Method and System for Detecting Transparent Obstacles and Map Reconstruction Based on LiDAR Point Cloud Data

Abstract

The present invention discloses a method and system for detecting transparent obstacles and reconstructing a map based on lidar point cloud data, including: scanning a target area using a lidar to obtain multiple sets of point cloud data; registering the multiple sets of obtained point cloud data to the same coordinate system to form a point cloud data set; dividing the point cloud data set in a dynamic window sampling manner and performing geometric structure analysis to find point cloud data that conforms to the characteristics of transparent obstacles; constructing and training an identification model, using the identification model to identify the point cloud data that conforms to the characteristics of transparent obstacles, and modifying the data according to the identification results; replacing the corresponding original data in the original point cloud data set with the modified output data, and constructing a grid map of the target area based on this. The present invention can solve the problem that in the existing map construction process, due to the refraction and diffuse reflection of laser by indoor glass doors, windows or walls, it is impossible to correctly identify obstacles, which affects the driving safety of unmanned vehicles.

Description

Technical Field

[0001] The present invention belongs to the technical field of driverless technology, and in particular relates to a method and system for detecting transparent obstacles and reconstructing a map based on lidar point cloud data. Background Art

[0002] With the development of artificial intelligence technology and driverless vehicle sensor technology, driverless vehicles are increasingly widely used in various fields of social life. When a driverless vehicle moves in an unknown environment, simultaneous localization and mapping (SLAM) has currently become the core technology for realizing the autonomous movement of driverless vehicles. Currently, SLAM technology is divided into two major categories according to the sensors it depends on: lidar-based laser SLAM and vision SLAM based on single / double binocular cameras. In indoor application scenarios, driverless vehicles generally use laser SLAM as the main method because the ranging of lidar is simple and stable, and the driverless vehicle can plan and navigate the path based on the point cloud information returned by the lidar to meet the basic needs of movement.

[0003] However, map construction of lidar in an indoor environment faces challenges, especially the detection of transparent obstacles such as glass doors, windows or walls. Since lidar relies on the reflection of laser signals, the optical properties of transparent obstacles will cause the transmission of laser signals, resulting in most of them being unable to return to the receiving sensor of the lidar, thus affecting the accuracy of indoor environment map construction and bringing difficulties to vehicle movement. Currently, for driverless vehicles in an indoor environment, in order to meet their low-power consumption requirements, usually only a single-line lidar is installed for mapping and navigation, which means that the point cloud information is even scarcer, further posing challenges to the analysis of point cloud data.

[0004] Fortunately, even though most of the laser signals penetrate glass objects, a small part of the signals can still return to the receiver, thus providing certain conditions for distinguishing glass objects. Therefore, how to extract the characteristics of glass objects from a small number of returned signals and identify them as obstacles when constructing a map, so that a driverless vehicle based on a single-line lidar can operate normally in an indoor environment, has become a difficult problem that needs to be solved urgently at present. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the related art to a certain extent.

[0006] An object of the present invention is to provide a method for detecting transparent obstacles and reconstructing a map based on lidar point cloud data, to solve the problem that lidar has poor effect in detecting transparent obstacles, and to ensure that a driverless vehicle can operate normally in an environment with transparent obstacles such as doors, windows or walls.

[0007] Another object of the present invention is to provide a transparent obstacle detection and map reconstruction system based on lidar point cloud data.

[0008] To achieve the above object, on the one hand, the present invention provides a method for transparent obstacle detection and map reconstruction based on lidar point cloud data, including:

[0009] S1. The driverless vehicle uses lidar to scan the entire target area including transparent obstacles to obtain multiple groups of point cloud data;

[0010] S2. The registration algorithm is used to register the obtained multiple groups of point cloud data to the same coordinate system to form a point cloud data set;

[0011] S3. The point cloud data set is divided by the method of dynamic window sampling, and geometric structure analysis is carried out to find the point cloud data that conforms to the characteristics of transparent obstacles;

[0012] S4. An identification model for identifying transparent obstacles is constructed and trained. The identification model is used to identify the point cloud data that conforms to the characteristics of transparent obstacles obtained in step S3. If the part identified as having a transparent obstacle, the point cloud density is supplemented in the output data to be equivalent to that of ordinary obstacles and regarded as an ordinary obstacle; for the part identified as having no transparent obstacle, the original sparse point cloud is removed in the output data and regarded as having no obstacle;

[0013] S5. The modified output data in step S4 is used to replace the corresponding original data in the original point cloud data set, and a grid map of the target area is constructed accordingly.

[0014] A further preferred technical solution of the present invention is that the point cloud data set is divided by the method of local dynamic window sampling in step S3, and geometric structure analysis is carried out to find the point cloud data that conforms to the characteristics of transparent obstacles. The specific method is as follows:

[0015] S31. Set the size of the initial window and traverse the point cloud data set with the initial window;

[0016] S32. Geometric structure analysis is carried out on the point cloud data in the current window to find the point cloud with the geometric characteristics of transparent obstacles. The geometric characteristics of transparent obstacles are manifested as: the front and rear point clouds of the data are relatively dense, reflecting the edges of the transparent obstacles; there are fewer and discontinuous point clouds in the middle, reflecting the transparent obstacles; if the number of point clouds within the straight line area fitted by the two edge coordinates is less than a given threshold, it is considered that this part of the point cloud data is suspected to be a transparent obstacle;

[0017] S33. Increase the window size and perform the next traversal of the point cloud data set until all suspected transparent obstacles of different sizes are completely detected.

[0018] Preferably, every time the window finishes sampling the point cloud dataset in step S3, the density gradient between adjacent sampling windows is calculated. If the window is sampled for the first time or the gradient density change exceeds the feature loss threshold, the window is enlarged by the magnification factor . The calculation formula of the magnification factor is as follows:

[0019]

[0020] where is an adjustable coefficient used to adjust the weight of the gradient density in the formula;

[0021] According to the calculated magnification factor , a new window size is set. The calculation formula of is as follows:

[0022]

[0023] In the formula, is the original window size.

[0024] Preferably, the density gradient is calculated as follows: The number of point clouds within the range of each sampling window is counted as the local density of the window, and the density gradient is obtained by calculating the density difference between adjacent windows.

[0025] Preferably, the recognition model for identifying transparent obstacles constructed in step S4 includes:

[0026] Input layer, which obtains the data input by the sampling window and processes it into a tensor format and then transfers it to the feature extraction layer;

[0027] Feature extraction layer, which first independently extracts features from each point cloud tensor using a multi-layer perceptron, and then obtains a fixed-length global feature vector through max pooling and fuses it with the local features, and outputs a feature matrix with a dimension of , represents the number of point clouds input to the network for each window, and represents the dimension of the coordinate system;

[0028] Output layer, which takes the feature matrix as input, and after processing by the fully connected layer and the softmax function, outputs the probability that the point cloud contains transparent obstacles. If the probability exceeds the threshold, it is determined that there is a point cloud with a transparent glass obstacle.

[0029] Preferably, the method for training the recognition model is as follows: After annotating the point cloud data set detected by the radar, a training set is generated, and the recognition model is trained multiple times with this training set, and the model parameters are adjusted to the optimal values. The loss function during the training process is the cross-entropy loss and the Dice loss.

[0030] Preferably, the data input into the sampling window in the input layer is preprocessed to a specified data length, and the preprocessing method is one of the farthest point sampling algorithm or the average sampling algorithm;

[0031] When the output layer outputs data, in the reverse processing mode of the preprocessing, the length of the output data is restored to the normal data length.

[0032] Preferably, the specific method of step S1 is as follows:

[0033] S11. Plan a path for the driverless vehicle so that the driving route of the driverless vehicle can cover the entire target area;

[0034] S12. The driverless vehicle uses a lidar to scan at the set scanning points, and obtains a set of point cloud data after each 360° scan; the point cloud data includes at least distance information and point cloud data intensity information, and the scanning resolution is determined by the radar model.

[0035] Preferably, after constructing the grid map, if there is a dynamic change in the open or closed state of the glass door, return to step S2, correct the real-time obtained point cloud data into the point cloud data set, re-identify the transparent obstacle, and update the grid map of the target area.

[0036] On the other hand, the present invention provides a transparent obstacle detection and map reconstruction system based on lidar point cloud data, including:

[0037] An information acquisition module for sensing the target area using the lidar of the driverless vehicle;

[0038] A calculation module for storing the data collected by the driverless vehicle, deploying a recognition model for identifying transparent obstacles, identifying transparent obstacles from the collected data, and then outputting corrected data;

[0039] A map construction module for correcting the data collected by the driverless vehicle according to the corrected data and constructing a grid map of the target area accordingly;

[0040] A communication module, connected to external devices through a wired or wireless network, for obtaining the real-time state of the driverless vehicle and remote control;

[0041] A power supply module for supplying power to the information acquisition module, calculation module, map construction module and communication module of the driverless vehicle.

[0042] Beneficial effects: The present invention provides a method and system for detecting transparent obstacles and map reconstruction based on lidar point cloud data. First, lidar point cloud data with glass in the surrounding environment is collected, and the processed point cloud data is used for deep learning model training. Secondly, the real-time collected point cloud data is directly input into the model. If glass doors, windows or walls are detected, the deep learning model will obtain their position information. Subsequently, the point cloud data is interpolated and complemented, and finally, a reconstructed map with glass doors, windows or walls as wall obstacles is output. The present invention can solve the problem that in the existing map construction process, due to the refraction and diffuse reflection of laser by indoor glass doors, windows or walls, it is impossible to correctly identify obstacles, which affects the driving safety of unmanned vehicles. The deep learning model directly uses point cloud data as input, and has the advantages of less calculation amount and simple deployment. Description of the drawings

[0043] Figure 1 It is a flowchart of the method for detecting transparent obstacles and map reconstruction based on lidar point cloud data of the present invention.

[0044] Figure 2 It is a comparison diagram of the lidar point cloud data before and after modification in Embodiment 1 of the present invention;

[0045] Figure 2 In (a), it is a demonstration diagram of obstacles before data modification; Figure 2 In (b), it is a demonstration diagram of obstacles after data modification.

[0046] Figure 3 It is a comparison diagram of the map reconstruction effects before and after in Embodiment 1 of the present invention;

[0047] Figure 3 In (a), it is the original grid map of the target area; Figure 3 In (b), it is the reconstructed grid map of the target area. Detailed implementation manners

[0048] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention, and they should not be construed as limiting the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. In the description of the present invention, it should be understood that the terms used are only for the purpose of description and cannot be construed as indicating or implying relative importance.

[0049] The following is combined with Figures 1-3Describe a method and system for transparent obstacle detection and map reconstruction based on lidar point cloud data provided by the present invention.

[0050] Embodiment 1: This embodiment provides a method for transparent obstacle detection and map reconstruction based on lidar point cloud data. The overall process of this method is as Figure 1 shown and includes:

[0051] S1. The driverless vehicle uses a lidar to scan the entire target area including transparent obstacles to obtain multiple sets of point cloud data;

[0052] S2. The registration algorithm is used to register the multiple sets of point cloud data obtained to the same coordinate system to form a point cloud data set;

[0053] S3. The point cloud data set is divided by the dynamic window sampling method and geometric structure analysis is performed to find the point cloud data that conforms to the characteristics of transparent obstacles;

[0054] S4. An identification model for identifying transparent obstacles is constructed and trained. The identification model is used to identify the point cloud data that conforms to the characteristics of transparent obstacles obtained in step S3. If the part identified as having a transparent obstacle, the point cloud density is supplemented in the output data to be equivalent to that of ordinary obstacles and regarded as an ordinary obstacle; for the part identified as having no transparent obstacle, the original sparse point cloud is removed in the output data and regarded as having no obstacle;

[0055] S5. The modified output data in step S4 is used to replace the corresponding original data in the original point cloud data set, and a grid map of the target area is constructed accordingly.

[0056] The above steps will be described in detail through specific examples below.

[0057] In step S1, first, any path planning algorithm can be adopted to make the driving route of the driverless vehicle cover the entire indoor space. The driverless vehicle uses a lidar to collect data from the indoor area containing transparent obstacles such as glass doors, windows, or walls.

[0058] In this embodiment, the driverless vehicle uses the A* algorithm for path planning and is equipped with a lidar device of the Ydlidar-X3 model for data collection. The point cloud data obtained by this lidar device consists of a frame header, lidar scanning parameters, scanning distance information, and point cloud data intensity information, etc. The scanning parameters are set as follows: the scanning start angle starts from +180° and ends at -180°, the angle between adjacent two scanning points increases by 0.5°, the scanning time increases by 0.0001 seconds, and the minimum distance that this lidar can detect is from 0.07 meters to the maximum distance of 16 meters; in the scanning distance information, since it is collected every 0.5° and the scanning range is 360°, the distance information is organized into a one-dimensional array containing 720 elements, and each element represents the distance (in meters) from the lidar center to the detected object at the corresponding angle.

[0059] The purpose of step S2 is to register the point cloud data.

[0060] In the point cloud data obtained by each lidar scan, the origin of the point cloud coordinates is the position of the lidar, resulting in the point cloud data obtained by each scan not being in the same coordinate system. At the same time, since the driverless vehicle may still be in a driving state, the point cloud data obtained under different scanning perspectives of the lidar may also not be in the same coordinate system. Therefore, it is necessary to perform a registration operation on the point cloud data to obtain a point cloud data set that can describe the complete indoor structure information under a unified coordinate system. The registration algorithms that can be used include but are not limited to the Iterative Closest Point algorithm, the Normal Distribution Transform algorithm, etc., and this embodiment does not make any limitations in this regard.

[0061] Step S3 is to perform geometric structure analysis on the registered point cloud data set in the way of dynamic window sampling to find the part that conforms to the characteristics of transparent obstacles.

[0062] First, determine the characteristic geometric features of the transparent obstacle point cloud as follows: the front and rear point clouds of the data are relatively dense, which reflects the edges of the transparent obstacle, and there are fewer and discontinuous point clouds in the middle, which reflects the transparent obstacle. If the number of point clouds within the straight line area fitted by the two edge coordinates is less than a given threshold , then it is considered that this part of the data may be a transparent obstacle, and the threshold can be obtained through data pre-training.

[0063] Secondly, in this embodiment, the way of dynamic window sampling is adopted to sample the point cloud data set. The specific method is as follows:

[0064] S31. Set the size of the initial window and traverse the point cloud data set with the initial window;

[0065] S32. Perform geometric structure analysis on the point cloud data within the current window to find the point cloud data with the geometric characteristics of transparent obstacles;

[0066] S33. Increase the window size and perform the next traversal of the point cloud dataset until all suspected transparent obstacles of different sizes are completely detected.

[0067] Among them, it is necessary to preset the window size adjustment rule in advance: after each sampling of the point cloud dataset by the window, calculate the density gradient between adjacent sampling windows , the density gradient is calculated as follows: count the number of point clouds within each sampling window range as the local density of the window, and calculate the density difference between adjacent windows to obtain the density gradient.

[0068] If the window is for the first sampling or the gradient density change exceeds the feature loss threshold , then expand the window by the magnification factor , the magnification factor is calculated as:

[0069]

[0070] Among them, is an adjustable coefficient used to adjust the weight of the gradient density in the formula, and this adjustment coefficient needs to be set according to the experimental scenario;

[0071] According to the calculated magnification factor , set the new window size , The calculation formula is:

[0072]

[0073] In the formula, is the original window size.

[0074] The purpose of step S4 is to further determine whether the point cloud data contains transparent obstacles through the neural network, and then modify the network output.

[0075] In this step, first, a recognition model for identifying transparent obstacles needs to be constructed based on the neural network as a neural network model and trained.

[0076] The architecture of the recognition model is:

[0077] Input layer, obtain the data input by the sampling window, and process it into a tensor format and then transfer it to the feature extraction layer;

[0078] Feature extraction layer, first use a multi-layer perceptron to independently extract features from each point cloud tensor, and then obtain a fixed-length global feature vector through max pooling and fuse it with the local features, and the output dimension is The feature matrix, represents the number of point clouds input to the network for each window, represents the dimension of the coordinate system.

[0079] The output layer takes the feature matrix as input, and after processing by the fully connected layer and the softmax function, outputs the probability that the point cloud contains a transparent obstacle. If the probability exceeds the threshold, it is determined that the point cloud with a transparent glass obstacle exists, where the probability threshold needs to be set according to the test scenario.

[0080] The parameters of the recognition network need to be obtained after network training. The point cloud dataset detected by the radar can be labeled to generate a training set, and the recognition model is trained multiple times with this training set, and the model parameters are adjusted to the optimal values to obtain the recognition model.

[0081] In this embodiment, the data acquisition scenario is an indoor laboratory scenario, which contains transparent glass doors of different lengths and sizes. The driverless vehicle is controlled to drive in this scenario and collect the point cloud data output by the lidar. In this embodiment, 360 groups of point cloud data are collected as the training data of the recognition network, and are divided into a training set and a validation set according to a ratio of 7:3. The optimizer used during network training is AdamW, the learning rate adjustment method is the cosine annealing algorithm, the loss function during the training process is the cross-entropy loss and the Dice loss, and the number of training rounds is set to Epochs, and the range of Epochs is between 100 and 300.

[0082] The data dimension of the sampling window input to the recognition network is ( , 2), represents the number of point clouds input to the network for each window, and 2 indicates that it contains two-dimensional coordinate position information; in the feature extraction layer, the data is upsampled to ( , 128) through a multi-layer perceptron layer for the input point cloud, that is, the original two-dimensional features of the point cloud are extended to 128-dimensional features to highlight the information features of the point cloud; the upsampled data is input to the max pooling layer, and after max pooling, data of size (1, 128) is obtained; finally, the data is input to the output layer, and in the output layer, data of size (1, 2) is obtained through the fully connected layer, where 2 represents the prediction score of whether there is a transparent obstacle, and then through the softmax function, these two scores are converted into a probability distribution, and each value represents the probability that the point cloud belongs to the corresponding category.

[0083] After the model training is completed, the recognition model is used to recognize the point cloud data that conforms to the characteristics of the transparent obstacle obtained in step S3. Before the data is input into the recognition model, preprocessing is performed on the window sampling data. The purpose of the preprocessing is to make the length of the data input into the recognition network consistent for easy processing. The preprocessing algorithms include, but are not limited to, the farthest point sampling algorithm, the average sampling algorithm, etc. In this embodiment, the farthest point sampling algorithm is used for the preprocessing of the point cloud. The sampling distance threshold in the farthest point sampling is 0.3 meters.

[0084] The recognition model outputs the probability of whether the point cloud in the window contains a transparent obstacle. If it exceeds the set probability threshold , it is determined that this part of the point cloud contains a transparent obstacle, and the two-dimensional coordinate information of the edge of the transparent obstacle in the map is output. In the embodiment, the threshold is set to 0.6.

[0085] For the point cloud containing a transparent obstacle, according to the edge information of the transparent obstacle point cloud, the missing point cloud is interpolated and complemented, and it is regarded as an ordinary obstacle. In this embodiment, linear interpolation is used, and the edge coordinate information is used as the input to complete the point cloud information. For the point cloud without a transparent obstacle, the original sparse point cloud is removed and regarded as an obstacle-free area. The differences before and after the data modification are as Figure 2 shown. Figure 2 In (a) is the demonstration diagram of the obstacle before the data modification, Figure 2 in (b) is the demonstration diagram of the obstacle after the data modification. Comparing the two figures, it is obvious that the transparent obstacle is complemented and the non-transparent obstacle is deleted.

[0086] The modified data is transformed in the reverse processing mode of the preprocessing, so that the length of the output data is restored to the normal data length, which is convenient for filling into the original point cloud data.

[0087] Step S5 is the map reconstruction step.

[0088] Replacing the corresponding original data in the original point cloud dataset with the output data of the recognition model, the corrected target area grid map can be obtained. The maps constructed before and after the data correction are as Figure 3 shown. Figure 3 In (a) is the original grid map of the target area; Figure 2 in (b) is the reconstructed grid map of the target area. When the driverless vehicle plans the path according to the corrected map, it can accurately avoid the glass obstacle.

[0089] In addition, during the subsequent operation of the driverless vehicle, the lidar still works continuously. If there are dynamic changes in the open or closed state of the glass door, the above steps are repeated to continuously update the target area grid map.

[0090] Example 2: This example provides a transparent obstacle detection and map reconstruction system based on lidar point cloud data. This system is deployed on an autonomous vehicle and serves as the physical carrier for implementing the method of Example 1. It includes:

[0091] An information acquisition module, which is used to sense the target area by using the lidar of the autonomous vehicle; corresponding to the execution of step S1 in Example 1. It should be noted that in this example, mainly a single-line lidar is used for data acquisition, but it is not limited to adding other sensors such as an IMU sensor and a GPS module for information fusion processing.

[0092] A calculation module, which is used to store the data collected by the autonomous vehicle, deploy an identification model for identifying transparent obstacles, identify transparent obstacles from the collected data, and then output corrected data; corresponding to the execution of steps S2 - S4 in Example 1.

[0093] A map construction module, which is used to correct the data collected by the autonomous vehicle according to the corrected data and construct a grid map of the target area accordingly; corresponding to the execution of step S5 in Example 1.

[0094] A communication module, which is connected to external devices through a wired or wireless network and is used for obtaining the real-time status and remote control of the autonomous vehicle.

[0095] A power supply module, which is used to supply power to the information acquisition module, calculation module, map construction module, and communication module of the autonomous vehicle.

[0096] The system examples described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this example. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0097] Through the description of the above implementation manners, those skilled in the art can clearly understand that each implementation manner can be realized by means of software plus a necessary general hardware platform, and of course, it can also be realized by hardware. Based on such an understanding, the above technical solutions, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each example or some parts of the examples.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A transparent obstacle detection and map reconstruction method based on laser radar point cloud data, characterized in that: include: S1. The unmanned vehicle uses laser radar to scan the entire target area including transparent obstacles to obtain multiple sets of point cloud data; S2, using a registration algorithm to register the acquired multiple sets of point cloud data to the same coordinate system to form a point cloud data set; S3. Divide the point cloud data set by dynamic window sampling, and perform geometric structure analysis to find point cloud data that meets the characteristics of transparent obstacles; the specific method is: S31, setting the size of the initial window, and traversing the point cloud data set with the initial window; S32, performing geometric structure analysis on the point cloud data in the current window, and finding point cloud data with geometric features of transparent obstacles. The geometric features of transparent obstacles are as follows: the point clouds at the front and back ends of the data are relatively dense, which are reflected as the edges of transparent obstacles; there are fewer and discontinuous point clouds in the middle, which are reflected as transparent obstacles; if the number of point clouds in the straight line area fitted by the coordinates of the two edges is less than a given threshold, it is considered that the part of the point cloud data is suspected to be a transparent obstacle; S33, increasing the window size, and performing the next traversal of the point cloud data set until all suspected transparent obstacles of different sizes are completely detected; In step S3, each time the window completes sampling of the point cloud data set, the density gradient between adjacent sampling windows is calculated. If the window is sampled for the first time or the gradient density changes exceed the feature loss threshold, the magnification factor To expand the window, the magnification factor The calculation formula is: ； in, is an adjustable coefficient used to adjust the weight of the gradient density in the formula; The calculated magnification factor , set the new window size , The calculation formula is: ； In the formula, is the original window size; S4. Construct and train a recognition model for identifying transparent obstacles, and use the recognition model to identify the point cloud data that meets the characteristics of transparent obstacles obtained in step S3. If a part is identified as having a transparent obstacle, the point cloud density is supplemented in the output data to be equivalent to that of an ordinary obstacle, and the part is regarded as an ordinary obstacle; if a part is identified as not having a transparent obstacle, the original sparse point cloud is removed from the output data, and the part is regarded as having no obstacle; S5. Replace the corresponding original data in the original point cloud data set with the modified output data of step S4, and construct a grid map of the target area based on the modified output data.

2. The transparent obstacle detection and map reconstruction method based on laser radar point cloud data according to claim 1 is characterized in that: The density gradient The calculation method is as follows: the number of point clouds within each sampling window is counted as the local density of the window, and the density gradient is obtained by calculating the density difference between adjacent windows.

3. The transparent obstacle detection and map reconstruction method based on laser radar point cloud data according to claim 1, characterized in that: The recognition model for identifying transparent obstacles constructed in step S4 is a neural network model, including: The input layer obtains the data input from the sampling window and processes it into a tensor format before passing it to the feature extraction layer; The feature extraction layer first uses a multi-layer perceptron to independently extract features from each point cloud tensor, then obtains a global feature vector of fixed length through maximum pooling and fuses it with local features. The output dimension is The characteristic matrix of Represents the number of point clouds input into the network for each window, Indicates the dimension of the coordinate system; The output layer takes the feature matrix as input and outputs the probability of containing transparent obstacles in the point cloud after processing by the fully connected layer and the softmax function. If the probability exceeds the threshold, it is determined that there is a point cloud with transparent glass obstacles.

4. The transparent obstacle detection and map reconstruction method based on laser radar point cloud data according to claim 3 is characterized in that: The method for training the recognition model is: annotating the point cloud data set detected by the radar to generate a training set, training the recognition model multiple times with the training set, adjusting the model parameters to the optimal values, and the loss functions of the training process are cross entropy loss and Dice loss.

5. The transparent obstacle detection and map reconstruction method based on laser radar point cloud data according to claim 3 is characterized in that: The data inputted by the sampling window in the input layer is preprocessed to a specified data length, and the preprocessing method is one of the farthest point sampling algorithm or the average sampling algorithm; When the output layer outputs data, it uses the reverse processing method of preprocessing to restore the length of the output data to the normal data length.

6. The transparent obstacle detection and map reconstruction method based on laser radar point cloud data according to claim 1, characterized in that: The specific method of step S1 is: S11. Planning a path for the unmanned vehicle so that the driving route of the unmanned vehicle can cover the entire target area; S12. The unmanned vehicle uses a laser radar to scan at a set scanning point, and obtains a set of point cloud data after each 360° scan; the point cloud data at least includes distance information and point cloud data intensity information, and the scanning resolution is determined by the radar model.

7. The transparent obstacle detection and map reconstruction method based on laser radar point cloud data according to claim 1, characterized in that: After constructing the grid map, if there is a dynamic change in the opening or closing state of the glass door, return to step S2, correct the point cloud data acquired in real time to the point cloud data set, re-identify the transparent obstacles, and update the target area grid map.

8. A transparent obstacle detection and map reconstruction system based on laser radar point cloud data using the method described in any one of claims 1 to 7, characterized in that: include: An information collection module, used to sense the target area using the laser radar of the unmanned vehicle; A computing module, which is used to store the data collected by the unmanned vehicle, deploy a recognition model for identifying transparent obstacles, identify the transparent obstacles from the collected data, and then output correction data; A map construction module is used to correct the data collected by the unmanned vehicle according to the correction data, and to construct a grid map of the target area accordingly; Communication module, which uses wired or wireless network to connect external devices for real-time status acquisition and remote control of unmanned vehicles; The power module is used to supply power to the information collection module, computing module, map building module and communication module of the unmanned vehicle.

Images