments, it will be understood to those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. Accordingly, the disclosed application is to be considered as illustrative and not restrictive, and the application is defined by the scope of the appended claims.

Claims

1. A laser beam-based AGV environment perception method, characterized in that, The method comprises the following steps: an AGV in a spinning workshop controls a vehicle-mounted laser radar to continuously scan the surrounding environment to obtain a set of time-sequentially continuous laser point cloud data frames; a plurality of dynamic targets are derived based on the set of laser point cloud data frames, and dynamic characteristic parameters of each dynamic target are extracted, at least including the existence duration and the motion trajectory feature of the dynamic target, and a lightweight dynamic target is identified from each dynamic target based on the dynamic characteristic parameters; motion center analysis is performed on the corresponding lightweight dynamic target based on the motion trajectory feature to obtain a motion center analysis result, including whether the lightweight dynamic target is in a motion state based on a constraint center and the number of constraint centers; a plurality of obstacle targets are determined based on the motion center analysis result, a driving trajectory of each obstacle target in the surrounding environment is planned, and the driving trajectory is executed. 2.The AGV environment perception method based on laser beam according to claim 1, wherein: After obtaining the set of time-sequentially continuous laser point cloud data frames, the method further comprises a step of data preprocessing of the set of point cloud data frames, specifically including: based on the structure prior information of the fixed equipment in the spinning workshop, static background filtering is performed on the point cloud data; for the point cloud data after filtering the static background, a statistical feature-based outlier filtering algorithm is used to remove discrete noise points caused by small-diameter suspended flying floss in the air that do not satisfy the spatial distribution continuity. 3.The AGV environment perception method based on laser beam according to claim 1, wherein: The lightweight dynamic target is identified from each dynamic target based on the dynamic characteristic parameters, including: the existence duration and the randomness index of the motion trajectory of each dynamic target are calculated; wherein the randomness index of the motion trajectory is quantified by calculating the standard deviation of the change amount of the motion direction of the dynamic target in the continuous frames; if the existence duration of the dynamic target is less than a first preset threshold and the randomness index of the motion trajectory is greater than a second preset threshold, the dynamic target is determined to be a lightweight dynamic target.

4. The AGV environment perception method based on laser beam according to claim 3, characterized in that: The motion center analysis result is obtained by performing motion center analysis on the corresponding lightweight dynamic target based on the motion trajectory feature, including: for each identified lightweight dynamic target, a historical motion trajectory point set within a time window is extracted, and a circular or elliptical fitting is performed on the historical motion trajectory point set; whether the motion of the lightweight dynamic target is around a stable geometric center is judged based on the fitting goodness, specifically: if the fitting goodness is higher than a preset threshold, it is determined that the lightweight dynamic target is in a motion state based on a constraint center, and the fitted center of the circle or ellipse is determined as the constraint center, and then the number of constraint centers is obtained according to the fitted shape.

5. The AGV environment perception method based on laser beam according to claim 4, characterized in that: The circular or elliptical fitting of the historical motion trajectory point set comprises: after completing the preliminary fitting of the circle or ellipse, multi-dimensional features in the fitting process are extracted, at least including the distribution density of the trajectory points, the distribution uniformity of the fitting residual, and the projection point distribution continuity of the trajectory point set on the fitted curve; the fitting goodness obtained by the preliminary fitting and the multi-dimensional features are jointly input into a pre-trained prediction model, the input features are comprehensively analyzed by the prediction model, and the corrected fitting goodness is output.

6. A laser beam based AGV environment perception system, characterized in that, The system comprises: a vehicle-mounted laser radar configured to continuously scan a surrounding environment of an AGV in a spinning mill, to obtain a set of time-sequentially continuous laser point cloud data frames; a target initial identification unit configured to derive a plurality of dynamic targets based on the set of laser point cloud data frames, to extract dynamic characteristic parameters of each dynamic target, including at least a duration of existence and a motion trajectory feature of the dynamic target, and to identify a lightweight dynamic target from each dynamic target based on the dynamic characteristic parameters; a target distinguishing unit configured to analyze a motion center of the corresponding lightweight dynamic target based on the motion trajectory feature, to obtain a motion center analysis result including whether the lightweight dynamic target is in a motion state based on a constraint center and a number of constraint centers; and a trajectory planning unit configured to determine a plurality of obstacle targets based on the motion center analysis result, to plan a driving trajectory of each obstacle target in the surrounding environment, and to execute.

7. The laser beam based AGV environment perception system of claim 6, wherein: The system further comprises a preprocessing unit configured to filter static backgrounds from the point cloud data based on structural prior information of fixed equipment in the spinning mill, and to remove discrete noise points that do not satisfy spatial distribution continuity caused by small-diameter suspended flying floss in the air from the point cloud data after filtering the static backgrounds, by using an outlier filtering algorithm based on statistical characteristics.

8. The laser beam based AGV environment perception system of claim 6, wherein: The target initial identification unit is configured to calculate a duration of existence and a randomness index of a motion trajectory of each dynamic target, wherein the randomness index of the motion trajectory is quantified by calculating a standard deviation of a change amount of a motion direction of the dynamic target in consecutive frames; if the duration of existence of the dynamic target is less than a first preset threshold and the randomness index of the motion trajectory is greater than a second preset threshold, the dynamic target is determined to be a lightweight dynamic target.

9. The laser beam based AGV environment perception system of claim 8, wherein: The target distinguishing unit is configured to extract a historical motion trajectory point set of each identified lightweight dynamic target within a time window, and to perform circular or elliptical fitting on the historical motion trajectory point set; and to determine whether the motion of the lightweight dynamic target is around a stable geometric center based on a fitting goodness, specifically: the target distinguishing unit is configured to determine that the lightweight dynamic target is in a motion state based on a constraint center if the fitting goodness is higher than a preset threshold, to determine a fitting center of a circle or an ellipse as the constraint center, and to determine a number of constraint centers according to the fitted shape.

10. The laser beam based AGV environment perception system of claim 9, wherein: The target distinguishing unit is further configured to extract multi-dimensional features during the fitting process after completing the preliminary circular or elliptical fitting, including at least a distribution density of the trajectory points, a distribution uniformity of the fitting residuals, and a projection point distribution continuity of the trajectory point set on the fitted curve; to input the fitting goodness obtained by the preliminary fitting and the multi-dimensional features into a pre-trained prediction model together; and to output a corrected fitting goodness by comprehensively analyzing the input features through the prediction model.

Images