A low, slow and small target detection and trajectory prediction tracking method based on laser radar

Abstract

The application discloses a low, slow and small target detection and track prediction method based on a laser radar, first acquires laser radar point cloud data, and carries out space modeling and coordinate transformation preprocessing on the laser radar point cloud data; carries out saliency screening; carries out Pillar construction and coding based on distance partition driving; constructs a deep learning detection network; carries out size adaptation on the weak aerial target in the fused features based on an anchor design and matching strategy; carries out time sequence prediction and observation update on the target state based on an extended Kalman filter (EKF), and completes low, slow and small target detection and track prediction. The application can retain the high-precision advantage of the laser radar, improve the identification capability of the laser radar on weak reflection, small volume and irregular motion targets, and has strong robustness and environmental adaptability, so that stable, accurate perception and continuous tracking of low, slow and small flight targets are realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of control theory and signal processing, and particularly relates to a low, slow and small target detection and trajectory prediction tracking method based on laser radar. BACKGROUND

[0002] With the wide application of small flight platforms such as unmanned aerial vehicles and model aircrafts, new flight targets with the typical characteristics of "low altitude, slow speed and small size" (referred to as "low, slow and small") frequently appear in the fields of city, transportation, energy and military. Due to the characteristics of small size, low flight altitude, slow speed, weak radar reflection characteristics and easy concealment, this kind of target increasingly highlights its potential threat in the current airspace safety management. In particular, in the urban environment, the low, slow and small target can easily bypass the monitoring of the traditional air defense system, forming a serious security risk. Therefore, how to realize the high-precision, all-weather and full-time domain detection and stable tracking of the low, slow and small target has become an important research direction in the field of intelligent perception and low-altitude security.

[0003] Among many detection technologies, laser radar (LiDAR) has great potential in target three-dimensional perception and spatial positioning due to its high precision, high spatial resolution and independence from light conditions. Compared with traditional millimeter wave radar, laser radar can obtain high-density point cloud data and has stronger object contour restoration capability, which is suitable for geometric feature extraction and motion trajectory tracking of small targets in complex environments. Especially in the multi-obstacle urban environment, laser radar can effectively obtain the spatial position information of the target, assisting in accurately judging its position, attitude and motion state.

[0004] However, laser radar also faces a series of challenges when used for low, slow and small target detection. On the one hand, due to the small size and irregular flight trajectory of the low, slow and small target, the reflected laser signal is weak, which is sparse and discontinuous in the point cloud, and is easily submerged by background ground points, building reflection points or tree crowns, etc., resulting in low detection accuracy. On the other hand, limited by the scanning frequency and field of view range of the laser radar, when the target moves quickly, changes direction suddenly or is temporarily blocked, the difference between the point cloud data frames is large, which brings significant difficulties to continuous tracking and state prediction. In addition, the laser echo in the low-altitude environment is also easily affected by factors such as multi-path reflection, rain and fog shielding, further increasing the complexity of target identification.

[0005] Although some studies attempt to combine point cloud segmentation, feature clustering, and deep learning methods to improve the detection ability of laser radar for low, slow, and small targets, there are still many challenges in practical applications. For example, when the reflection characteristics of low, slow, and small targets are weak, the size is small, or the motion trajectory is complex, existing methods often rely on strong prior information or ideal environmental conditions, making it difficult to ensure the stability of detection and tracking in general scenarios. At the same time, many existing algorithms are mainly designed for ground targets (such as vehicles and pedestrians), which have essential differences in target size, motion pattern, spatial distribution, and other aspects from low, slow, and small targets in the air, resulting in limited generalization ability in the context of flying targets.

[0006] In contrast to laser radar, traditional visible light imaging systems can provide rich texture and color information, but their performance drops sharply in unstable lighting, backlight, or night conditions, and they lack three-dimensional spatial positioning capability. Especially in cases where the target and background colors are not clearly contrasted, false positives and false negatives can easily occur, and target loss can occur in short occlusions or high-speed flight.

[0007] Event cameras, with their high temporal resolution and sensitivity to dynamic targets, can quickly respond to sudden movement events, but they have limited ability to express the structure and texture features of targets and are only suitable for assisting in determining target trajectories, not as the main detection method. Infrared thermal imaging is suitable for night and low-visibility environments, but it has limited accuracy in long-range target recognition due to the weak thermal radiation and low thermal contrast of low, slow, and small targets. Radio frequency and voiceprint detection have the advantages of all-weather and low cost, but their identification accuracy is not high in noisy environments, and they cannot provide accurate spatial location information, making it difficult to achieve high-precision tracking.

[0008] In summary, in complex urban or open low-altitude environments, the detection and tracking of low, slow, and small targets require a system that has high spatial resolution, large field-of-view coverage, strong anti-interference capability, target geometric structure restoration capability, and stable tracking capability. Laser radar is one of the important technical paths to achieve this goal due to its rich three-dimensional spatial information and ability to perceive non-cooperative targets. However, existing laser radar detection techniques still have problems such as insufficient perception of sparse targets, frequent tracking interruptions, and low identification accuracy when applied to low, slow, and small target scenarios, making it difficult to meet the urgent need for continuous and reliable monitoring of low, slow, and small targets in real-world environments. SUMMARY

[0009] The purpose of the present application is to provide a laser radar-based low, slow, and small target detection and trajectory prediction tracking method. It can retain the high-precision advantages of laser radar while improving its ability to identify weakly reflective, small-volume, and irregularly moving targets, and has strong robustness and environmental adaptability, thereby achieving stable and accurate perception and continuous tracking of low, slow, and small flying targets.

[0010] Technical scheme: a low, slow and small target detection and trajectory prediction tracking method based on laser radar, comprising the following steps: step 1, collecting laser radar point cloud data, and preprocessing the laser radar point cloud data by space modeling and coordinate transformation;

[0011] Step 2, significant screening is performed on the preprocessed data to obtain screened point cloud data.

[0012] Step 3, based on distance partition driving, the screened point cloud data is subjected to Pillar construction and coding to obtain a pillar expression vector.

[0013] Step 4, a deep learning detection network is constructed, taking the pillar expression vector as input and outputting fused features.

[0014] Step 5, based on the anchor design and matching strategy, the size of the aerial weak target in the fused features is adapted.

[0015] Step 6, based on the extended Kalman filter EKF, the target state is time series predicted and observed updated, and the low, slow and small target detection and trajectory prediction tracking are completed.

[0016] Further, step 1 is specifically: measuring the vertical height of the radar from the installation reference surface to the bottom of the sensor , recording the pitch angle of the radar at the initial installation , horizontal rotation deflection , and horizontal position offset of the radar relative to the origin of the carrier coordinate system , inputting the data into the calibration module, and constructing the extrinsic matrix of the radar sensor under the platform coordinate system ; in the point cloud collection process, each frame of original point cloud data is represented in local coordinates as:

[0017] .

[0018] Among them, is the original data of each frame of point cloud, , , is the horizontal coordinate, vertical coordinate and height information of each point cloud in the local coordinate system, is the reflection intensity of each point cloud, and N is the number of point clouds in each frame of original point cloud data; read the extrinsic parameter , and perform coordinate transformation on each point:

[0019] .

[0020] The position of the point in the global space of the platform after transformation .

[0021] The vertical height information of the calculation point relative to the ground: .

[0022] The spatial distance of each point is calculated: . According to the threshold value set by experience , the point cloud is divided into three spatial regions: .

[0023] Wherein, is the spatial region of the point cloud, is the threshold limit of the point cloud near distance, middle distance and far distance set by experience.

[0024] Further, step 2 is specifically: after each frame of point cloud collection is completed, the reflection intensity value of all points is scanned, and the maximum value and the minimum value of the frame are extracted , and the reflection intensity of each point is normalized: .

[0025] The KD-Tree neighborhood search is used to count the number of neighborhood points in the three-dimensional neighborhood of each point : .

[0026] Wherein, is the target point currently being calculated for saliency, is any other point in the point cloud ;

[0027] The number of neighborhood points is normalized: .

[0028] Wherein, is the normalized local density value, is the maximum value of the local density value of all points in the current frame point cloud.

[0029] After normalization in two dimensions, the saliency score S i is constructed for each point: .

[0030] Wherein, ; An adaptive weight adjustment mechanism is introduced: when a point belongs to the Far Field region, the weight is adjusted upwards, and the saliency threshold is lowered: .

[0031] Wherein, and represent the dynamic basic threshold coefficient, Ts far The significance threshold indicating that the point cloud is in the Far Field region. Finally, according to the threshold, it is judged whether to retain the point into the subsequent pillar construction: .

[0032] Further, step 3 is specifically: reading the pillar parameter table, using the grid index method to divide the pillar, and marking the pillar number according to the selected region Grid divides the BEV space.

[0033] Each grid point corresponds to a pillar column, and the screened and retained point cloud is projected to the BEV grid, and the corresponding pillar number is marked. If the number of points in a certain pillar is greater than or equal to the corresponding pillar number in the region, it is considered as a legal pillar, and enters the next stage of coding, and the remaining point cloud is removed; for each point in the legal pillar, the offset of the point and the center of the pillar and the offset of the point and the grid position of the pillar center are calculated, and then the feature vector is spliced, the feature of each point is coded by sharing MLP, and then the maximum pooling is performed in the pillar dimension, forming a fixed-length pillar expression vector.

[0034] Further, step 4 is specifically: in the shallow and middle layer feature output of the backbone network, local convolution response is introduced for self-attention estimation, and the significance weight of each spatial position on the feature map is calculated , and it is used as a scale factor to reweight the original feature map: .

[0035] Wherein, represents the original BEV feature map, is the enhanced result. The resolution recovery mechanism from top to bottom is adopted, the high-level semantic features are gradually up-sampled, and the low-level detail features are spliced and fused; the fusion strategy is expressed in the following form: .

[0036] Wherein, is the current scale feature map, is the feature map of the upper level, and after splicing, the feature dimension is integrated by convolution to generate a unified channel fusion result.

[0037] Further, step 5 is specifically: loading multiple sets of preset anchor frame templates through a parameter file, each Anchor containing three-dimensional size of length, width and height, center point height and direction angle information; according to the actual envelope size of the aerial weak target point cloud, the mean value is obtained by fitting the 3D bounding box of multiple frames of actual measurement data; the positive and negative sample division mechanism based on 3D IoU is used to calculate the volume intersection ratio between each Anchor prediction frame and the real target frame GT; the IoU lower limit dynamic adjustment strategy based on distance interval is as follows: .

[0038] wherein, is the spatial distance from the Anchor center to the radar, , is the partition threshold set in the system; and the center height of each Anchor is adjusted.

[0039] Further, step 6 is specifically: for the spatial flight characteristics of low, slow and small targets, a six-dimensional state vector is used to model the trajectory, the first three dimensions representing the three-dimensional spatial position of the target in the current frame, and the last three dimensions representing the velocity components of the target.

[0040] Assuming that the target satisfies the uniform straight flight model in a short time, a state prediction model is established: .

[0041] wherein, X k-1 is the posterior estimation value of the target state vector at the last time, is a state transition matrix, defined as: .

[0042] wherein, is the sampling period determined according to the scanning frequency of the laser radar.

[0043] Process noise obeys zero-mean Gaussian distribution, which is used to absorb nonlinear disturbances or control errors outside the model; the observation model is constructed based on the detection results of the laser radar, and the spatial coordinates of the detected target in the current frame are used as the observation input: .

[0044] wherein: , is the observation noise, which models the uncertainty of the detection result.

[0045] Whenever a new frame of point cloud detection result is generated, the filter completes state prediction and update: state prediction: , .

[0046] wherein, is the prior state estimate at time k, which is predicted from the state at time k-1 and the motion model, is the posterior state estimate at time k-1, which is the optimal estimate incorporating all observations up to time k-1, is the prior error covariance matrix at time k, which represents the uncertainty of the predicted state, is the posterior error covariance matrix at time k-1, which represents the uncertainty of the optimal estimate at the previous time step, is the process noise covariance matrix, which represents the inaccuracy of the state transition model; F k T is the transpose of the state transition matrix F k .

[0047] Observation update: .

[0048] .

[0049] .

[0050] where, is the Kalman gain at time k, is the observation matrix, H k T is the transpose of the observation matrix , which maps the six-dimensional state space to the three-dimensional observation space, is the observation noise covariance matrix, which represents the measurement error and uncertainty of the LiDAR, is the posterior state estimate at time k, which is the optimal, final state output after the observation update, is the posterior error covariance matrix at time k, which represents the uncertainty of the final optimal estimate.

[0051] After completing the EKF state update for each frame, the latest estimated target position is converted to the yaw angle Yaw and the pitch angle Pitch relative to the gimbal origin, which are used to drive the gimbal to real-time turn: Yaw angle calculation: .

[0052] Pitch angle calculation: .

[0053] The application also discloses a computer device, which comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to realize the steps of the method.

[0054] The application also discloses a computer readable storage medium, which stores computer programs / instructions, and the computer programs / instructions are executed by a processor to realize steps of the method.

[0055] The application also discloses a computer program product, which comprises computer programs / instructions, and the computer programs / instructions are executed by a processor to realize steps of the method.

[0056] Advantages: Compared with the prior art, the application has the following advantages: 1. Laser radar point cloud collection and preprocessing: the application designs a point cloud preprocessing strategy suitable for small air targets according to the characteristics of small radar scattering cross section and weak reflection signal of low, slow and small targets, and improves the target point cloud saliency in a low signal-to-noise ratio background. The retention ability of weak echoes is enhanced in the data collection and filtering stage, and background interference points such as ground clutter, trees and buildings are suppressed, so that clearer and more stable point cloud input is provided for the subsequent recognition model.

[0057] 2. Deep learning model design: the application optimizes the deep learning network structure, adjusts the mainstream three-dimensional detection model based on ground targets (such as cars and pedestrians) to a structure suitable for low, slow and small flying targets in the air, and specially designs the model in terms of volume size adaptation, feature extraction level and detection frame size matching, so as to better adapt to the characteristics of small target size, irregular shape and sparse point cloud, thereby improving the detection accuracy and robustness.

[0058] 3. Extended Kalman filter tracking and prediction: considering the characteristics of variable motion trajectory and unstable attitude of low, slow and small targets, the application introduces an extended Kalman filter algorithm to estimate the state and predict the trajectory of the detection result. The method combines multi-frame point cloud data of the laser radar and the target dynamic model to update the state variables such as the spatial position, velocity and acceleration of the target in real time, effectively solves the problems of short-time occlusion, detection jitter and temporary missed detection, and improves the tracking continuity and prediction accuracy.

[0059] 4. The application constructs a laser radar perception and tracking system for low, slow and small flying targets, has good target detection accuracy, tracking stability and environmental adaptability, and can provide effective technical support for low-altitude safety monitoring, unmanned aerial vehicle control and other application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The flowchart of the application. DETAILED DESCRIPTION

[0061] The technical solutions of the application are further described below with reference to the drawings.

[0062] The present application is directed to the characteristics of low, slow and small flying targets in point cloud data, combined with the data input requirements of the PointPillars model, a point cloud preprocessing framework including multiple sub-modules such as high-frequency laser radar collection, data space regularization, feature enhancement, background removal and voxelization construction is designed, so as to realize efficient extraction and structure expression of low, slow and small targets at the data input layer. Figure 1 As shown in the technical scheme of the present application, the main steps include the following: Step 1: Laser radar point cloud collection and distance perception preprocessing.

[0063] 1.1 Distance perception point cloud region modeling: The laser radar device used in the present application is fixedly installed on a two-degree-of-freedom holder with pitch and rotation capability, providing dynamic perception capability for subsequent space modeling and target positioning. In order to ensure accurate interpretation of the three-dimensional structure of the point cloud by the system, the measurement of the installation position of the radar and the configuration of the space parameters need to be completed at the initial stage of device deployment. The specific operation is as follows: First, use a tape measure and other measuring tools to accurately measure the vertical height from the installation reference surface (such as the platform ground) to the bottom of the sensor, denoted as installation height , in meters. At the same time, with the help of a level and a protractor, record the pitch angle (Pitch Angle) , horizontal rotation deflection (Yaw) , and horizontal position offset relative to the origin of the carrier coordinate system .

[0064] After completing the above measurements, input the above data into the calibration module to construct the extrinsic matrix of the radar sensor in the platform coordinate system . This matrix is stored in the form of four-dimensional homogeneous coordinate transformation and written into the system configuration file for real-time calling during runtime. In the point cloud collection process, each frame of original point cloud data is represented in local coordinates as: .

[0065] In order to convert it to the absolute coordinate system, the system reads the extrinsic matrix during data preprocessing and performs coordinate transformation on each point: .

[0066] After transformation, the position of the point in the global space of the platform is obtained , which is used for subsequent height calculation and space screening.

[0067] In order to further identify the potential point cloud features of low-altitude flying targets, the vertical height information of the point relative to the ground is calculated: .

[0068] The height will be one of the input dimensions for the subsequent saliency score and classification decision, used to determine whether the target is likely to belong to a low, slow and small aircraft. At the same time, the spatial distance of each point (i.e. its three-dimensional Euclidean distance to the radar center) is calculated: The threshold value is set according to experience , and the point cloud is divided into three spatial regions: .

[0069] This area information will run through the subsequent saliency screening and pillar construction process, as an important basis for distance-aware adaptive processing.

[0070] 1.2 Adaptive saliency screening mechanism: In actual tasks, the point cloud data collected by the laser radar is mixed with a large number of non-target points, such as ground, building edges and background vegetation, etc. Especially in the long distance area, there are only a few sparse points of target points, which are easily mistaken or submerged. Therefore, it is particularly important to reasonably screen the point cloud before voxel encoding.

[0071] Therefore, the present application designs a saliency scoring mechanism that combines "reflection characteristics" and "local geometric structure", aiming to preferentially retain key points that may belong to target objects, especially giving additional retention opportunities to sparse points of small targets at a distance.

[0072] (1) Reflection intensity normalization: After each frame of point cloud is collected, the system first scans the reflection intensity values of all points, extracts the maximum value and the minimum value of the frame. Then the reflection intensity of each point is normalized: .

[0073] This processing unifies the intensity range difference between different frames, enhancing the stability of the scoring mechanism.

[0074] (2) Local density estimation: Count the number of neighborhood points in the three-dimensional neighborhood (sphere radius ) of each point. KD-Tree fast neighborhood search is used to achieve high computational efficiency, suitable for large-scale point cloud:

[0075] For easy comparison between different points, the number of neighborhood points is normalized: .

[0076] The density value reflects the geometric structure complexity of the local point, and high-density points usually belong to clear structure targets or their boundary regions.

[0077] (3) saliency score and adaptive enhancement mechanism: after the normalization of the above two dimensions, a saliency score is constructed for each point: .

[0078] wherein, In the system configuration, the parameter group can be specified by the user, for example, the default value is .

[0079] In order to improve the retention ability of long-distance sparse targets, the present application introduces an "adaptive weight adjustment" mechanism: when a point belongs to the "Far Field" region, the system will automatically increase the weight (pay more attention to the intensity of reflection), while reducing the saliency threshold , for example: .

[0080] wherein and Through the optimization of measured data, the system dynamically adjusts the parameters according to the distance distribution of the detected target, ensuring that the long-distance sparse points are more easily screened. Finally, according to the threshold, it is judged whether to retain the point to enter the subsequent pillar construction: .

[0081] 1.3 Distance partition driven Pillar construction and encoding strategy: In the task of laser radar three-dimensional target detection, the PointPillars encoding structure effectively improves the feature processing efficiency by mapping the spatial point cloud to a two-dimensional Bird's Eye View (BEV) plane. However, when facing low, slow and small flying targets, especially when they appear in the far distance area and the point cloud is sparse, the traditional uniform size voxel division strategy often has the following problems: Far distance area: sparse point distribution, insufficient points in voxel, difficult to form a column.

[0082] Near distance area: point number is dense but pillar resolution is too low, resulting in loss of boundary details.

[0083] Uniform parameter strategy: unable to balance the feature expression needs of targets at different distances.

[0084] To solve these problems, the present application proposes an adaptive pillar construction mechanism based on distance perception, the core idea of which is to "dynamically set pillar structure parameters according to distance regions" to enhance the universality and effectiveness of pillar generation.

[0085] (1) Region classification and parameter loading: this module automatically reads the preset pillar parameter table in the configuration file at system startup, as shown in Table 1.

[0086] .

[0087] These parameter values are optimized based on a large number of flight measured point clouds, and are set according to the point number characteristics and structure distribution characteristics of the target at different distances.

[0088] In each frame processing, the system assigns a region label (near, middle, far) to each point according to the previously calculated value (i.e. the spatial distance of the point), and uses different pillar division parameters accordingly.

[0089] (2) Pillar generation operation process: this module uses the grid indexing method for pillar division, divides the BEV space according to the grid in the selected region; each grid point corresponds to a pillar, and the filtered and retained point cloud is projected to the BEV grid to mark the pillar number, if the number of points in a pillar is greater than or equal to the corresponding number of points in the region, it is considered as a legal pillar, which enters the next stage of coding, and the remaining point cloud is discarded or cached for global optimization.

[0090] (3) Feature coding mode: for each point in the legal pillar , the system calculates the following local features in turn: the offset of the point and the centroid in the pillar: .

[0091] the offset of the point and the center grid position of the pillar: .

[0092] wherein is the coordinate of the center of the pillar in the BEV plane.

[0093] Complete point feature vector splicing: .

[0094] Feature coding: each point feature is up-coded by a shared MLP, and then maximum pooling is performed in the pillar dimension to form a fixed-length pillar expression vector.

[0095] Finally, all pillar coding results form a sparse tensor input, which is input to the backbone network for spatial perception and target detection.

[0096] Step 2: Design of detection network based on multi-scale structure and enhancement module mechanism: based on the foregoing sparse BEV coding result, the present application constructs a multi-scale perception backbone network, and embeds a spatial enhancement mechanism and a channel attention mechanism at a key level to significantly improve the robustness and sensitivity of detection.

[0097] 2.1 Multi-scale pyramid backbone network: In the point cloud detection scene, the volume scale of low, slow and small targets changes significantly at different distances: in the near distance area, the voxel is dense, the target features are clear, and the edges and geometric shapes are easy to capture; while in the long distance area, the point number is sparse, the target contour is blurred, the feature expression is limited, and it is easy to be covered by the background.

[0098] To adapt to this distribution characteristic, the BEV backbone network constructed by the application adopts a pyramid structure to expand the feature extraction path, which is constructed by multiple down-sampling levels in turn, and different scale feature layers have different receptive fields:

[0099] The first layer maintains the original BEV size, focusing on preserving the spatial details of small targets; the second layer expands the receptive field by 2 times down-sampling, capturing medium-scale structures; the third layer further down-samples to obtain large-scale background and context information.

[0100] Each layer is constructed by a combination of convolutional layers, and the number of channels increases step by step from bottom to top, which is used to enhance the semantic expression ability. The feature maps output by each layer will be cascaded or up-sampled in the subsequent fusion module to realize information backflow and take into account multi-scale structure representation.

[0101] 2.2 Spatial attention and edge-guided enhancement mechanism: When processing sparse point cloud features, the target itself has weak structure information, so a spatial attention mechanism needs to be introduced to highlight the potential target area and suppress background interference. The application designs a spatial enhancement sub-module to apply spatial weight adjustment to the BEV feature map, thereby dynamically enhancing the structure-sensitive area.

[0102] In the shallow and middle layer feature output of the backbone network, local convolution response is introduced for self-attention estimation, and the significance weight of each spatial position on the feature map is calculated And the significance weight is calculated as And the significance weight is calculated as .

[0103] Wherein, represents the original BEV feature map, is the enhanced result. This mechanism is particularly obvious for sparse feature response in the long distance area, and can highlight the contour shape of weak targets.

[0104] In addition, considering that small targets are often distributed in the edge position of the background, a boundary response map is constructed using local gradient information in the BEV plane to guide the feature to focus on the edge details, thereby improving the retention ability of micro-structures.

[0105] 2.3 Feature fusion and resolution backflow strategy: To solve the problem of semantic inconsistency between multi-scale features, a top-down resolution recovery mechanism is adopted to gradually up-sample high-level semantic features and splice and fuse them with low-level detailed features.

[0106] The fusion strategy can be expressed in the following form: .

[0107] wherein, is the current scale feature map, is the upper level (low resolution) feature map, and after splicing, the feature dimension is integrated by convolution to generate a fusion result with uniform channel number. Through step-by-step upward fusion, the system obtains multi-scale feature expression with details and context before the final detection head input.

[0108] Step 3: Anchor design and matching strategy based on aerial weak targets: In the point cloud target detection of BEV representation, the anchor mechanism is the basis for establishing the association between prediction and real target, and its matching quality directly affects the regression accuracy and positive and negative sample balance of the subsequent detection network. Traditional anchor design is mostly based on fixed size and biased towards large target vehicle / pedestrian task, and for low, slow and small aerial targets, there are problems such as low matching rate, insufficient positive samples and large positioning error.

[0109] The present application faces the characteristics of aerial flight micro-targets, redesigns the anchor size system, and introduces distance partition adjustment strategy and IoU dynamic lower limit adjustment mechanism in the matching mechanism, to comprehensively improve the anchor matching rate and training sample quality of weak targets.

[0110] 3.1 Anchor geometric parameter design and pre-configuration: To cover the shape features of various scale flight targets (such as micro unmanned aerial vehicles, remote control gliders, etc.), the present application loads multiple preset anchor templates through parameter files in the initialization stage of the detection network. Each Anchor contains three-dimensional size, center point height and direction angle information. The anchor template style is shown in Table 2.

[0111] .

[0112] In the anchor design process, according to the actual envelope size of aerial weak targets, the mean value is obtained by fitting the 3D bounding box of multiple frames of measured data. At the same time, multiple orientation angles are configured for each type of Anchor to improve the matching flexibility of multi-direction flight targets.

[0113] 3.2 Anchor matching mechanism and IoU lower limit adjustment: To improve the efficiency of Anchor matching, the system uses a positive and negative sample division mechanism based on 3D IoU. The volume intersection over union (IoU) is calculated for each Anchor prediction box and the real target box (Ground Truth, GT): .

[0114] The traditional matching mechanism often sets a fixed IoU threshold (such as a positive sample threshold of 0.6 and a negative sample threshold of 0.3), but this performs poorly in the far distance target sparse point scene. Due to the weak target point cloud envelope blur and the easy shift of the anchor box and GT center, a large number of effective Anchors are misjudged as negative samples. Therefore, the present invention proposes a dynamic adjustment strategy for the IoU lower limit in the distance interval, as follows: .

[0115] wherein, is the spatial distance from the Anchor center to the radar, , is the partition threshold set in the system. This mechanism makes it easier for far distance sparse targets to match to the appropriate Anchor, increasing the number of positive samples in the training process and reducing the detection blind area.

[0116] 3.3 Air target center height alignment strategy: Low, slow, and small targets are generally higher than ground static targets in flight state. To avoid detection bias caused by mismatch between Anchor height and target center, the system adjusts the center alignment of each Anchor's in the matching stage: for GT center height , the Anchor closest to it is selected as the preferred one. If the height difference between the matching Anchor and the GT center exceeds the threshold (such as ±0.3m), it is not matched and is considered as a structure that does not fit.

[0117] This height alignment strategy can significantly improve the matching stability of Anchors in the z-axis direction and has an important contribution to three-dimensional detection accuracy.

[0118] 3.4 Anchor balance and training sample construction: To prevent class imbalance, the present invention uses the following strategy when constructing positive and negative samples: for each GT box, at least one Anchor is guaranteed to match successfully; for Anchors with an overlap rate close to the threshold, the one closest to the center is retained; and the ratio of positive and negative samples in each frame is controlled within a reasonable range (such as 1:3) to balance the loss calculation.

[0119] After construction, all positive samples will participate in the regression loss and classification loss calculation, and the network parameters will be optimized through error backpropagation.

[0120] ​

[0121] Step 4: Low, slow, and small target trajectory association and extended Kalman filter tracking method based on laser radar point cloud: The system uses an extended Kalman filter (EKF) to predict and update the target state in time sequence, and uses the estimated target three-dimensional space position to drive the two-degree-of-freedom cloud turning, forming a "detection-prediction-turning" closed loop, ensuring the continuous tracking and stable pointing of the unmanned aerial vehicle target.

[0122] 4.1 State modeling and target representation: For the spatial flight characteristics of low, slow, and small targets, the system uses a six-dimensional state vector to model the trajectory: .

[0123] Among them, the first three dimensions represent the three-dimensional space position of the target in the current frame, and the last three dimensions represent the velocity components of the target. The state vector can accurately express the dynamic behavior of the target in the radar coordinate system, supporting the prediction of the target motion trend.

[0124] In the system initialization phase, when the detection module continuously and stably detects a potential target for multiple frames, the system will establish the EKF trajectory object corresponding to the target and initialize its state vector and covariance matrix.

[0125] 4.2 State transition and observation model: The invention assumes that the target satisfies the uniform straight flight model in a short time, and establishes a state prediction model: .

[0126] Among them, is the state transition matrix, defined as: .

[0127] Process noise obeys a zero-mean Gaussian distribution, which is used to absorb nonlinear disturbances or control errors outside the model.

[0128] At the same time, the observation model is constructed based on the laser radar detection results, and the spatial coordinates of the detected target in the current frame are used as the observation input: .

[0129] Among them: .

[0130] is the observation noise, which models the uncertainty of the detection results.

[0131] 4.3 Filter running process and trajectory prediction update: During the system running process, whenever a new frame of point cloud detection results is generated, the filter completes the state prediction and update according to the following steps.

[0132] (1) State prediction: , .

[0133] (2) Observation update: .

[0134] .

[0135] .

[0136] Through the above prediction and update process, the system can realize continuous estimation of the target three-dimensional space position and velocity, and alleviate the unstable trajectory problem caused by short-time detection missing frames, false detection or occlusion.

[0137] 4.4 PTZ control and target tracking driving mechanism: After the system completes the EKF state update in each frame, the latest estimated target position is converted into the yaw angle (Yaw) and pitch angle (Pitch) relative to the PTZ origin, which is used to drive the PTZ to turn in real time.

[0138] Yaw angle calculation: .

[0139] Pitch angle calculation: .

[0140] The system converts the above angle values into PWM signals or motor control instructions, transmits them to the control board (such as STM32 or Jetson platform), and drives two servos to control the horizontal and vertical rotation of the PTZ, so that the laser radar is directed towards the target position, maintaining a continuous tracking state.

[0141] When the target moves near the edge of the field of view or appears to be short-time occluded, the system can still perform PTZ turning according to the EKF prediction value to maintain the tracking direction, improving the anti-interference and responsiveness.

[0142] Embodiment

[0143] First, use the RayGod intelligent laser radar to receive raw data at a rotation scanning frequency of 10Hz. Each frame of data contains tens of thousands of three-dimensional point clouds, and the information of each point is a data vector According to the calibrated extrinsic matrix , the local coordinates of each point are converted to the global coordinate system to obtain the position of the point in the global space , the spatial distance of each point is calculated , the maximum / minimum reflection intensity of the frame is extracted by scanning the entire frame of point clouds, and is normalized to , the density of local points is calculated by using KD-Tree to count the number of neighborhood points, the significance score is calculated, and whether to retain the point into the subsequent pillar construction is judged according to the threshold; the screened points are projected to the BEV plane, the region where the target point is located is detected according to the adaptive mechanism, and the corresponding grid parameters are loaded, the offset of the point from the center of mass of all points in the Pillar and the offset of the grid center position are calculated, and the original coordinates and the reflection intensity and the two offsets are spliced into a 6-dimensional feature vector ; the feature vectors of all points in the Pillar are input into a shared MLP, each point is upgraded to 128 dimensions, and Max Pooling is performed, the maximum value on 128 channels is taken out, and the processed Pillar features are scattered back to the BEV grid according to their spatial positions, forming a pseudo-image of HxWx128; the pseudo-image is input into a pyramid CNN, and a multi-scale feature map is output, the high-level features are upsampled and spliced with the bottom-level features to obtain the final feature map fused with details and semantics; preset anchors are used on the final feature map for classification and regression to determine whether each Anchor frame is "background" or "low, slow and small target"; and the first three-dimensional data of the detected small target information is input into Kalman filtering; the Kalman gain is calculated according to the observation covariance matrix and the observation noise , the predicted value is corrected by state updating, and the optimal estimation of trajectory prediction is output ; finally, the predicted target position is converted into an azimuth angle and a pitch angle relative to the gimbal origin, and the motor drives the gimbal to realize real-time tracking of the low, slow and small target.

Claims

1. A laser radar-based low, slow and small target detection and trajectory prediction tracking method, characterized in that, Comprise the following steps: Step 1, collect laser radar point cloud data, and carry out spatial modeling and coordinate transformation preprocessing on the laser radar point cloud data; Step 2, significant screening is carried out on the preprocessed data, and the screened point cloud data is obtained; Step 3, based on distance partition driving, the screened point cloud data is subjected to Pillar construction and coding, and the pillar expression vector is obtained; Step 4, a deep learning detection network is constructed, and the pillar expression vector is taken as the input, and the fused features are output; Step 4 is specifically: in the shallow and middle layer feature output of the backbone network, local convolution response is introduced for self-attention estimation, and each spatial position on the feature map The significance weight thereof is calculated and is used as a scale factor to reweight the original feature map: ， wherein, denotes the original BEV feature map, is the enhanced result; A top-down resolution recovery mechanism is adopted to gradually up-sample high-level semantic features and splice and fuse with low-level detail features; The fusion strategy is expressed in the following form: ， wherein, is the current scale feature map, is the upper level feature map, and after splicing, the feature dimension is integrated through convolution to generate a fusion result with a unified channel number; Step 5, based on the anchor design and matching strategy, the size of the weak target in the air is adapted in the fused features; Step 6, based on the extended Kalman filter EKF, the target state is time series predicted and observed updated, and the low, slow and small target detection and trajectory prediction tracking are completed.

2. The low, slow and small target detection and trajectory prediction tracking method based on laser radar according to claim 1, characterized in that, Step 1 specifically: measure the vertical height from the installation reference surface to the bottom of the sensor , record the initial installation pitch angle of the radar , horizontal rotation deflection , and the horizontal position offset of the radar relative to the origin of the carrier coordinate system , input the data into the calibration module, and construct the external parameter matrix of the radar sensor in the platform coordinate system ; during the point cloud collection process, each frame of original point cloud data is expressed in the form of local coordinates: ， wherein, is the original data of each frame of point cloud, , , is the horizontal coordinate, vertical coordinate and height information of each point cloud in the local coordinate system, is the reflection intensity of each point cloud, and N is the number of point clouds in the original point cloud data of each frame. Reading the external parameters and performing coordinate transformation for each point: ， transformed to get the position of the point in the global space of the platform ; The vertical height information of the point relative to the ground is calculated: ， The spatial distance of each point is calculated: ， Thresholds set empirically , the point cloud is divided into three spatial regions: ， wherein, is a spatial region of the point cloud, are empirically set threshold limits for the near, mid, and far distances of the point cloud.

3. The low, slow and small target detection and trajectory prediction tracking method based on laser radar according to claim 1, characterized in that, Step 2 is specifically: After each frame of point cloud acquisition is completed, the reflection intensity values ​​of all points are scanned, and the maximum value of that frame is extracted. and minimum value The reflection intensity at each point Normalize: ， KD-Tree neighborhood search is used to count the number of neighborhood points within a three-dimensional neighborhood of each point : ， wherein, is the current point for which saliency is being computed, is the point cloud is any other point in the point cloud; The number of neighborhood points is normalized: ， wherein, is the normalized local density value, is the maximum value of the local density values of all points in the current frame point cloud; After normalization in both dimensions, a saliency score S is constructed for each point i : ， wherein, ; introduce adaptive weight adjustment mechanism: at runtime, when a point belongs to Far Field region, up-regulate weight, while lowering saliency threshold ; ， wherein, and denotes a dynamic base threshold coefficient, T s far denotes a significance threshold that the point cloud is in the Far Field region; Finally, according to the threshold value, it is judged whether the point is retained to enter the subsequent pillar construction: 。 4. The low, slow and small target detection and trajectory prediction tracking method based on laser radar according to claim 1, characterized in that, Step 3 is specifically: reading the pillar parameter table, using the grid index method to divide the pillars, and dividing the pillars in the selected area according to the pillar parameter table Grid division of the BEV space; Each grid point corresponds to a pillar column, and the screened and retained point cloud is projected to the BEV grid, and the corresponding pillar number is marked. If the number of points in a certain pillar is greater than or equal to the corresponding number of points in the region , it is regarded as a legal pillar, and the next stage of coding is entered, and the remaining point cloud is removed. For each point within a legal pillar , the offset of the point to the centroid of the pillar and the offset of the point to the center grid location of the pillar are calculated, and then the feature vectors are spliced, each point feature is dimensionally coded through a shared MLP, and then maximum pooling is performed in the pillar dimension to form a fixed-length pillar expression vector.

5. The low, slow and small target detection and trajectory prediction tracking method based on laser radar according to claim 1, characterized in that, Step 5 is specifically: a plurality of preset anchor box templates are loaded through a parameter file, each Anchor contains three-dimensional size, center point height and direction angle information; According to the actual envelope size of the weak target in the air, the mean value is obtained by fitting the 3D bounding box of multiple measured data; A positive and negative sample division mechanism based on 3D IoU is adopted to calculate the volume intersection ratio between each Anchor prediction box and the real target box GT; The IoU lower limit based on distance interval is dynamically adjusted, which is as follows: ， wherein, is the spatial distance from the Anchor center to the radar, , is the partition threshold set in the system; and the center height of each Anchor is adjusted in alignment.

6. The low, slow and small target detection and trajectory prediction tracking method based on laser radar according to claim 1, characterized in that, Step 6 is specifically: for the spatial flight characteristics of the low, slow and small target, a six-dimensional state vector is used to model its trajectory, the first three dimensions represent the three-dimensional spatial position of the target in the current frame, and the last three dimensions represent the velocity component of the target; Assuming that the target satisfies the uniform straight flight model in a short time, the state prediction model is established: ， where X k-1 is the posterior estimate of the state vector at the previous time step, is the state transition matrix, defined as: ， wherein, is a sampling period determined from the laser radar scan frequency; Process noise The observation model is constructed based on the detection results of the lidar, and the spatial coordinates of the detected targets in the current frame are taken as the observation input: ， Wherein: ， To observe noise, the uncertainty of the detection result is modeled; Every time a new frame of point cloud detection result is generated, the filter completes state prediction and update: State prediction: ， ， wherein, is the prior state estimate at time k, representing the result predicted from the state at time k-1 and the motion model, is the posterior state estimate at time k-1, representing the optimal estimate incorporating all observations up to time k-1, is the prior error covariance matrix at time k, representing the uncertainty of the predicted state, is the posterior error covariance matrix at time k-1, representing the uncertainty of the optimal estimate at the previous time, is the process noise covariance matrix, representing the inaccuracy of the state transition model; F k T is the state transition matrix F k is the transpose of the state transition matrix F Observation update: ， ， ， wherein, is the Kalman gain at time k, is the observation matrix, H k T is the observation matrix is the transpose of H, mapping the six-dimensional state space to the three-dimensional observation space, is the observation noise covariance matrix, representing the measurement error and uncertainty of the LiDAR, is the posterior state estimate at time k, which is the optimal, final state output after the observation update, is the posterior error covariance matrix at time k, representing the uncertainty of the final optimal estimate; After each frame completes the EKF state update, the latest estimated target position Convert to azimuth Yaw and pitch Pitch relative to the gimbal origin, used to drive the gimbal real-time steering: Azimuth angle calculation: ， Pitch angle calculation: 。 7. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, causes the processor to perform the method of any one of claims 1 to 6. The processor executes the computer program to realize the steps of the method of claim 1.

8. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to realize the steps of the method of claim 1.

9. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to realize the steps of the method of claim 1.

Images