A method and system for estimating the density of prairie dogs based on multiple sensors

Abstract

The application discloses a kind of grassland pika density estimation method and system based on multiple sensors, belong to grassland rodent control technical field, wherein the method includes: by unmanned aerial vehicle platform obtains multiple sensor data, carries out global map registration and obtains current pose data, then updates global point cloud map, current frame feature target is matched with target in global feature target library and global feature target library is maintained, current frame tracking target image coordinate is transformed into global coordinate, then according to the global position and historical position of current frame feature target, update the final global coordinate of feature target in global point cloud map, based on the global coordinate and quantity of feature target in global point cloud map, determine the density distribution of plateau pika in region by quantitative relationship estimation model.The application realizes the accurate positioning and tracking of plateau pika feature target, improves the automation degree and estimation accuracy of large-scale plateau pika density monitoring.

Description

Technical Field

[0001] This invention belongs to the field of grassland rodent control technology, specifically relating to a method and system for estimating grassland pika density based on multiple sensors. Background Technology

[0002] The plateau pika is one of the most widespread grassland rodent pests on the Qinghai-Tibet Plateau. The burrows they dig and the resulting bald patches have become important indicators for assessing the severity of their damage. Accurate and efficient monitoring of pika population density and distribution is crucial for early warning and precise control of grassland ecological security.

[0003] Early technologies primarily relied on traditional manual survey methods such as hole-blocking and mapping. While these methods could provide detailed plot information, they were time-consuming and labor-intensive, making large-scale monitoring difficult. Furthermore, data accuracy was affected by the subjective factors of the surveyors, failing to meet the demands for high-frequency, real-time monitoring. In addition, while satellite remote sensing technology possesses large-scale coverage capabilities, its spatial resolution limitations make it difficult to accurately identify details such as sub-meter-level mouse holes and bald patches, thus failing to meet the needs of refined monitoring.

[0004] In recent years, drone technology has been applied to grassland monitoring. Existing methods include rodent burrow identification based on drone aerial photography and machine learning, and rodent population analysis combining rodent monitoring vehicles and drone imagery. However, the former relies on digital elevation models (DEMs), limiting its applicability in unfamiliar areas without DEM data; the latter requires ground vehicles and is difficult to implement in areas with complex terrain. Furthermore, existing technologies mostly rely on single visible light sensors, lacking multi-source data fusion verification, making detection accuracy susceptible to lighting conditions, and failing to achieve precise global positioning and unique identification of characteristic targets, easily leading to duplicate counting or omissions. Therefore, there is an urgent need to develop a large-scale, high-precision plateau rodent density monitoring method applicable to any unfamiliar area to achieve early warning and precise control of grassland ecological security. Summary of the Invention

[0005] To address the aforementioned technical problems and the issues of high statistical cost and low statistical accuracy in existing technologies, this invention proposes a multi-sensor-based method for estimating grassland pika density, comprising the following steps: S1: Deploy a drone equipped with multiple sensors to a designated grassland, and fly according to preset parameters and routes to acquire multi-sensor data; S2: Perform global map registration based on the multi-sensor data obtained in S1 and obtain the sensor pose data at the current moment. Then, use the pose data to obtain and update the global point cloud map. S3: Based on the multi-sensor data obtained in S1, feature targets are identified through the target detection model. Then, the feature targets in the current frame are matched with targets in the global feature target library by combining target tracking and feature matching methods, and it is determined whether they are the same target. Finally, the global feature target library is maintained synchronously. S4: Calculate the global pose of the current frame image based on the multi-sensor data obtained in S1, transform the image coordinates of the target being tracked in the current frame into global coordinates, then update the historical position of the current frame and the feature target through extended Kalman filtering, and update the final global coordinates of the feature target in the global point cloud map; S5: Based on the global coordinates and quantity of feature targets in the global point cloud map obtained in S4, the density distribution of plateau pikas in the region is determined by a quantitative relationship estimation model.

[0006] Preferably, in S1, the multiple sensors include a camera, a lidar, and an inertial sensor, and the multiple sensor data includes image data of the current frame, point cloud data, and odometer data.

[0007] The preferred operation of S1 includes: S101: Based on the terrain features of the grassland to be monitored, set the drone's flight altitude and plan its flight path; S102: Control the drone to take off, and after reaching the designated altitude, start collecting multi-sensor data of the grassland to be monitored. The drone flies along the preset route. S103: The acquired current frame image data, point cloud data, and odometer data are marked with timestamps and stored respectively.

[0008] The preferred S2 operation includes: S201: Obtain the odometry data during the generation period of the original point cloud of the current frame, and perform motion distortion correction operation on the original point cloud of the current frame; S202: Based on the odometry data between the previous frame point cloud and the current frame point cloud, perform LiDAR pose pre-estimation, and then register the current frame point cloud and the global point cloud map based on the pre-estimation results to obtain the LiDAR pose data at the current moment. S203: Based on the current lidar pose data, transform the current frame point cloud to the global coordinate system of the global point cloud map, add the transformed point cloud to the global point cloud map, and update the global point cloud map.

[0009] The preferred option, S3, specifically includes: S301: Acquire the image data of the current frame and combine it with the camera distortion correction parameters to complete the distortion correction of the current frame image; S302: Using a target detection model, identify feature targets such as mouse holes and bald spots in the current frame image and obtain the location of the feature targets in the image; S303: Using the target tracking model, calculate the feature vector of each feature target in the current frame image, and compare the feature vector with the corresponding feature vector of the global feature targets: When the similarity of the feature vectors is greater than or equal to the preset similarity value, the feature target in the current frame and the global feature target being compared are determined to be the same target, and the global number of the global feature target is used. When the similarity of the feature vectors is less than the preset similarity value, the feature target of the current frame is determined to be a new global feature target and a new global number is assigned.

[0010] The preferred option, S4, specifically includes: S401: Acquire the LiDAR pose data closest to the current frame image time, the odometry data between the LiDAR pose time and the current frame image time, and the external parameters between the LiDAR and the camera. Integrate the odometry data and the LiDAR pose data, and combine the external parameter transformation to obtain the current frame image pose data. S402: Combine the pose data of the current frame image with the position data of the feature target in the current frame image to calculate the global coordinates of the feature target in the current frame; S403: Combine the global position of the feature target with the global coordinates corresponding to the current frame, and update the global position of the feature target by extended Kalman filtering.

[0011] The superior S5 specifically includes: S501: Based on the coordinate and quantity information of the feature targets in the global point cloud map, a grid is drawn with the location of each feature target as the center, and the number of points in the grid is calculated. S502: Based on the number of feature targets within the grid, according to Calculate the number of valid openings within the grid, where Indicates the number of valid openings. This represents the conversion coefficient between the number of effective openings and the number of characteristic targets. Indicates the number of feature targets within a grid; where The results were obtained by setting up quadrats in the monitoring area and counting the number of effective openings and characteristic targets; S503: Based on the number of effective openings within the grid, according to Estimate the pika population size within the grid, of which Indicates the population size of pikas within the grid. Indicates the conversion coefficient between the burrow entrance and the pika; S504: Generate a continuous density raster map based on the number of pikas in each grid cell to obtain a density distribution map of plateau pikas in the monitoring area.

[0012] To implement the above method, the present invention also provides a multi-sensor-based grassland pika density estimation system, comprising the following modules: Drone detection module: Drones equipped with cameras, lidar, and inertial sensors fly according to preset parameters and routes to acquire image data, point cloud data, and odometry data of the area to be monitored; Global point cloud map module: Perform motion compensation based on the obtained multi-sensor data, perform global map registration on the motion-compensated multi-sensor data and obtain the sensor pose data at the current moment, and use the pose data to obtain and update the global point cloud map. Feature target recognition module: Identifies feature targets through a target detection model, combines target tracking and feature matching methods to match the feature targets in the current frame with the feature targets in the global feature target library, and determines whether they are the same feature target. Finally, it synchronously maintains the global feature target library. Global position determination module: Calculates the global pose of the current frame image based on multi-sensor data, transforms the image coordinates of the tracked target in the current frame into global coordinates, updates the historical position of the current frame and the feature target through extended Kalman filtering, and updates the final global coordinates of the feature target in the global point cloud map; Density distribution estimation module: Based on the global coordinates and quantity of feature targets in the obtained global point cloud map, the density distribution of plateau pikas in the region is determined by a quantitative relationship estimation model.

[0013] Compared with the prior art, the technical solution of the present invention has the following advantages: 1. This invention utilizes a multi-sensor fusion approach, combining the high-precision ranging capability of lidar with the rich semantic information of cameras, to achieve accurate localization and tracking of plateau pika characteristic targets, thereby improving the automation level and estimation accuracy of large-scale plateau pika density monitoring.

[0014] 2. This invention combines target detection and tracking to achieve continuous identification and globally unique identification of feature targets, avoids duplicate counting, ensures the accuracy of mouse and rabbit density estimation, and reduces statistical overhead.

[0015] 3. This invention integrates spatiotemporal alignment and extended Kalman filtering, making full use of historical observation information to optimize the current estimate, thereby improving the accuracy and robustness of global localization of feature targets. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a multi-sensor-based method for estimating grassland pika density. Figure 2 Flowchart of the method for constructing a global point cloud map; Figure 3 Here is a flowchart of the target detection and tracking method; Figure 4 This is a schematic diagram illustrating the alignment of timestamps from multiple sensors. Figure 5 This is a schematic diagram of the structure of a multi-sensor-based grassland pika density estimation system. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention are described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention. Therefore, the detailed description of the embodiments of this invention provided below is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention.

[0019] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it need not be further defined and explained in subsequent figures.

[0020] Example 1

[0021] like Figure 1As shown, this embodiment is a multi-sensor-based method for estimating grassland pika density. The method includes the following steps: S1: Deploy a drone equipped with multiple sensors (camera, lidar, inertial sensor) to the grassland to be monitored. The drone flies according to preset parameters and a route to acquire multi-sensor data, including image data, point cloud data, and odometry data. Before conducting the monitoring task, the preset parameters and route need to be determined based on the terrain features, altitude, and vegetation cover of the grassland to be monitored. This includes planning the drone's flight altitude (usually set to 30-80 meters above the ground to ensure ground resolution and flight safety), flight speed (usually 3-5 m / s to balance data quality and monitoring efficiency), and flight path (usually a serpentine or meandering route to ensure full coverage of the monitoring area). After the drone takes off and reaches the designated altitude, the multi-sensor data acquisition program is initiated. During flight, the camera acquires image data at a frame rate of approximately 15 Hz, the lidar acquires point cloud data at a frame rate of approximately 10 Hz, and the inertial sensor acquires odometry data at a frame rate of approximately 200 Hz. Each sensor's data is individually timestamped during acquisition and stored in rosbag format for subsequent offline or real-time processing. The timestamp accuracy must be at the millisecond level to ensure spatiotemporal consistency during multi-sensor data fusion.

[0022] S2: Perform global map registration based on the multi-sensor data obtained in S1 and obtain the sensor pose data at the current moment. Then, use the pose data to obtain and update the global point cloud map.

[0023] The global point cloud map construction process in this embodiment is as follows: Figure 2 As shown, due to the high-frequency vibration and displacement of the UAV during flight, the coordinate system of the lidar undergoes significant changes within a single frame scan cycle (e.g., 100ms), resulting in point cloud motion distortion. Therefore, motion distortion correction is the first step. The system acquires inertial sensor odometry data during the generation period of the original point cloud in the current frame, performs motion distortion correction on each original point in the point cloud, and then obtains the lidar pose at the corresponding time by linear interpolation or spherical linear interpolation based on its precise timestamp. All points are corrected to the unified coordinate system at the end of the frame to obtain the motion-compensated point cloud.

[0024] Then, pose pre-estimation is performed: pre-integration is performed based on the inertial sensor odometry data between the point cloud of the previous frame and the point cloud of the current frame to obtain the relative pose transformation between the two frames, which is used as the initial pose estimate of the point cloud of the current frame.

[0025] Based on the initial pose estimation, the current frame point cloud is registered with the constructed global point cloud map (which is initially empty or a pre-built map). This embodiment uses the Iterative Closest Point (ICP) algorithm or its variants (such as GeneralizedICP) for registration. By iteratively solving the objective function that minimizes the distance between point clouds, the precise pose data of the lidar in the global coordinate system at the current moment (including 3D position and attitude angle) is obtained.

[0026] Based on the current pose data from the inertial sensor, the motion-compensated point cloud of the current frame is projected onto the global coordinate system of the global point cloud map through rigid body transformation. The transformed point cloud is then fused with the existing map, duplicate or noisy points are removed, and the global point cloud map is updated. M i The global point cloud map serves as the benchmark for subsequent global localization of feature targets.

[0027] S3: Based on the multi-sensor data obtained in S1, feature targets are identified through a target detection model. Then, the feature targets in the current frame are matched with targets in the global feature target library by combining target tracking and feature matching methods, and it is determined whether they are the same target. Finally, the global feature target library is maintained synchronously.

[0028] S4: Calculate the global pose of the current frame image based on the sensor data obtained in S1, transform the image coordinates of the target being tracked in the current frame into global coordinates, then update the historical position of the current frame and the feature target through extended Kalman filtering, and update the final global coordinates of the feature target in the global point cloud map.

[0029] S5: Based on the global coordinates and quantity of feature targets in the global point cloud map obtained in S4, the density distribution of plateau pikas in the region is determined by a quantitative relationship estimation model.

[0030] In S1, a drone platform equipped with multiple sensors is deployed over a designated grassland monitoring area. The drone flies according to set parameters and routes to acquire data from multiple sensors.

[0031] In S2, based on odometry data and current frame point cloud data, motion compensation is performed on the original point cloud data through motion estimation to obtain motion-compensated point cloud; the motion-compensated point cloud is registered with the global map to obtain the pose data of the lidar at the current moment; according to the lidar pose data, the motion-compensated point cloud is projected and transformed into the global point cloud map, and the global point cloud map is updated.

[0032] In S3, based on the current frame image data, the target detection model identifies feature targets such as mouse holes and bald spots, and obtains their positions in the image. Combining target tracking and feature matching methods, the feature targets in the current frame are matched with targets in the global feature target library to determine whether they are the same target, and the global numbering is updated and the global feature target library is maintained synchronously.

[0033] The target detection and tracking method process used in Example S3 is as follows: Figure 3 As shown, the specific process is as follows: First, obtain the image data of the current frame. C i By combining the pre-calibrated camera intrinsic parameters (including focal length and principal point coordinates) and distortion correction parameters (such as radial distortion coefficient and tangential distortion coefficient), the distortion correction of the current frame image is completed through the inverse distortion model, eliminating the influence of lens distortion on target position extraction.

[0034] Subsequently, the corrected image is input into the object detection model. In this embodiment, the object detection model is a one-stage YOLO object detection model trained on a self-constructed grassland mouse hole-bald spot dataset. The model is trained on a large number of grassland images labeled with mouse holes and bald spots as feature targets, and can output the category (mouse hole or bald spot), confidence score, and bounding box coordinates (i.e., image location) of the feature targets in the image in real time.

[0035] For the detected feature targets, the confidence threshold is first determined: set a confidence threshold (e.g., 0.7), retain only valid targets with a confidence level greater than or equal to the threshold, and filter out false positive targets with low confidence (e.g., falsely detected pits, stone shadows, etc.).

[0036] For the retained valid targets, a normalized feature vector for each target is calculated using a target tracking model. In this embodiment, a deep convolutional neural network (such as ResNet or a lightweight MobileNet) is used to extract the appearance features of the targets, and combined with the target's position information (center coordinates, width and height) in the image, a feature description database is constructed.

[0037] The feature vector of each target in the current frame is compared with the feature vectors of targets in the global feature target library, and the cosine similarity is calculated. When the feature vector similarity is greater than or equal to the preset similarity value (e.g., 80%) and is the only highest similarity, the target in the current frame is determined to be the same target as the compared global feature target, and the global number of the global feature target is retained; when the feature vector similarity is less than the preset similarity value, the target in the current frame is determined to be a new global feature target, and a new unique global number is assigned.

[0038] The global feature target library stores information such as the confirmed feature targets in historical frames, their global numbers, feature vectors, and historical trajectories.

[0039] After the matching is completed, the features of the target in the global feature target library are updated. For example, the feature vector is updated to a weighted fusion of the current frame and historical features. Then the image position of the current frame is recorded, and the global feature target library is maintained synchronously.

[0040] Methods for maintaining a global feature target library include removing invalid targets that have not been matched for a long time.

[0041] In S4, based on the extrinsic parameters of the lidar and camera, the lidar pose at the current moment, and the odometry data, the global pose of the current frame image is calculated, and the image coordinates of the target being tracked in the current frame are transformed to global coordinates. The observed value and the target's historical position are updated by extended Kalman filtering, and the final global position of the feature target in the global point cloud map is updated.

[0042] like Figure 4 The diagram shown illustrates the timestamp alignment of multiple sensors in this embodiment. Because the camera (15Hz) and LiDAR (10Hz) have different data frame rates and their acquisition times are offset, time synchronization is necessary. For the current frame image data... C i Its timestamp is t ci Find the lidar pose data with the closest timestamp (e.g.) L i timestamp t li )as well as t ci and t li The inertial sensor odometer data between them.

[0043] Based on the external parameters (rotation matrix) between the lidar and the camera R cl Translation vector t cl ), and the current lidar pose ( R wl , t wl ), calculate the pose of the current frame image ( R wc , t wc Specifically, through a coordinate transformation chain: world coordinate system → lidar coordinate system → camera coordinate system, combined with inertial sensor odometer... t ci arrive t li The integral results within the time period are used for pose correction to obtain the precise pose of the current frame image in the global coordinate system.

[0044] Combining the current frame image pose data and the pixel coordinates of the feature target in the current frame image ( u , v Using the camera's intrinsic parameters and the depth information obtained from the point cloud projection at the corresponding time, the global 3D coordinates of the feature target in the current frame are calculated. X , Y , Z ).

[0045] The global coordinates obtained from the current frame observation are used as the observation value. Combined with the historical position of the target in the global target database (as the state prediction), the state is updated using an extended Kalman filter. The state vector of the extended Kalman filter includes the target's three-dimensional position and velocity. The observation model is the coordinate observation of the current frame. The optimal estimate is calculated using Kalman gain to update the global position of the target, effectively suppressing single-frame observation noise and improving positioning accuracy and trajectory smoothness.

[0046] In S5, based on the global coordinates and number of mouse burrows and bald spots in the global point cloud map, the density distribution of plateau pikas in the region is determined by a quantitative relationship estimation model.

[0047] Specifically, after obtaining the precise global coordinates of all mouse burrows and bald patches within the monitoring area, spatial rasterization is performed. Using the location of each feature target as the center, a circular area with a certain radius (e.g., 5 meters) is determined based on the mouse's activity range as a grid unit, and the number of feature targets within each grid is counted. N 特征目标 .

[0048] Based on the number of feature targets within the grid, according to N 洞 = αN 特征目标 Calculate the number of valid openings within the grid. Among them, α The effective portal-feature target quantity conversion coefficient is used to correct potential missed or false detections in the detection model, and to distinguish between active and abandoned portals. α The number of effective openings and the number of feature targets detected by images / point clouds were manually counted in the field by randomly setting up several quadrats (such as 1m×1m or larger) in the monitoring area, and then calibrated by linear regression or statistical averaging methods.

[0049] Based on the number of effective openings within the grid, according to D = KN 洞 Estimate the pika population size within the grid. Among them, K The conversion coefficient between burrow entrance and pika (i.e., the average number of pika individuals corresponding to each active burrow entrance, usually 2-5, depending on the pika species and season).K The method of setting up quadrats in the monitoring area, capturing pikas using trapping or fencing methods, and recording the ratio of the actual number of pikas to the number of effective burrows is used to determine the number of pikas.

[0050] Based on the estimated pika population within each grid cell, a continuous density raster map is generated using spatial interpolation algorithms (such as inverse distance weighted interpolation (IDW) or Kriging interpolation). This yields a spatial distribution map of the plateau pika density within the monitoring area, expressed in pikas per hectare. The density distribution map can be used to identify high-incidence areas of rodent infestation, providing data support for early warning and precise control of grassland ecological security.

[0051] This embodiment is based on a multi-sensor method for estimating grassland pika density, which can be widely applied to the dynamic monitoring of pika populations in ecologically fragile areas such as alpine grasslands and alpine meadows. To facilitate further understanding of the above scheme, the UAV monitoring platform, consisting of a UAV and sensors, involved in this embodiment, will be further described. The UAV monitoring platform includes a UAV flight carrier and multiple sensors (hereinafter referred to as multi-sensors) mounted on it, as well as a multi-sensor data acquisition system mounted on the carrier. This data acquisition system utilizes commonly available data acquisition systems, which are usually provided by the supplier when purchasing sensors. The multi-sensors include: a visible light camera for acquiring high-resolution images of the grassland surface, a lidar for acquiring three-dimensional point cloud information of the surface, and an inertial sensor for acquiring high-frequency attitude and position changes of the carrier. The camera, lidar, and inertial sensor are fixed to the UAV gimbal via rigid connectors, and the external parameters (i.e., extrinsic parameters, including rotation matrices and translation vectors) between the sensors need to be accurately calibrated in advance using a joint calibration method. Simultaneously, the sensors need to achieve timestamp alignment through hardware or software synchronization to ensure the temporal consistency of multi-source data.

[0052] Example 2: Figure 5 As shown, this embodiment provides a grassland pika density estimation system for implementing the above method. The system includes the following modules: Drone detection module: A drone platform equipped with multiple sensors is deployed over a designated grassland monitoring area. The drone platform flies according to preset parameters and routes to acquire image data, point cloud data, and odometer data.

[0053] Global point cloud map module: Based on the obtained odometry data and the current frame point cloud data, motion compensation is performed on the original point cloud data through motion estimation to obtain motion-compensated point cloud data; the motion-compensated point cloud data is registered with the global map to obtain the pose data of the LiDAR at the current moment; the motion-compensated point cloud data is projected and transformed onto the global point cloud map based on the LiDAR pose data, and the global point cloud map is updated.

[0054] Feature target recognition module: Based on the image data of the current frame, the module identifies mouse holes and bald spots as feature targets through a target detection model and obtains the position of the feature targets in the image data; it combines target tracking and feature matching methods to match the feature targets of the current frame with targets in the global feature target library, then determines whether they are the same target, completes the global number allocation update, and finally synchronously maintains the global feature target library.

[0055] Global Position Determination Module: Based on the extrinsic parameters of the LiDAR and camera, the current LiDAR pose, and odometry data, the module calculates the global pose of the current frame image and transforms the image coordinates of the tracked target in the current frame to global coordinates. It then updates the observed value and the target's historical position using an extended Kalman filter, ultimately updating the final global position of the feature target in the global point cloud map. The specific process is as follows: First, it acquires the LiDAR pose data closest in time to the current frame image, the odometry data between the LiDAR pose time and the current frame image time, and the extrinsic parameters between the LiDAR and camera. It then integrates the odometry data and the LiDAR pose data, and combines this with the extrinsic parameter transformation to obtain the current frame image pose data.

[0056] Then, by combining the pose data of the current frame image and the position data of the feature target in the current frame image, the global coordinates of the feature target in the current frame are calculated.

[0057] Finally, by combining the global position of the feature target with the global coordinates corresponding to the current frame, the global position of the feature target is updated through extended Kalman filtering.

[0058] Density distribution estimation module: Based on the global coordinates and quantity of mouse burrows and bald patches in the global point cloud map, the density distribution of plateau pikas in the region is determined through a quantitative relationship estimation model. The specific process of density distribution integration is as follows: First, based on the coordinate and quantity information of the feature targets in the global point cloud map, a region with a preset radius is delineated as a grid with each feature target location as the center, and the number of points within the grid is calculated.

[0059] Then, based on the number of feature targets within the grid, according to Calculate the number of valid openings within the grid, where Indicates the number of valid openings. This represents the conversion coefficient between the number of effective openings and the number of characteristic targets. Indicates the number of feature targets within a grid; where The number of effective openings and characteristic targets were obtained by setting up quadrats in the monitoring area.

[0060] Then, based on the number of effective openings within the grid, according to Estimate the pika population size within the grid, of which Indicates the population size of pikas within the grid. This represents the conversion coefficient between the burrow entrance and the pika.

[0061] Finally, based on the number of pikas in each grid cell, a continuous density grid map is generated to obtain a density distribution map of plateau pikas in the monitoring area.

[0062] The above are merely preferred embodiments of the present invention. It should be noted that the above preferred embodiments should not be considered as limitations on the present invention, and the scope of protection of the present invention should be determined by the scope defined in the claims. For those skilled in the art, several improvements and modifications can be made without departing from the spirit and scope of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-sensor-based method for estimating grassland pika density, characterized in that, Includes the following steps: S1: Deploy a drone equipped with multiple sensors to the grassland to be monitored. The multiple sensors include cameras, lidar, and inertial sensors. The drone flies according to preset parameters and routes to acquire data from the multiple sensors. S2: Perform motion compensation on the multi-sensor data obtained in S1, and then perform global map registration on the motion-compensated multi-sensor data to obtain the pose data of the sensors at the current moment. Then, use the pose data to obtain and update the global point cloud map. S3: Based on the multi-sensor data obtained in S1, feature targets are identified through the target detection model. Then, the feature targets in the current frame are matched with the feature targets in the global feature target library by combining target tracking and feature matching methods, and it is determined whether they are the same feature targets. Finally, the global feature target library is maintained synchronously. S4: Calculate the global pose of the current frame image based on the multi-sensor data obtained in S1, transform the image coordinates of the target being tracked in the current frame into global coordinates, then update the historical position of the current frame and the feature target through extended Kalman filtering, and update the final global coordinates of the feature target in the global point cloud map; S5: Based on the global coordinates and quantity of feature targets in the global point cloud map obtained in S4, the density distribution of plateau pikas in the region is determined by a quantitative relationship estimation model.

2. The method for estimating grassland pika density based on multiple sensors according to claim 1, characterized in that, The multi-sensor data includes image data, point cloud data, and odometer data for the current frame.

3. The method for estimating grassland pika density based on multiple sensors according to claim 2, characterized in that, The process of acquiring multi-sensor data is as follows: S101: Based on the terrain features of the grassland to be monitored, set the drone's flight altitude and plan its flight path; S102: Control the drone to take off, reach the designated altitude, and fly along the preset route to start collecting multi-sensor data of the grassland to be monitored; S103: The acquired current frame image data, point cloud data, and odometer data are marked with timestamps and stored respectively.

4. The method for estimating grassland pika density based on multiple sensors according to claim 3, characterized in that, The specific process of S2 is as follows: S201: Obtain the odometry data during the generation period of the original point cloud of the current frame, and perform motion distortion correction operation on the original point cloud of the current frame; S202: Based on the odometry data between the previous frame point cloud and the current frame point cloud, the LiDAR pose is pre-estimated, and the current frame point cloud and global point cloud map are registered according to the pre-estimated results to obtain the LiDAR pose data at the current moment. S203: Based on the lidar pose data of the current frame, transform the point cloud of the current frame to the global coordinate system of the global point cloud map, add the transformed point cloud to the global point cloud map, and update the global point cloud map.

5. The method for estimating grassland pika density based on multiple sensors according to claim 4, characterized in that, The specific process of S3 is as follows: S301: Acquire the image data of the current frame and combine it with the camera distortion correction parameters to complete the distortion correction of the current frame image; S302: Using a target detection model, identify mouse holes and bald patches in the current frame image, and obtain the location of the target in the image; S303: Using the target tracking model, calculate the feature vector of each feature target in the current frame image, and compare the feature vector with the corresponding feature vector of the global feature targets: When the similarity of the feature vectors is greater than or equal to the preset similarity value, the feature target in the current frame and the global feature target being compared are determined to be the same target, and the global number of the global feature target is used. When the similarity of the feature vectors is less than the preset similarity value, the feature target of the current frame is determined to be a new global feature target and a new global number is assigned.

6. The method for estimating grassland pika density based on multiple sensors according to claim 1, characterized in that, The specific process of S4 is as follows: S401: Acquire the LiDAR pose data closest to the current frame image time, the odometry data between the LiDAR pose time and the current frame image time, and the external parameters between the LiDAR and the camera. Integrate the odometry data and the LiDAR pose data, and combine the external parameter transformation to obtain the current frame image pose data. S402: Combine the pose data of the current frame image with the position data of the feature target in the current frame image to calculate the global coordinates of the feature target in the current frame; S403: Combine the global position of the feature target with the global coordinates corresponding to the current frame, and update the global position of the feature target by extended Kalman filtering.

7. The method for estimating grassland pika density based on multiple sensors according to claim 1, characterized in that, The specific process of S5 is as follows: S501: Based on the coordinate and quantity information of the feature targets in the global point cloud map, a grid is drawn with the location of each feature target as the center, and the number of points in the grid is calculated. S502: Based on the number of feature targets within the grid, according to Calculate the number of valid openings within the grid, where Indicates the number of valid openings. This represents the conversion coefficient between the number of effective openings and the number of characteristic targets. Indicates the number of feature targets within a grid; where The results were obtained by setting up quadrats in the monitoring area and counting the number of effective openings and characteristic targets; S503: Based on the number of effective openings within the grid, according to Estimate the pika population size within the grid, of which Indicates the population size of pikas within the grid. Indicates the conversion coefficient between the burrow entrance and the pika; S504: Generate a continuous density raster map based on the number of pikas in each grid cell to obtain a density distribution map of plateau pikas in the monitoring area.

8. A multi-sensor-based grassland pika density estimation system, characterized in that, Includes the following modules: Drone detection module: Drones equipped with cameras, lidar, and inertial sensors fly according to preset parameters and routes to acquire image data, point cloud data, and odometry data of the area to be monitored; Global point cloud map module: Perform motion compensation based on the obtained multi-sensor data, perform global map registration on the motion-compensated multi-sensor data and obtain the sensor pose data at the current moment, and use the pose data to obtain and update the global point cloud map. Feature target recognition module: Identifies feature targets through a target detection model, combines target tracking and feature matching methods to match the feature targets in the current frame with the feature targets in the global feature target library, and determines whether they are the same feature target. Finally, it synchronously maintains the global feature target library. Global position determination module: Calculates the global pose of the current frame image based on multi-sensor data, transforms the image coordinates of the tracked target in the current frame into global coordinates, updates the historical position of the current frame and the feature target through extended Kalman filtering, and updates the final global coordinates of the feature target in the global point cloud map; Density distribution estimation module: Based on the global coordinates and quantity of feature targets in the obtained global point cloud map, the density distribution of plateau pikas in the region is determined by a quantitative relationship estimation model.

9. A multi-sensor-based grassland pika density estimation system according to claim 8, characterized in that, The global location determination module performs the global location determination process as follows: First, acquire the LiDAR pose data closest to the current frame image time, the odometry data between the LiDAR pose time and the current frame image time, and the external parameters between the LiDAR and the camera. Integrate the odometry data and the LiDAR pose data, and combine them with the external parameter transformation to obtain the current frame image pose data. Then, by combining the pose data of the current frame image and the position data of the feature target in the current frame image, the global coordinates of the feature target in the current frame are calculated. Finally, by combining the global position of the feature target with the global coordinates corresponding to the current frame, the global position of the feature target is updated through extended Kalman filtering.

10. A multi-sensor-based grassland pika density estimation system according to claim 9, characterized in that, The density distribution estimation module performs the following estimation process: First, based on the coordinate and quantity information of the feature targets in the global point cloud map, a grid is drawn with the location of each feature target as the center, and the number of points in the grid is calculated. Then, based on the number of feature targets within the grid, according to Calculate the number of valid openings within the grid, where Indicates the number of valid openings. This represents the conversion coefficient between the number of effective openings and the number of characteristic targets. Indicates the number of feature targets within a grid; where The results were obtained by setting up quadrats in the monitoring area and counting the number of effective openings and characteristic targets; Then, based on the number of effective openings within the grid, according to Estimate the pika population size within the grid, of which Indicates the population size of pikas within the grid. Indicates the conversion coefficient between the burrow entrance and the pika; Finally, based on the number of pikas in each grid cell, a continuous density grid map is generated to obtain a density distribution map of plateau pikas in the monitoring area.

Images