Indoor environment dynamic target detection method based on vehicle-mounted sensor system

Abstract

The invention discloses an indoor environment dynamic target detection method based on a vehicle-mounted sensor system. The method comprises the following steps: collecting laser radar data and millimeter wave radar data; performing pre-clustering on the preprocessed laser radar data by using a DBSCAN algorithm, updating an Euclidean distance calculation formula according to a point cloud distribution variance mean value in each direction, and performing improved DBSCAN clustering on the preprocessed laser radar data by using the updated Euclidean distance calculation formula to obtain a final clustering cluster set of laser radar targets; constructing an extended Kalman filter tracker to track the position of a laser radar target, and calculating the displacement, speed and motion direction information of the target; doppler velocity and azimuth angle information are extracted from millimeter wave radar data, mapping of the Doppler velocity in the vehicle movement direction is calculated, a velocity value distribution curve in the vehicle movement direction is calculated, and a velocity value at the peak value of the distribution curve is searched as the vehicle self-movement velocity; calculating the absolute movement speed of the target, and judging whether the target is a moving target. The algorithm overhead can be reduced, and the detection real-time performance is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic target state detection, and in particular to a method for detecting dynamic targets in an indoor environment based on a vehicle-mounted sensor system. Background Art

[0002] As an uncertain factor in the scene, moving obstacles are the key perception objects in the indoor intelligent driving process. In the SLAM mapping process of autonomous driving, the presence of dynamic targets will lead to inaccurate mapping. In view of the dilemma of poor mapping stability of existing vehicle-mounted sensor systems in dynamic indoor scenes, the present invention proposes a method for detecting dynamic targets in indoor environments based on a vehicle-mounted sensor system.

[0003] At present, the research on dynamic target detection based on vehicle-mounted sensor systems includes: Patent "A laser SLAM method integrating dynamic target estimation". In this invention, the point cloud target detection network PointRCNN is used to perform three-dimensional target detection on point cloud data, the laser odometer is used to estimate the motion of the unmanned vehicle based on the nearest point iteration method, and the target state estimation method is designed to divide the targets in the scene into dynamic and static.

[0004] The application scenario of this method is outdoor. In outdoor scenes, there are many feature points of LiDAR data, while indoors there are highly repetitive scenes such as corridors. If this method is used to process indoor scenes, the laser odometer based on the nearest point iteration method cannot perform accurate motion estimation. At the same time, the nearest point iteration algorithm has a huge amount of calculation and cannot meet the real-time requirements.

[0005] In view of the problems existing in the existing algorithms, the present invention has built an algorithm framework that uses laser radar to track targets to obtain the relative motion state of targets, and uses millimeter-wave radar to estimate vehicle self-motion. The algorithm can adapt well to high-similarity indoor scenes and achieve accurate moving target detection, and the extremely low performance overhead of the millimeter-wave radar self-motion compensation algorithm ensures the real-time performance of the system. At the same time, the present invention has adapted the algorithm to low-performance hardware, and can use low-beam laser radar to obtain accurate target relative position and speed information. Summary of the invention

[0006] The purpose of the present invention is to provide a method for detecting dynamic targets in an indoor environment based on a vehicle-mounted sensor system in order to solve the problems existing in the existing indoor environment dynamic target detection.

[0007] The technical solution to achieve the purpose of the present invention is: a method for detecting dynamic targets in an indoor environment based on a vehicle-mounted sensor system, comprising the following steps:

[0008] Step 1: Use the vehicle-mounted sensor system to collect data and obtain data packets, including lidar data and millimeter-wave radar data;

[0009] Step 2: extract the lidar data in the data packet and preprocess the lidar data, including: ROI area extraction and point cloud downsampling;

[0010] Step 3: Use the DBSCAN algorithm to pre-cluster the pre-processed lidar data, calculate the mean variance of the point cloud distribution in all directions of all clusters in the pre-clustering results, and update the Euclidean distance calculation formula based on the mean variance of the point cloud distribution in all directions. Use the updated Euclidean distance calculation formula to perform improved DBSCAN clustering on the pre-processed lidar data to obtain the final cluster set of lidar targets;

[0011] Step 4: Calculate the centroid of each cluster in the final cluster set of the laser radar target as the position of the laser radar target corresponding to the cluster, build an extended Kalman filter tracker to track the position of the laser radar target, and calculate the displacement, speed and movement direction information of the target according to the target movement trajectory obtained by tracking;

[0012] Step 5, extracting the millimeter wave radar data in the data packet, extracting the Doppler velocity and azimuth information from the millimeter wave radar data, calculating the mapping of the Doppler velocity in the vehicle moving direction, obtaining the speed value in the vehicle moving direction, calculating the speed value distribution curve in the vehicle moving direction, finding the speed value at the peak of the distribution curve, and determining this speed value as the vehicle self-moving speed;

[0013] Step 6: The target speed and direction of motion information obtained by the extended Kalman filter tracker and the vehicle's self-motion speed are integrated to calculate the target's absolute motion speed and determine whether it is a moving target.

[0014] Furthermore, in step 2, the laser radar data in the data packet is extracted and preprocessed, including ROI area extraction and point cloud downsampling. The specific method is as follows:

[0015] Extract the lidar data in the data packet and set the ROI area extraction range threshold x according to the effective detection range of the low-beam lidar in the indoor scene min 、x max ,y min ,y max 、z min 、z max ;

[0016] Use the ROI region extraction range threshold to judge each point cloud in the LiDAR data, retain the points within the ROI extraction range threshold, filter out the points within the extraction range threshold, and obtain the LiDAR data within the ROI range;

[0017] The voxel-based downsampling method is used to downsample the lidar data within the ROI range to obtain the preprocessed lidar data.

[0018] Furthermore, in step 3, the pre-processed lidar data is pre-clustered using the DBSCAN algorithm, the mean of the variance of the point cloud distribution in each direction of all clusters in the pre-clustering results is calculated, and the Euclidean distance calculation formula is updated according to the mean of the variance of the point cloud distribution in each direction. The updated Euclidean distance calculation formula is used to perform improved DBSCAN clustering on the pre-processed lidar data to obtain the final cluster set of the lidar target. The specific method is:

[0019] Step 3.1, preset the pre-clustering parameters neighborhood radius eps' and minimum number of points Minpts', use the DBSCAN algorithm to pre-cluster the pre-processed lidar data, and the pre-clustering result is expressed as {M'1,M'2,…,M' n}, M' n ={(x'1,y'1,z'1),(x'2,y'2,z'2),…,(x' m ,y' m ,z' m )} is a cluster in the pre-clustering result, which contains multiple lidar points; based on each cluster M' n The coordinates of all points in the x-axis direction are used to calculate the mean variance of the point cloud distribution of all clusters. Mean variance of point cloud distribution in the y-axis direction The mean variance of the point cloud distribution in the z-axis direction The calculation process is described as follows:

[0020]

[0021]

[0022]

[0023] Step 3.2: Distribute the variance mean of the point cloud in each direction The weight parameters α, β, and δ of the distances in the x, y, and z axes in the Euclidean distance calculation formula are updated to obtain an improved Euclidean distance calculation formula. The specific design of the improved Euclidean distance calculation formula and the parameter calculation process are as follows:

[0024]

[0025] in,

[0026]

[0027] Step 3.3, set the improved DBSCAN clustering parameters neighborhood radius eps and minimum number of points Minpts. In the improved DBSCAN clustering process, use the improved Euclidean distance calculation formula to calculate the distance r between points in the preprocessed lidar data. Two points that satisfy r<eps belong to the same temporary cluster. The set of temporary clusters {M1,M2,…,M n}, which is the final clustering cluster set of the improved DBSCAN clustering algorithm, where M n ={(x1,y1,z1),(x2,y2,z2),…,(x m ,y m ,z m )}.

[0028] Furthermore, in step 4, the centroid of each cluster in the final cluster set of the laser radar target is calculated as the position of the laser radar target corresponding to the cluster, and an extended Kalman filter tracker is constructed to track the position of the laser radar target. According to the target movement trajectory obtained by tracking, the speed and movement direction information of the target are calculated. The specific method is:

[0029] Step 4.1, traverse the final cluster set of laser radar targets {M1,M2,…,M n Each cluster M in n , by obtaining the cluster M n The centroid (x, y, z) of the point set in the cluster is obtained by calculating the mean of the coordinates of each point in the x, y, and z axes as the corresponding target position, and the cluster M n The maximum coordinate x in the x, y, and z directions max ,y max 、z max and the minimum coordinate x min ,y min 、z min Calculate the target three-dimensional size information (l, w, h). The specific calculation process is as follows:

[0030]

[0031] Step 4.2, concatenate the target position and target 3D size information to obtain the current target state matrix (x, y, z, l, w, h). Associate the current target state matrix with the predicted value of the existing trajectory, and use the centroid displacement ratio and dimensional similarity as the basis for data association. If the predicted value of the existing trajectory and the centroid displacement ratio of the target at the current moment are less than the threshold and the dimensional similarity is less than the threshold, then the two are considered to match, otherwise they are considered to be mismatched; the state matrix of the existing trajectory prediction value is (x A ,y A ,zA ,l A ,w A ,h A ), the centroid displacement ratio is calculated as:

[0032]

[0033] The dimension similarity is calculated as follows:

[0034]

[0035] Among them, x_dim is the x-dimensional similarity, y_dim is the y-dimensional similarity, and z_dim is the z-dimensional similarity.

[0036] Perform corresponding actions based on the matching results:

[0037] If the currently detected pedestrian target matches the existing trajectory successfully, the trajectory is updated using the current target state matrix based on the extended Kalman filter update function;

[0038] If the current detected pedestrian target is not matched successfully, a new tracking trajectory with (x, y, z, l, w, h) as the starting position is initialized;

[0039] If an existing track is not successfully matched, its lost association count coast_cycles is increased by one and compared with the threshold. If it is greater than the threshold, the track is deleted;

[0040] According to the target state matrix at the current moment in the matching result, the tracking trajectory list is updated and the moving trajectory of the pedestrian target is output. The two-dimensional coordinates of the current detected target are (x, y), and the coordinates of the target corresponding to the trajectory N frames ago are (x -N ,y -N ), the displacement calculation method of the current detection target within N frames is:

[0041]

[0042] The speed calculation method of the current detected target within N frames is:

[0043]

[0044] Among them, Δt is the time difference, and the motion direction of the current detection target is calculated as:

[0045]

[0046] Further, step 5 extracts the millimeter wave radar data in the data packet, extracts the Doppler velocity and azimuth information from the millimeter wave radar data, calculates the mapping of the Doppler velocity in the vehicle moving direction, obtains the speed value in the vehicle moving direction, calculates the speed value distribution curve in the vehicle moving direction, finds the speed value at the peak of the distribution curve, and determines that this speed value is the vehicle self-moving speed. The specific method is:

[0047] Extract the Doppler velocity and azimuth information from the millimeter wave radar data in the data packet. Use the radial velocity and azimuth to calculate the speed of each point in the direction of vehicle movement:

[0048]

[0049] Among them, v 车 is the speed of the vehicle in the direction of motion, v 多普勒 is the Doppler velocity in the millimeter-wave radar data, β is the azimuth in the millimeter-wave radar data;

[0050] v 车 Sort and use Gaussian kernel function for v 车 The kernel density estimation of the vehicle's driving direction speed v 车 The probability density function of . Get the v corresponding to the peak position of this probability density function 车 , which is the vehicle's own speed v ego .

[0051] Further, step 6, integrating the target speed and movement direction information obtained by the extended Kalman filter tracker and the vehicle's self-motion speed, calculating the target's absolute movement speed, and determining whether it is a moving target, the specific steps are:

[0052] Calculate the relative velocity in the x-axis direction:

[0053]

[0054] Calculate the relative velocity along the y axis:

[0055]

[0056] Calculate the absolute velocity in the x-axis direction:

[0057]

[0058] Calculate the relative velocity along the y axis:

[0059]

[0060] Determine whether the target is a dynamic target based on speed:

[0061]

[0062] A vehicle-mounted sensor system-based indoor environment dynamic target detection system implements the vehicle-mounted sensor system-based indoor environment dynamic target detection method to achieve vehicle-mounted sensor system-based indoor environment dynamic target detection.

[0063] Compared with the prior art, the present invention has the following significant advantages: using millimeter-wave radar as an odometer to estimate vehicle self-motion can be well adapted to indoor high-repetitive scenes, and compared with the feature point matching algorithm of the laser radar odometer, the millimeter-wave radar odometer greatly reduces the algorithm overhead and ensures the real-time performance of the algorithm. At the same time, the present invention improves the performance of the dynamic target detection algorithm by improving the DBSCAN algorithm, and can be well adapted to low-performance laser radars. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is the overall scheme diagram of the indoor environment dynamic target detection method based on the vehicle-mounted sensor system.

[0065] Figure 2 A specific flow chart of the indoor environment dynamic target detection method based on the vehicle-mounted sensor system

[0066] Figure 3 This is the principle of millimeter wave radar self-motion estimation.

[0067] Figure 4 This is the result diagram of millimeter wave radar self-motion estimation.

[0068] Figure 5 This is the result diagram of dynamic target detection in indoor environment. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0070] The present invention performs filtering preprocessing on the original point cloud data collected by the lidar to reduce the data volume, proposes an improved DBSCAN algorithm based on the anisotropy of the point cloud sparsity, uses the improved DBSCAN algorithm to cluster the preprocessed point cloud, and then uses the extended Kalman filter and Hungarian matching to track the target, calculate the displacement, speed and movement direction of the target, and construct a millimeter-wave radar odometer to estimate the unmanned vehicle's own motion state. The unmanned vehicle self-motion compensation is performed using a low-computational method that does not require feature point matching, and the target motion state change information is obtained to perform dynamic target detection.

[0071] The method for detecting dynamic targets in indoor environments based on a vehicle-mounted sensor system comprises the following steps:

[0072] Step 1: Use the vehicle-mounted sensor system to collect data and obtain data packets, including lidar data and millimeter-wave radar data.

[0073] Step 2: extract the lidar data in the data packet and preprocess the lidar data, including ROI area extraction and point cloud downsampling. The specific method is as follows:

[0074] Extract the lidar data in the data packet and set the ROI area extraction range threshold x according to the effective detection range of the low-beam lidar in the indoor scene min 、x max ,y min ,y max 、z min 、z max ;

[0075] Use the ROI region extraction range threshold to judge each point cloud in the LiDAR data, retain the points within the ROI extraction range threshold, filter out the points within the extraction range threshold, and obtain the LiDAR data within the ROI range;

[0076] The voxel-based downsampling method is used to downsample the lidar data within the ROI range to obtain the preprocessed lidar data.

[0077] Step 3: Use the DBSCAN algorithm to pre-cluster the pre-processed lidar data, calculate the mean variance of the point cloud distribution in all directions of all clusters in the pre-clustering results, and update the Euclidean distance calculation formula based on the mean variance of the point cloud distribution in all directions. Use the updated Euclidean distance calculation formula to perform improved DBSCAN clustering on the pre-processed lidar data to obtain the final cluster set of lidar targets. The specific method is as follows:

[0078] Step 3.1, preset the pre-clustering parameters neighborhood radius eps' and minimum number of points Minpts', use the DBSCAN algorithm to pre-cluster the pre-processed lidar data, and the pre-clustering result is expressed as {M'1,M'2,…,M' n}, M' n ={(x'1,y'1,z'1),(x'2,y'2,z'2),…,(x' m ,y' m ,z' m )} is a cluster in the pre-clustering result, which contains multiple lidar points; based on each cluster M' n The coordinates of all points in the x-axis direction are used to calculate the mean variance of the point cloud distribution of all clusters. Mean variance of point cloud distribution in the y-axis direction The mean variance of the point cloud distribution in the z-axis direction The calculation process is described as follows:

[0079]

[0080]

[0081]

[0082] Step 3.2: Distribute the variance mean of the point cloud in each direction The weight parameters α, β, and δ of the distances in the x, y, and z axes in the Euclidean distance calculation formula are updated to obtain an improved Euclidean distance calculation formula. The specific design of the improved Euclidean distance calculation formula and the parameter calculation process are as follows:

[0083]

[0084] in,

[0085]

[0086] Step 3.3, set the improved DBSCAN clustering parameters neighborhood radius eps and minimum number of points Minpts. In the improved DBSCAN clustering process, use the improved Euclidean distance calculation formula to calculate the distance r between points in the preprocessed lidar data. Two points that satisfy r<eps belong to the same temporary cluster. The set of temporary clusters {M1,M2,…,M n}, which is the final clustering cluster set of the improved DBSCAN clustering algorithm, where M n ={(x1,y1,z1),(x2,y2,z2),…,(x m ,y m ,z m )}.

[0087] Step 4: Calculate the centroid of each cluster in the final cluster set of the LiDAR target as the position of the LiDAR target corresponding to the cluster, build an extended Kalman filter tracker to track the position of the LiDAR target, and calculate the speed and direction of movement of the target based on the target movement trajectory obtained by tracking. The specific method is as follows:

[0088] Step 4.1, traverse the final cluster set of laser radar targets {M1,M2,…,M n Each cluster M in n , by obtaining the cluster M n The centroid (x, y, z) of the point set in the cluster is obtained by calculating the mean of the coordinates of each point in the x, y, and z axes as the corresponding target position, and the cluster M nThe maximum coordinate x in the x, y, and z directions max ,y max 、z max and the minimum coordinate x min ,y min 、z min Calculate the target three-dimensional size information (l, w, h). The specific calculation process is as follows:

[0089]

[0090] Step 4.2, concatenate the target position and target 3D size information to obtain the current target state matrix (x, y, z, l, w, h). Associate the current target state matrix with the predicted value of the existing trajectory, and use the centroid displacement ratio and dimensional similarity as the basis for data association. If the predicted value of the existing trajectory and the centroid displacement ratio of the target at the current moment are less than the threshold and the dimensional similarity is less than the threshold, then the two are considered to match, otherwise they are considered to be mismatched; the state matrix of the existing trajectory prediction value is (x A ,y A ,z A ,l A ,w A ,h A ), the centroid displacement ratio is calculated as:

[0091]

[0092] The dimension similarity is calculated as follows:

[0093]

[0094] Among them, x_dim is the x-dimensional similarity, y_dim is the y-dimensional similarity, and z_dim is the z-dimensional similarity.

[0095] Perform corresponding actions based on the matching results:

[0096] If the currently detected pedestrian target matches the existing trajectory successfully, the trajectory is updated using the current target state matrix based on the extended Kalman filter update function;

[0097] If the current detected pedestrian target is not matched successfully, a new tracking trajectory with (x, y, z, l, w, h) as the starting position is initialized;

[0098] If an existing track is not successfully matched, its lost association count coast_cycles is increased by one and compared with the threshold. If it is greater than the threshold, the track is deleted;

[0099] According to the target state matrix at the current moment in the matching result, the tracking trajectory list is updated and the moving trajectory of the pedestrian target is output. The two-dimensional coordinates of the current detected target are (x, y), and the coordinates of the target corresponding to the trajectory N frames ago are (x -N ,y -N ), the displacement calculation method of the current detection target within N frames is:

[0100]

[0101] The speed calculation method of the current detected target within N frames is:

[0102]

[0103] Among them, Δt is the time difference, and the motion direction of the current detection target is calculated as:

[0104]

[0105] Step 5: extract the millimeter wave radar data in the data packet, extract the Doppler velocity and azimuth information from the millimeter wave radar data, calculate the mapping of the Doppler velocity in the direction of vehicle movement, obtain the speed value in the direction of vehicle movement, calculate the distribution curve of the speed value in the direction of vehicle movement, find the speed value at the peak of the distribution curve, and determine that this speed value is the self-motion speed of the vehicle. The specific method is as follows:

[0106] Extract the Doppler velocity and azimuth information from the millimeter wave radar data in the data packet. Use the radial velocity and azimuth to calculate the speed of each point in the direction of vehicle movement:

[0107]

[0108] Among them, v 车 is the speed of the vehicle in the direction of motion, v 多普勒 is the Doppler velocity in the millimeter-wave radar data, β is the azimuth in the millimeter-wave radar data;

[0109] v 车 Sort and use Gaussian kernel function for v 车 The kernel density estimation of the vehicle's driving direction speed v 车 The probability density function of . Get the v corresponding to the peak position of this probability density function 车 , which is the vehicle's own speed v ego .

[0110] Step 6: Fusion the target speed and direction information obtained by the extended Kalman filter tracker and the vehicle's self-motion speed to calculate the target's absolute speed and determine whether it is a moving target. The specific steps are as follows:

[0111] Calculate the relative velocity in the x-axis direction:

[0112]

[0113] Calculate the relative velocity along the y axis:

[0114]

[0115] Calculate the absolute velocity in the x-axis direction:

[0116]

[0117] Calculate the relative velocity along the y axis:

[0118]

[0119] Determine whether the target is a dynamic target based on speed:

[0120]

[0121] The present invention also proposes an indoor environment dynamic target detection system based on a vehicle-mounted sensor system, implements the indoor environment dynamic target detection method based on a vehicle-mounted sensor system, and realizes indoor environment dynamic target detection based on a vehicle-mounted sensor system.

[0122] The present invention uses a self-motion estimation odometer that is more suitable for indoor scenes, and integrates multiple sensors for target speed measurement and self-motion estimation. It can adapt well to indoor scenes, reduce algorithm overhead, and ensure the real-time performance of the algorithm. At the same time, the present invention proposes an improved DBSCAN algorithm, which improves the performance of the dynamic target detection algorithm, can adapt well to low-performance laser radars, and has huge potential commercial value.

[0123] Example

[0124] In order to verify the effectiveness of the scheme of the present invention, the following experiment was carried out.

[0125] (1) Experimental preparation

[0126] The experimental vehicle hardware includes RS-LiDAR-16-line laser radar and ARS548 laser radar. The software is deployed with Ubuntu20.04 and ros noetic. The improved DBSCAN algorithm is used for point cloud clustering, the extended Kalman filter and Hungarian matching are used for target tracking, the vehicle self-motion estimation is based on the millimeter-wave radar, and the sensor data is fused to obtain the target motion state classification result.

[0127] (2) Experimental Architecture

[0128] The present invention is based on the ROS system, subscribes to the original laser radar point cloud information based on ROS, pre-processes the data, tracks the target and measures the speed, and subscribes to the millimeter wave radar information at the same time, estimates the motion of the unmanned vehicle, and publishes the speed result through the ROS node. The speed measurement data and the self-motion estimation result are integrated to obtain the accurate target motion state. Figure 1 This is a diagram of the experimental architecture.

[0129] (3) Indoor experiment

[0130] The method of the present invention is used to detect dynamic targets in indoor scenes. Figure 4 is the millimeter wave radar self-motion estimation result, Figure 5 The red point cloud represents the dynamic target, and the white point cloud represents the non-dynamic target.

[0131] (4) Experimental results analysis

[0132] Without the use of the present invention, dynamic targets in indoor scenes cannot be well perceived, affecting the accuracy and safety of unmanned driving. After using the present invention, accurate target tracking can be achieved in indoor scenes, and accurate self-motion estimation can be performed by integrating millimeter-wave radar data to further identify dynamic targets. The invention can adapt well to low-performance lidar and complex indoor scenes, and can achieve accurate perception under unstable lighting conditions.

[0133] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0134] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for detecting dynamic targets in indoor environments based on a vehicle-mounted sensor system, characterized in that: The following steps are involved: Step 1: Use the vehicle-mounted sensor system to collect data and obtain data packets, including lidar data and millimeter-wave radar data; Step 2: extract the lidar data in the data packet and preprocess the lidar data, including: ROI area extraction and point cloud downsampling; Step 3: Use the DBSCAN algorithm to pre-cluster the pre-processed lidar data, calculate the mean variance of the point cloud distribution in all directions of all clusters in the pre-clustering results, and update the Euclidean distance calculation formula based on the mean variance of the point cloud distribution in all directions. Use the updated Euclidean distance calculation formula to perform improved DBSCAN clustering on the pre-processed lidar data to obtain the final cluster set of lidar targets; Step 4: Calculate the centroid of each cluster in the final cluster set of the laser radar target as the position of the laser radar target corresponding to the cluster, build an extended Kalman filter tracker to track the position of the laser radar target, and calculate the displacement, speed and movement direction information of the target according to the target movement trajectory obtained by tracking; Step 5, extracting the millimeter wave radar data in the data packet, extracting the Doppler velocity and azimuth information from the millimeter wave radar data, calculating the mapping of the Doppler velocity in the vehicle moving direction, obtaining the speed value in the vehicle moving direction, calculating the speed value distribution curve in the vehicle moving direction, finding the speed value at the peak of the distribution curve, and determining this speed value as the vehicle self-moving speed; Step 6: The target speed and direction of motion information obtained by the extended Kalman filter tracker and the vehicle's self-motion speed are integrated to calculate the target's absolute motion speed and determine whether it is a moving target.

2. The method for detecting dynamic targets in indoor environments based on a vehicle-mounted sensor system according to claim 1, characterized in that: Step 2: extract the lidar data in the data packet and preprocess the lidar data, including ROI area extraction and point cloud downsampling. The specific method is as follows: Extract the lidar data in the data packet and set the ROI area extraction range threshold x according to the effective detection range of the low-beam lidar in the indoor scene min 、x max ,y min ,y max 、z min 、z max ; Use the ROI region extraction range threshold to judge each point cloud in the LiDAR data, retain the points within the ROI extraction range threshold, filter out the points within the extraction range threshold, and obtain the LiDAR data within the ROI range; The voxel-based downsampling method is used to downsample the lidar data within the ROI range to obtain the preprocessed lidar data.

3. The method for detecting dynamic targets in indoor environments based on a vehicle-mounted sensor system according to claim 2, characterized in that: Step 3: Use the DBSCAN algorithm to pre-cluster the pre-processed lidar data, calculate the mean variance of the point cloud distribution in all directions of all clusters in the pre-clustering results, and update the Euclidean distance calculation formula based on the mean variance of the point cloud distribution in all directions. Use the updated Euclidean distance calculation formula to perform improved DBSCAN clustering on the pre-processed lidar data to obtain the final cluster set of lidar targets. The specific method is as follows: Step 3.1, preset the pre-clustering parameters neighborhood radius eps' and minimum number of points Minpts', use the DBSCAN algorithm to pre-cluster the pre-processed lidar data, and the pre-clustering result is expressed as {M'1,M'2,…,M' n }, M' n ={(x'1,y'1,z'1),(x'2,y'2,z'2),…,(x' m ,y' m ,z' m )} is a cluster in the pre-clustering result, which contains multiple lidar points; based on each cluster M' n The coordinates of all points in the x-axis direction are used to calculate the mean variance of the point cloud distribution of all clusters. Mean variance of point cloud distribution in the y-axis direction The mean of the variance of the point cloud distribution in the z-axis direction The calculation process is described as follows: Step 3.2: Distribute the variance mean of the point cloud in each direction The weight parameters α, β, and δ of the distances in the x, y, and z axes in the Euclidean distance calculation formula are updated to obtain an improved Euclidean distance calculation formula. The specific design of the improved Euclidean distance calculation formula and the parameter calculation process are as follows: in, Step 3.3, set the improved DBSCAN clustering parameters neighborhood radius eps and minimum number of points Minpts. In the improved DBSCAN clustering process, use the improved Euclidean distance calculation formula to calculate the distance r between points in the preprocessed lidar data. Two points that satisfy r<eps belong to the same temporary cluster. The set of temporary clusters {M1,M2,…,M n }, which is the final clustering cluster set of the improved DBSCAN clustering algorithm, where M n ={x1,y1,z1),(x2,y2,z2),…,(x m ,y m ,z m )}.

4. The method for detecting dynamic targets in indoor environments based on a vehicle-mounted sensor system according to claim 3, characterized in that: Step 4: Calculate the centroid of each cluster in the final cluster set of the LiDAR target as the position of the LiDAR target corresponding to the cluster, build an extended Kalman filter tracker to track the position of the LiDAR target, and calculate the speed and direction of movement of the target based on the target movement trajectory obtained by tracking. The specific method is as follows: Step 4.1, traverse the final cluster set of laser radar targets {M1,M2,…,M n Each cluster M in n , by obtaining the cluster M n The centroid (x, y, z) of the point set in the cluster is obtained by calculating the mean of the coordinates of each point in the x, y, and z axes as the corresponding target position, and the cluster M n The maximum coordinate x in the x, y, and z directions max ,y max 、z max and the minimum coordinate x min ,y min 、z min Calculate the target three-dimensional size information (l, w, h). The specific calculation process is as follows: Step 4.2, concatenate the target position and target 3D size information to obtain the current target state matrix (x, y, z, l, w, h). Associate the current target state matrix with the predicted value of the existing trajectory, and use the centroid displacement ratio and dimensional similarity as the basis for data association. If the predicted value of the existing trajectory and the centroid displacement ratio of the target at the current moment are less than the threshold and the dimensional similarity is less than the threshold, then the two are considered to match, otherwise they are considered to be mismatched; the state matrix of the existing trajectory prediction value is (x A ,y A ,z A ,l A ,w A ,h A ), the centroid displacement ratio is calculated as: The dimension similarity is calculated as follows: Among them, x_dim is the x-dimensional similarity, y_dim is the y-dimensional similarity, and z_dim is the z-dimensional similarity. Perform corresponding actions based on the matching results: If the currently detected pedestrian target matches the existing trajectory successfully, the trajectory is updated using the current target state matrix based on the extended Kalman filter update function; If the current detected pedestrian target is not matched successfully, a new tracking trajectory with (x, y, z, l, w, h) as the starting position is initialized; If an existing track is not successfully matched, its lost association count coast_cycles is increased by one and compared with the threshold. If it is greater than the threshold, the track is deleted; According to the target state matrix at the current moment in the matching result, the tracking trajectory list is updated and the moving trajectory of the pedestrian target is output. The two-dimensional coordinates of the current detected target are (x, y), and the coordinates of the target corresponding to the trajectory N frames ago are (x -N ,y -N ), the displacement calculation method of the current detection target within N frames is: The speed calculation method of the current detected target within N frames is: Among them, Δt is the time difference, and the motion direction of the current detection target is calculated as:

5. The method for detecting dynamic targets in indoor environments based on a vehicle-mounted sensor system according to claim 4, characterized in that: Step 5: extract the millimeter wave radar data in the data packet, extract the Doppler velocity and azimuth information from the millimeter wave radar data, calculate the mapping of the Doppler velocity in the direction of vehicle movement, obtain the speed value in the direction of vehicle movement, calculate the distribution curve of the speed value in the direction of vehicle movement, find the speed value at the peak of the distribution curve, and determine that this speed value is the self-motion speed of the vehicle. The specific method is as follows: Extract the Doppler velocity and azimuth information from the millimeter wave radar data in the data packet. Use the radial velocity and azimuth to calculate the speed of each point in the direction of vehicle movement: Among them, v 车 is the speed of the vehicle in the direction of motion, v 多普勒 is the Doppler velocity in the millimeter-wave radar data, β is the azimuth in the millimeter-wave radar data; v 车 Sort and use Gaussian kernel function for v 车 The kernel density estimation of the vehicle's driving direction speed v 车 The probability density function of . Get the v corresponding to the peak position of this probability density function 车 , which is the vehicle's own speed v ego .

6. The method for detecting dynamic targets in indoor environments based on a vehicle-mounted sensor system according to claim 5, characterized in that: Step 6: Fusion the target speed and direction information obtained by the extended Kalman filter tracker and the vehicle's self-motion speed to calculate the target's absolute speed and determine whether it is a moving target. The specific steps are as follows: Calculate the relative velocity in the x-axis direction: Calculate the relative velocity in the y-axis direction: Calculate the absolute velocity in the x-axis direction: Calculate the relative velocity in the y-axis direction: Determine whether the target is a dynamic target based on speed:

7. A dynamic target detection system for indoor environment based on a vehicle-mounted sensor system, characterized in that: Implement the indoor environment dynamic target detection method based on the vehicle-mounted sensor system as described in any one of claims 1-5 to realize indoor environment dynamic target detection based on the vehicle-mounted sensor system.

Images