A lidar sensor for detection

Abstract

The application belongs to the technical field of photoelectric measurement, and particularly relates to a laser radar sensor for detection, which firstly carries out dynamic target extraction and scene structure analysis processing on a real-time acquired point cloud sequence, generates a real-time scene understanding result, introduces target situation correlation and global perception state evaluation, judges a high-value scanning area in the scene, dynamically generates adaptive scanning track parameters according to the high-value scanning area, realizes accurate focusing scanning on a key target and a blind area, solves the problem of resource allocation imbalance and insufficient point cloud density in a key area of an existing laser radar fixed scanning mode in a dynamic complex scene, significantly improves resource utilization efficiency and data effectiveness, and further introduces an angle sensor feedback and tracking error compensation mechanism in a driving beam deflection device focusing scanning process, realizes closed-loop control, guarantees scanning accuracy, avoids equipment vibration and response delay, and is suitable for a high-speed dynamic scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photoelectric measurement, and specifically relates to a laser radar sensor for detection. BACKGROUND

[0002] As an active photoelectric detection system, the laser radar can obtain high-precision three-dimensional point cloud data of the surrounding environment in real time by emitting laser pulses and accurately measuring the return time or phase difference, and has become a core sensor for realizing environment perception and target recognition in the fields of automatic driving, robot navigation and intelligent mapping.

[0003] The current mainstream laser radar system generally adopts a fixed scanning mechanism, and a beam deflection device such as a mechanical rotation or a MEMS mirror is used to uniformly cover the field of view range according to a preset scanning track, such as linear array scanning and spiral scanning. This design ensures the completeness of the basic environment data acquisition and has good applicability in static or simple scenes.

[0004] However, the fixed scanning mode has significant defects: the indiscriminate resource allocation strategy causes a large number of laser pulses to be consumed in non-key static areas, resulting in data redundancy and resource waste. At the same time, the scanning density of dynamic key targets is insufficient, which seriously affects the target tracking accuracy and system response speed. The contradiction between this resource allocation imbalance and the mismatch with the scene demand has become the main technical bottleneck restricting the perception performance improvement of the laser radar in dynamic complex environments. SUMMARY

[0005] In order to overcome the shortcomings in the background art, the embodiments of the present application provide a laser radar sensor for detection, which can effectively solve the problems involved in the above background art.

[0006] The purpose of the present application can be achieved by the following technical scheme: a laser radar sensor for detection, comprising: a scene perception module, a value area determination module, a scanning strategy decision module and a scanning control module.

[0007] The scene perception module is connected with the value area determination module, the value area determination module is connected with the scanning strategy decision module, and the scanning strategy decision module is connected with the scanning control module.

[0008] The scene perception module performs dynamic target extraction and scene structure analysis processing on the real-time acquired point cloud sequence to generate a real-time scene understanding result.

[0009] The value area determination module evaluates the target situation correlation and global perception state according to the real-time scene understanding result to determine a high-value scanning area in the scene and output an identifier.

[0010] A scanning strategy decision module receives the identifier and dynamically generates adaptive scanning trajectory parameters for the high-value scanning area in combination with a preset scanning resource allocation strategy.

[0011] A scanning control module drives a beam deflection device to perform focused scanning processing on the high-value scanning area according to the adaptive scanning trajectory parameters, and generates optimized point cloud data.

[0012] Compared with the prior art, embodiments of the present application have at least the following advantages or beneficial effects: (1) The present application first performs dynamic target extraction and scene structure analysis processing on a real-time acquired point cloud sequence, generates a real-time scene understanding result, introduces target situation correlation and global perception state evaluation, determines a high-value scanning area in the scene, dynamically generates adaptive scanning trajectory parameters accordingly, realizes precise focused scanning on a key target and a blind area, solves the problem of unbalanced resource allocation and insufficient point cloud density in a key area of a fixed scanning mode of a laser radar in a dynamic complex scene, and significantly improves resource utilization efficiency and data effectiveness.

[0013] (2) The present application introduces an angle sensor feedback and tracking error compensation mechanism in the scanning control module, constructs a high-precision closed-loop control; at the same time, a smooth and continuous adaptive scanning trajectory is generated through kinematics optimization, and a transition trajectory meeting the dynamic constraint is inserted between multi-region scanning paths. This design guarantees the scanning accuracy of the system while effectively suppressing equipment vibration and response delay, and significantly improves the adaptability to high-speed dynamic scenes. BRIEF DESCRIPTION OF DRAWINGS

[0014] The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled persons in the art, other drawings can be obtained without creative labor on the basis of the following drawings.

[0015] Figure 1 The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled persons in the art, other drawings can be obtained without creative labor on the basis of the following drawings.

[0016] Figure 2 The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled persons in the art, other drawings can be obtained without creative labor on the basis of the following drawings.

[0017] Figure 3 The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled persons in the art, other drawings can be obtained without creative labor on the basis of the following drawings. DETAILED DESCRIPTION

[0018] In order to further clarify the technical means and effects taken by the present application to achieve the predetermined inventive purpose, the specific embodiments, structure, features and effects of a laser radar sensor for detection according to the present application are described in detail below in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0020] The specific scheme of a laser radar sensor for detection provided by the present application is described in detail below in combination with the drawings.

[0021] Please refer to Figure 1 which shows a module connection diagram of a laser radar sensor for detection provided by an embodiment of the present application, specifically including: a scene perception module, a value area determination module, a scanning strategy decision module and a scanning control module.

[0022] The scene perception module is connected with the value area determination module, the value area determination module is connected with the scanning strategy decision module, and the scanning strategy decision module is connected with the scanning control module.

[0023] The scene perception module performs dynamic target extraction and scene structure analysis processing on the real-time acquired point cloud sequence to generate real-time scene understanding results.

[0024] In an embodiment of the present application, laser pulses are emitted by a laser radar sensor and laser signals reflected back by targets are received to obtain original laser radar data. Due to factors such as environmental interference, sensor noise and motion distortion in the scanning process existing in the original data, noise suppression, data calibration and initial point cloud construction processing need to be performed on the original laser radar data to generate real-time point cloud sequences. It should be noted that the above preprocessing operations of the original laser radar data are prior art and will not be described here.

[0025] Considering that motion targets need to be perceived and tracked in a dynamic complex scene, based on this, in a preferred embodiment of the present application, the dynamic target extraction includes: dividing the point cloud sequence into multiple continuous point cloud frames in time sequence.

[0026] Each point cloud frame is subjected to semantic segmentation to define the semantic category of each point cloud in the scene.

[0027] Based on the semantic segmentation result of the current point cloud frame, point cloud data belonging to a specific dynamic object category and being spatially adjacent are clustered to form several independent candidate dynamic targets.

[0028] It should be noted that the specific dynamic object category refers to an object category associated with the current monitoring task of the sensor.

[0029] The candidate dynamic targets identified in the current point cloud frame are associated and matched with the target trajectories established based on historical frames, and the motion state parameters of the successfully associated targets over time are estimated.

[0030] Specifically, the association and matching process is as follows: for each established target trajectory, the expected position of the target in the current point cloud frame is calculated based on the historical position sequence of the target in the previous frames by linear extrapolation, wherein the linear extrapolation calculates the instantaneous velocity vector according to the positions of the target in the last two frames, and the product of this instantaneous velocity vector and the frame time interval is added to the position of the last frame to calculate the expected position in the current point cloud frame.

[0031] The candidate dynamic targets of the current point cloud frame are taken as rows, and the established target trajectories are taken as columns to construct an association cost matrix, wherein each element in the matrix represents the difference between the predicted position of a candidate dynamic target and an existing target trajectory in the current point cloud frame. The calculation process of the difference is as follows: the Euclidean distance between the center position of the dynamic candidate target and the expected position of the trajectory is calculated as a spatial distance measure.

[0032] The intersection-over-union of the three-dimensional bounding box of the dynamic candidate target and the historical three-dimensional bounding box of the target trajectory in the length, width and height dimensions is calculated, and the difference between 1 and the average of the intersection-over-union is taken as a geometric feature difference measure.

[0033] The normalized spatial distance measure and the geometric feature difference measure are added to obtain the difference.

[0034] Any dynamic candidate target and any target trajectory are randomly combined to generate an association pair, a data association algorithm is applied to globally optimally match and solve the association cost matrix, the association cost value of each association pair is obtained, if the association cost value of a certain association pair is less than a preset association threshold, it is determined that the matching relationship between the dynamic candidate target and the target trajectory in this association pair is valid, and finally a candidate dynamic target with the lowest cost value and below the preset association threshold is assigned to each established target trajectory, otherwise it is determined that the trajectory is mismatched in the current point cloud frame. It should be noted that the data association algorithm can use the Hungarian algorithm or the KM algorithm of the prior art, and its solving process is not described here.

[0035] It is also necessary to explain that the data source of the preset association threshold is based on the collection of positive and negative sample pairs according to historical data distribution, wherein the positive sample pair is a successfully matched association pair, and the negative sample pair is an association pair that should not be matched but is determined to be valid by the algorithm.

[0036] The association cost of all the above positive and negative sample pairs is calculated, and the distribution curves of the positive and negative sample association costs are fitted respectively, and the preset association threshold is set as the intersection point of the two distribution curves.

[0037] Based on the above association matching result, it also includes trajectory maintenance and management: trajectory initialization: creating a new trajectory for a candidate dynamic target that is not matched to any existing trajectory.

[0038] Trajectory confirmation: when a trajectory is continuously associated and updated for a preset number of point cloud frames, it is marked as a confirmed dynamic target.

[0039] Trajectory deletion: when a trajectory is continuously associated for a preset number of point cloud frames, the trajectory is deleted.

[0040] Continuously output the category information, three-dimensional contour and real-time updated motion state parameters of each confirmed dynamic target.

[0041] In a preferred embodiment of the present application, the scene structure analysis includes: based on the dynamic target extraction result, removing the point cloud identified as a dynamic target from the current point cloud frame point cloud to obtain a static scene point cloud.

[0042] Based on the spatial geometric features, a dominant plane is segmented from the static scene point cloud.

[0043] The above dominant plane segmentation process is realized by iteratively performing the following steps: randomly sampling a minimum point set from the static scene point cloud to establish a candidate plane model, then calculating the geometric distance between all other points in the point cloud and the candidate plane, identifying points with a distance less than a preset distance threshold as inliers of the model to form a support point set.

[0044] Repeat the above sampling and verification process to finally select the candidate plane model with the largest support point set as the optimal plane model.

[0045] In the obtained optimal plane model, further compare its normal vector with the known gravity direction vector, select the plane with the smallest angle between the normal vector and the gravity direction, and determine it as the dominant plane.

[0046] Cluster the non-dominant plane point cloud to generate multiple independent point cloud clusters.

[0047] Analyze the spatial geometric properties of each point cloud cluster and its relative pose with the dominant plane, and classify the point cloud clusters into different static structure elements.

[0048] It should be noted that the spatial geometric properties include at least the size of the three-dimensional bounding box of the point cloud cluster, and the relative pose of the point cloud cluster to the dominant plane includes at least the included angle between the main direction of the point cloud cluster and the normal of the dominant plane.

[0049] The static structural elements include at least facade structure, rod structure and body structure.

[0050] The specific process of classifying the point cloud cluster into different static structural elements is as follows: in view of the vertical extension and high saliency characteristics of the facade structure in space, the ratio of the height to the width of the three-dimensional bounding box of the point cloud cluster is calculated, if the ratio is greater than or equal to 2:1, and the included angle between the main direction of the point cloud cluster and the normal of the dominant plane is in the range of 80 degrees to 100 degrees, the point cloud cluster is preferentially classified as a facade structure, for further clarification, the facade structure can be exemplified as a wall, a fence or a large sign.

[0051] Also considering that the rod structure has a significant extension characteristic in a certain dimension, if the size of two dimensions of the point cloud cluster is less than the size of the third dimension and the ratio of the size of the two dimensions to the size of the third dimension is 2 or more, the point cloud cluster is preferentially classified as a rod structure, and the rod structure can be exemplified as a tree trunk, a lamp pole or a power pole.

[0052] Furthermore, due to the balanced characteristics of the body structure in three-dimensional space, the ratio of the longest side to the shortest side of the three-dimensional bounding box of the point cloud cluster is less than or equal to 1.5:1, the point cloud cluster is preferentially classified as a body structure.

[0053] It should be explained that the various numerical values used in the classification process of the static structural elements are typical configuration parameters verified. Those skilled in the art understand that these parameters can be appropriately adjusted and customized according to the actual monitoring task requirements and the characteristics of the specific application scene without departing from the core classification principles of the present application.

[0054] The structured description information of the scene is generated by combining the dominant plane and the classified static structural elements.

[0055] The value area determination module evaluates the target situation relevance and the global perception state according to the real-time scene understanding result, determines a high-value scanning area in the scene and outputs an identifier.

[0056] In a preferred embodiment of the present application, the target situation relevance evaluation process includes: according to the mapping relationship between the category information of the dynamic target and the inherent importance of the current monitoring task of the sensor, a basic value weight is assigned to each dynamic target.

[0057] Based on the relative motion relationship between the dynamic target and the sensor, the potential influence degree of the dynamic target on the current monitoring task of the sensor is quantified and converted into a motion situation value score.

[0058] The quantification process of the potential influence degree is: monitoring the relative speed vector, the distance between the dynamic target and the sensor, the predicted collision time of the motion direction, and establishing a fuzzy rule base based on the above relative motion parameters to map the rule relationship between different relative motion parameter value combinations and the potential influence degree of the monitoring task. Each mapping rule is assigned a corresponding potential influence degree and confidence.

[0059] According to the relative motion relationship between the current dynamic target and the sensor, all mapping rules matching the relative motion relationship in the fuzzy rule base are retrieved, and the potential influence degree value of the mapping rule corresponding to the maximum confidence is selected as the final quantified potential influence degree result.

[0060] Considering that a higher potential influence degree of a dynamic target means that it has a greater negative impact on the completion of the monitoring task of the sensor, it is necessary to allocate higher attention and scanning resources to deal with it. In the logic of resource allocation, the potential influence degree should be positively correlated with the motion posture value score. The specific conversion method can exemplarily use a monotonically increasing function to multiply the potential influence degree by a preset scaling coefficient to obtain the motion posture value score.

[0061] The product of the motion posture value score and the basic value weight is used as the posture relevance evaluation index of the dynamic target.

[0062] Referring to Figure 2 As shown in the preferred embodiment of the present application, the global perception state assessment process includes: constructing a state prediction covariance matrix of each dynamic target according to the real-time updated motion state parameters of the dynamic target, and taking the trace of the covariance matrix as an index for quantifying the uncertainty of the dynamic target.

[0063] The information potential score of each planning area is calculated by combining the point cloud density, the key target distance, and the map state, and planning the static occlusion blind area, the historical sparse area, and the dynamic target forward area in the scene.

[0064] It should be noted that the static occlusion blind area is an occlusion area formed by the identified static structural elements under the sensor view.

[0065] The historical sparse area is an area where the point cloud density is continuously lower than the average density of the scene in the past continuous preset number of point cloud frames.

[0066] The dynamic target forward area is a potential area extending along the motion direction of the dynamic target with high posture relevance. The high posture relevance target refers to the target whose motion posture value score is in the top preset percentage among all dynamic targets. The generation method of the area is to take the current position of the target as the starting point and extend a distance along the velocity vector direction. The distance is obtained by multiplying the current speed value of the target by the preset forward prediction time window.

[0067] In calculating the information potential score of each planning area, the following process is performed: set the observation history factor according to the average coverage density of the planning area in the past preset number of point cloud frames, the lower the average coverage density, the smaller the value of the observation history factor.

[0068] Set the spatial correlation factor according to the Euclidean distance of the planning area relative to the high-situation-related dynamic target, the closer the distance, the greater the spatial correlation factor.

[0069] Set the prior knowledge factor according to the state of the planning area in the existing environment map, the state including one of known occupancy state, known free state and unknown state, the known occupancy state corresponding to the highest value of the prior knowledge factor, the known free state second, and the unknown state lowest.

[0070] Among them, the specific value of each factor can be determined by the implementer in the system development stage, according to the typical characteristics of the target application scene and the monitoring task requirements, through simulation test and parameter tuning process, to adaptively determine the value strategy or function parameter of each factor. The following gives two examples as a reference, which are not elaborated here: I. Predefined value table, mapping different ranges of input parameters to corresponding factor values.

[0071] Ii. Normalization function processing, linear or nonlinear conversion of input parameters to a preset value interval.

[0072] Add the observation history factor, spatial correlation factor and prior knowledge factor to obtain the information potential score.

[0073] Among them, a 2D grid map or 3D voxel map data structure is used to establish and maintain the environment map, and the map state of each planning area is captured in the actual maintenance process of the map. Specifically: set the initial state of the environment map to global unknown.

[0074] For each frame of point cloud data, the following operations are performed: mark the state of the map grid cell hit by the point cloud as known occupancy.

[0075] The grid cells through which the ray between the sensor origin and the hit grid cell passes are marked as known free if their state is not known occupancy.

[0076] When evaluating the information potential of the region, the current state distribution of the grid cells in each planning area is obtained by querying the environment map, and the state with the largest proportion of grid cells is taken as the map state of the region.

[0077] The quantitative indicator of dynamic target uncertainty is mapped to its estimated position in the form of a point heat map.

[0078] Each planning area information potential score is mapped to its corresponding continuous spatial range in the form of a planar heat map.

[0079] The planar heat map and the point heat map are superimposed to generate a global awareness state map.

[0080] In a preferred embodiment of the present application, the high-value scanning area determination process in the scene includes: extracting the area in which the information potential score or the quantitative indicator of dynamic target uncertainty in the global awareness state map exceeds the corresponding preset threshold, and marking the area as a candidate area.

[0081] The potential state correlation evaluation indicator, the quantitative indicator of uncertainty, and the information potential score of all dynamic targets in the candidate area are fused to obtain a comprehensive value score of each candidate area.

[0082] In an embodiment of the present application, the above fusion process can be specifically implemented by using a geometric mean method. First, considering that the quantitative indicator of uncertainty is negatively correlated with the comprehensive value, the quantitative indicator of uncertainty needs to be converted into the absolute difference between the quantitative indicator of uncertainty and 1, so that the numerical value is positively correlated with the value. Subsequently, the processed uncertainty indicator, the potential state correlation evaluation indicator, and the information potential score are multiplied, and the geometric mean value is calculated as the comprehensive value score of the candidate area. This method can effectively capture the synergistic and balanced effect between indicators, and a significant decrease in any indicator will directly inhibit the final score.

[0083] In other embodiments, the implementer can also use other fusion methods, such as a Bayesian fusion method: regarding uncertainty as probability, using the Bayesian theorem to update information potential and fuse potential state correlation, or a conventional linear weighted combination method, which will not be described again.

[0084] According to the upper limit of the scanning resource of the sensor, all candidate areas are sorted according to their comprehensive value scores, and one or more areas with the highest ranking are selected, and their spatial coordinates and priorities are encoded into high-value scanning area identifiers and output.

[0085] The scanning strategy decision module receives the identifiers, combines a preset scanning resource allocation strategy, and dynamically generates adaptive scanning trajectory parameters for the high-value scanning areas.

[0086] In a preferred embodiment of the present application, the preset scanning resource allocation strategy has the following constraint conditions, including: different scanning resource quotas are allocated to different areas according to the priorities in the high-value scanning area identifiers, and the highest priority area is granted the highest point cloud density quota and the longest scanning time quota.

[0087] In the scanning resource quota process, the cumulative scanning time of all regions is required to be less than the single-frame period budget, and the planned scanning trajectory has a kinetic characteristic that does not exceed the maximum angular velocity and the maximum angular acceleration of the beam deflection device.

[0088] Referring to Figure 3 As shown in the preferred embodiment of the present application, the adaptive scanning trajectory parameter dynamic generation process includes: according to the priority in the high-value scanning region identifier, combining the single-frame resource configuration constraint, and allocating the scanning time and point cloud density quota according to the preset weight ratio.

[0089] It should be noted that the above-mentioned preset weight ratio is specifically calculated, and the calculation method is: based on the geometric boundary box form, automatically mapping the optimal scanning path from the pre-configured scanning mode library.

[0090] The pre-configured scanning mode library contains a plurality of optimal scanning path templates matched with typical geometric shapes, and the automatic mapping process of the pre-configured scanning mode library includes: obtaining the shape feature vector of the candidate region geometric boundary box, which includes but is not limited to the long-short axis ratio, the rectangularity, and the circularity, calculating the cosine similarity of the shape feature vector and the standard feature vector of each template in the mode library, and finally selecting the template with the smallest cosine similarity as the matching result.

[0091] As an example of the above optimal scanning path template, when the region geometric boundary box is identified as a circle or a circle-like shape, a spiral filling scanning path will be automatically mapped and adopted, which starts from the center of the circle and spirals outward at a constant angular rate or a constant time increment until the entire circular region is covered. This path planning can ensure that the scanning point cloud has high and uniform density distribution in the radial and circumferential directions.

[0092] When the boundary box is rectangular or square, the arch-shaped grid scanning path will be called. This path starts from a corner of the rectangle, performs linear scanning along the long side direction, and after reaching the boundary, moves one resolution unit along the short side direction in a step-by-step manner, and then scans in the opposite direction, thereby achieving efficient and non-missing coverage of the entire rectangular region.

[0093] When the boundary box is an irregular convex polygon, a traversal scanning path based on polygon triangulation is used. This path first triangulates the polygon region into several triangular sub-regions, and then performs the aforementioned grid scanning on each triangular sub-region in an optimal order, and finally connects the scanning trajectories of each sub-region by using an existing path planning algorithm to form an optimal path that covers the entire polygon and has the shortest empty movement.

[0094] In other embodiments, the implementer can also preset more complex adaptive paths in the scanning mode library according to the specific performance of the sensor and the task requirements, such as generating a trajectory online using model predictive control for a dynamically deformed region, and no longer be described in detail.

[0095] Based on the quota and the scanning path, a preliminary scanning trajectory is generated.

[0096] The preliminary scanning trajectory is kinematically optimized to obtain an adaptive scanning trajectory, and is simultaneously converted into a sequence of beam deflection instructions. It needs to be particularly pointed out that the sequence of beam deflection instructions is specifically an angle-time instruction sequence of the beam deflection device coded in time sequence.

[0097] In a preferred embodiment of the present application, the kinematic optimization process includes: smoothing the preliminary scanning trajectory to generate a smooth trajectory with continuous angular velocity and angular acceleration.

[0098] The smoothing process is specifically performed according to the following steps: discretizing the preliminary scanning trajectory into a sequence of path points, calculating the minimum passing time of each path segment according to the total length of the path and the maximum angular velocity of the beam deflection device, and summarizing the total time budget for trajectory execution.

[0099] Based on the curvature of the discrete path points, the maximum allowed angular velocity that each point can reach without exceeding the maximum angular acceleration constraint is calculated, and a velocity curve that meets the dynamics of the beam deflection device is generated.

[0100] The discrete path points and the corresponding planned velocities are jointly generated using a B-spline curve fitting algorithm to generate a smooth scanning trajectory with continuous angular velocity and angular acceleration.

[0101] The execution process of the B-spline curve fitting algorithm is as follows: first, the control vertices are calculated by taking the discrete path points as the type value points to generate a geometrically smooth path curve; then, a time-parameter mapping function that follows the velocity and acceleration constraints is constructed; finally, a smooth trajectory with continuous angular velocity and angular acceleration is synthesized through a composite function. Since this algorithm is a mature technology, the present application will not be described in detail.

[0102] Between the scanning trajectories of multiple regions, a transition path that meets the dynamics constraints is inserted.

[0103] The dynamic constraint specifically refers to that the angular velocity and the angular acceleration in the multi-region transition process do not exceed the maximum allowable limit of the beam deflection device, and under this constraint condition, the transition path insertion process specifically performs the following: between the scan trajectories of multiple regions, the initial planning of the transition path takes time optimization as the primary goal. Specifically, the system uses a straight line connection or a simplified geometric curve such as a circular arc as the initial transition path. The path directly connects the end point of the previous scan trajectory and the starting point of the next scan trajectory, aiming to theoretically achieve the shortest moving time between the two points.

[0104] After generating the initial transition path, the angular velocity and the angular acceleration of each discrete path point in the initial transition path are calculated, and the maximum angular velocity and the maximum angular acceleration are screened and compared with the maximum allowable limit of the beam deflection device.

[0105] If none of them exceeds the limit, the initial transition path is used.

[0106] If the limit is exceeded, on the basis of the initial transition path, transition curves that meet the maximum angular acceleration and the maximum angular velocity constraint are inserted near the starting point and the ending point of the transition path. The core idea of the transition curve insertion is to increase the path length in exchange for the improvement of motion stability.

[0107] Exemplarily, the transition curve can be a polynomial curve, a Bezier curve or a B-spline curve.

[0108] The embodiment of the application first generates real-time scene understanding results by dynamically extracting targets and analyzing scene structures from the real-time acquired point cloud sequence, introduces target situation correlation and global perception state evaluation, determines high-value scanning areas in the scene, dynamically generates adaptive scanning trajectory parameters accordingly, realizes accurate focusing scanning of key targets and blind areas, solves the problem of unbalanced resource allocation and insufficient point cloud density in key areas of the existing fixed scanning mode of the laser radar in a dynamic complex scene, and significantly improves resource utilization efficiency and data effectiveness.

[0109] The scanning control module drives the beam deflection device to perform focusing scanning processing on the high-value scanning area according to the adaptive scanning trajectory parameters, and generates optimized point cloud data.

[0110] In a preferred embodiment of the application, the scanning control module, when performing focusing scanning, acquires the actual deflection angle of the beam deflection device in real time through the built-in angle sensor.

[0111] The actual deflection angle and the angle tracking error of the adaptive scanning trajectory are calculated.

[0112] A compensation signal is generated based on the tracking error and combined with a feedforward control signal.

[0113] It should be noted that the compensation signal is generated by a PID controller, which independently calculates the proportional term, integral term and differential term of the angle tracking error, and sums the three terms to output the compensation signal.

[0114] The synthesized signal is converted into a driving signal to drive the beam deflection device to perform a focusing scanning action.

[0115] In the scanning control module, the embodiment of the application introduces an angle sensor feedback and tracking error compensation mechanism to construct a high-precision closed-loop control; at the same time, a smooth and continuous adaptive scanning trajectory is generated through kinematic optimization, and a transition trajectory meeting the dynamic constraint is inserted between the multi-region scanning paths. This design effectively suppresses the device vibration and response delay while ensuring the scanning accuracy of the system, and significantly improves the adaptability to high-speed dynamic scenes.

[0116] The above is merely an example and description of the structure of the application, and those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, as long as they do not deviate from the structure of the application or exceed the scope defined by the application, which shall belong to the protection scope of the application.

Claims

1. A lidar sensor for detection, characterized in that Comprise: A scene perception module, which performs dynamic target extraction and scene structure analysis on a real-time acquired point cloud sequence to generate real-time scene understanding results; A value area determination module, which evaluates target situation relevance and global perception state based on the real-time scene understanding results to determine high-value scanning areas in the scene and output identifiers; A scanning strategy decision module, which receives the identifiers and dynamically generates adaptive scanning trajectory parameters for the high-value scanning areas in combination with a preset scanning resource allocation strategy; A scanning control module, which drives a beam deflection device to perform focused scanning on the high-value scanning areas according to the adaptive scanning trajectory parameters to generate optimized point cloud data; Different scanning resource quotas are allocated to different areas according to the priorities in the high-value scanning area identifiers; The adaptive scanning trajectory parameter dynamic generation process comprises: allocating scanning time and point cloud density quotas to the high-value scanning areas in a preset weight ratio based on the priorities in the high-value scanning area identifiers and in combination with single-frame resource configuration constraints; Based on the area geometric bounding box morphology, the optimal scanning path is automatically mapped from a preconfigured scanning mode library; based on the quotas and the scanning path, the preliminary scanning trajectory is generated; the preliminary scanning trajectory is kinematically optimized to obtain the adaptive scanning trajectory, which is then converted into a beam deflection instruction sequence.

2. The lidar sensor for detection according to claim 1, characterized in that The dynamic target extraction comprises: The point cloud sequence is divided into multiple continuous point cloud frames in chronological order; Semantic segmentation is performed on each point cloud frame to define the semantic category of each point cloud in the scene; Based on the semantic segmentation results of the current frame, point cloud data belonging to the dynamic object category and spatially adjacent are clustered to form several independent candidate dynamic targets; The candidate dynamic targets identified in the current frame are associated and matched with the target trajectories established based on historical frames, and the motion state parameters of the successfully associated targets are estimated over time; The category information, three-dimensional contour, and real-time updated motion state parameters of each confirmed dynamic target are continuously output.

3. The lidar sensor for detection according to claim 2, characterized in that The scene structure analysis comprises: Based on the dynamic target extraction results, the point clouds identified as dynamic targets are removed from the current point cloud frame to obtain static scene point clouds; The dominant plane is segmented from the static scene point clouds based on spatial geometric features; The non-dominant plane point clouds are clustered to generate multiple independent point cloud clusters; The spatial geometric properties of each point cloud cluster and its relative pose with the dominant plane are analyzed to classify the point cloud clusters into different static structural elements; The structured description information of the scene is generated by integrating the dominant plane and the classified static structural elements.

4. The lidar sensor for detection according to claim 2, characterized in that The target situation relevance evaluation process comprises: Based on the category information of the dynamic targets and the inherent importance mapping relationship of the current monitoring task of the sensor, a basic value weight is assigned to each dynamic target; Based on the relative motion relationship between the dynamic targets and the sensor, the potential influence degree of the dynamic targets on the current monitoring task of the sensor is quantified and converted into a motion situation value score; The product of the motion situation value score and the basic value weight is taken as the situation relevance evaluation index of the dynamic target.

5. A lidar sensor for detection according to claim 4, characterized in that The global perception state evaluation process comprises: According to the real-time updated motion state parameters of the dynamic target, a state prediction covariance matrix of each dynamic target is constructed, and a trace of the covariance matrix is taken as an index for quantifying the uncertainty of the dynamic target; A static occlusion blind area, a historical sparse area, and a dynamic target forward-looking area in the planning scene are planned, and the information potential score of each planning area is calculated in combination with the point cloud density, the key target spacing, and the map state; The quantified index of the uncertainty of the dynamic target is mapped to the estimated position of the dynamic target in the form of a point heat map; The information potential score of each planning area is mapped to the corresponding continuous space range in the form of a surface heat map; The surface heat map and the point heat map are superimposed to generate a global perception state map.

6. A lidar sensor for detection according to claim 5, characterized in that The high-value scanning area determination process in the scene includes: Extracting the area in which the information potential score or the quantified index of the uncertainty of the dynamic target exceeds the corresponding preset threshold in the global perception state map, and marking the area as a candidate area; Fusing the potential state correlation evaluation index, the quantified index of the uncertainty, and the information potential score of all dynamic targets in the candidate area to obtain a comprehensive value score of each candidate area; According to the upper limit of the scanning resource of the sensor, all candidate areas are sorted according to their comprehensive value scores, and one or more areas with the highest ranking are selected, and their spatial coordinates and priority are encoded as a high-value scanning area identifier and output.

7. The lidar sensor for detection according to claim 1, characterized in that The preset scanning resource allocation strategy has the following constraint conditions, including: The highest priority area is granted the highest point cloud density quota and the longest scanning time quota; During the scanning resource allocation process, the cumulative scanning time of all areas is required to be less than the single-frame period budget, and the kinetic characteristics of the planned scanning trajectory are required to be less than the maximum angular velocity and the maximum angular acceleration of the beam deflection device.

8. The lidar sensor for detection according to claim 7, characterized in that The kinematic optimization process includes: Smooth the preliminary scanning trajectory to generate a smooth trajectory with continuous angular velocity and angular acceleration; Insert a transition path that meets the dynamic constraints between the scanning trajectories of multiple areas.

9. The lidar sensor for detection according to claim 1, characterized in that The scanning control module includes the following when performing focused scanning: Real-time acquisition of the actual deflection angle of the beam deflection device through the built-in angle sensor; Calculating the angle tracking error of the actual deflection angle and the adaptive scanning trajectory; Generating a compensation signal based on the tracking error and combining it with the feedforward control signal; Convert the combined signal into a driving signal to drive the beam deflection device to perform focused scanning actions.

Images