Travelable area determination method, computer-readable storage medium, and driving device

Abstract

The present application relates to the technical field of intelligent driving, and particularly provides a drivable area determination method, a computer readable storage medium and a driving device, the drivable area determination method comprises the following steps: inputting point cloud data into a neural network model, obtaining ground height based on a branch ground height acquisition model of the neural network model; and determining a drivable area according to at least the ground height and coordinate data of the point cloud data. The ground height acquisition model is obtained by calculating a loss function of ground predicted height and ground height true value, and training an initial ground height acquisition model based on the loss function, and the ground height true value is obtained by fitting the ground height of point cloud training data. This method effectively avoids determining the ground height true value by labeling information, thereby saving the labeling cost, and effectively improving the convenience and generalization ability of determining the drivable area.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, and particularly provides a drivable area determination method, a computer readable storage medium and a driving device. BACKGROUND

[0002] The drivable area detection technology is one of important technologies for assisting in realizing automatic driving. The drivable area detection technology is used to determine the drivable area in a driving environment, so as to plan a driving path for a driving device. Therefore, a convenient, low-cost and strong generalization capability drivable area detection technology has become a focus.

[0003] In the prior art, when the drivable area is detected, the point cloud data is usually classified and labeled by using semantic segmentation, and then the ground truth of the point cloud data is determined based on the labeled information of the point cloud data. This method needs to be labeled, and has high cost, is not convenient, and has poor generalization capability. In addition, it takes a long time to deploy the drivable area detection method in a central processor. SUMMARY

[0004] The present application aims to solve the above technical problems, i.e., to solve the problem that the existing drivable area detection technology needs to be labeled, has high cost, is not convenient, and has poor generalization capability.

[0005] In a first aspect, the present application provides a drivable area determination method, which comprises:

[0006] acquiring point cloud data of a driving environment;

[0007] inputting the point cloud data into a trained neural network model, wherein the neural network model at least comprises a ground height obtaining model, and obtaining the ground height of the point cloud data based on the ground height obtaining model;

[0008] determining a drivable area according to at least the ground height of the point cloud data and coordinate data of the point cloud data;

[0009] wherein the ground height obtaining model is obtained by the following steps:

[0010] acquiring point cloud training data;

[0011] inputting the point cloud training data into an initial ground height obtaining model to obtain the ground predicted height of the point cloud training data;

[0012] ground height fitting is performed on the point cloud training data to obtain the ground truth of the point cloud training data;

[0013] compute a loss function of the ground predicted height and the ground height truth value, and train the initial ground height acquisition model based on the loss function to obtain the trained ground height acquisition model.

[0014] In some embodiments, the ground height fitting on the point cloud training data to obtain the ground height truth value of the point cloud training data comprises:

[0015] dividing the point cloud training data into grids to obtain a plurality of divided grid point cloud data sets;

[0016] fitting a plane to the point cloud training data in each grid point cloud data set respectively, and determining the type of the grid point cloud data set according to the fitted plane, wherein the type of the grid point cloud data set comprises a ground point grid point cloud data set and / or a non-ground point grid point cloud data set;

[0017] determining the ground height truth value of the grid point cloud data set based on the type of the grid point cloud data set.

[0018] In some embodiments, the determining the ground height truth value of the grid point cloud data set based on the type of the grid point cloud data set comprises:

[0019] when the type of the grid point cloud data set is the ground point grid point cloud data set, determining the ground height truth value of the ground point grid point cloud data set according to the coordinate data of the point cloud training data in the ground point grid point cloud data set;

[0020] when the type of the grid point cloud data set is the non-ground point grid point cloud data set, determining the ground height truth value of the non-ground point grid point cloud data set based on the ground point grid point cloud data set satisfying a preset distance condition and a preset orientation condition with the non-ground point grid point cloud data set.

[0021] In some embodiments, the determining the ground height truth value of the non-ground point grid point cloud data set based on the ground point grid point cloud data set satisfying a preset distance condition and a preset orientation condition with the non-ground point grid point cloud data set comprises:

[0022] selecting the ground point grid point cloud data set closest to the non-ground point grid point cloud data set and located at four different orientations of the non-ground point grid point cloud data set;

[0023] obtaining the ground height truth value of the ground point grid point cloud data set at the four different orientations;

[0024] perform weighted summation and average on the ground height true values of the ground point grid point cloud data sets in four different orientations and the weights corresponding to the ground point grid point cloud data sets in each orientation to obtain the ground height true value of the non-ground point grid point cloud data set, wherein the weight corresponding to the ground point grid point cloud data set in each orientation is inversely proportional to the distance between the ground point grid point cloud data set and the non-ground point grid point cloud data set.

[0025] In some embodiments, the type of the grid point cloud data set is determined according to the fitted plane, comprising:

[0026] obtaining the normal of the plane;

[0027] determining whether the included angle between the normal of the plane and the normal of the real ground is less than a preset included angle threshold;

[0028] if yes, determining that the type of the grid point cloud data set is a ground point grid point cloud data set; if no, determining that the type of the grid point cloud data set is a non-ground point grid point cloud data set.

[0029] In some embodiments, the type of the grid point cloud data set is determined according to the fitted plane, comprising:

[0030] obtaining the normal of the plane;

[0031] determining whether the included angle between the normal of the plane and the normal of the real ground is less than a preset included angle threshold;

[0032] if yes, determining the average value or variance of the plurality of point cloud training data in the grid point cloud data set projected in the direction perpendicular to the real ground;

[0033] when the average value of the plurality of point cloud training data projected in the direction perpendicular to the real ground is less than or equal to a first preset threshold, or the variance of the plurality of point cloud training data projected in the direction perpendicular to the real ground is less than or equal to a second preset threshold, determining that the type of the grid point cloud data set is a ground point grid point cloud data set;

[0034] when the average value of the plurality of point cloud training data projected in the direction perpendicular to the real ground is greater than the first preset threshold, or the variance of the plurality of point cloud training data projected in the direction perpendicular to the real ground is greater than the second preset threshold, determining that the type of the grid point cloud data set is a non-ground point grid point cloud data set.

[0035] In some embodiments, the drivable area is determined according to at least the ground height of the point cloud data and the coordinate data of the point cloud data, comprising:

[0036] determine a coordinate value of the point cloud data projected in a direction perpendicular to a real ground surface direction according to coordinate data of the point cloud data; and screen the point cloud data with the coordinate value greater than the ground height and take the point cloud data as non-collision point cloud data.

[0037] determine a drivable area based on the non-collision point cloud data.

[0038] In some embodiments, the neural network model further comprises an obstacle detection model, the obstacle detection model and the ground height acquisition model share a backbone network, and the obstacle detection model and the ground height acquisition model are obtained by synchronous training; after the point cloud data is input into the neural network model, the method further comprises:

[0039] obtain an obstacle detection result based on the obstacle detection model;

[0040] determining a drivable area according to at least the ground height of the point cloud data and coordinate data of the point cloud data comprises:

[0041] determining a drivable area according to the obstacle detection result and the ground height and the coordinate data of the point cloud data.

[0042] In a second aspect, the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by a processor to implement the drivable area determination method of any one of the above.

[0043] In a third aspect, the present application provides a driving device, which comprises a driving device body, a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to implement the drivable area determination method of any one of the above.

[0044] In the above technical solution, the point cloud data is input into the neural network model, the ground height is obtained based on a branch ground height acquisition model of the neural network model, and the drivable area is determined according to at least the ground height and the coordinate data of the point cloud data. The ground height acquisition model is obtained by calculating the loss function of the ground predicted height and the ground height true value and training the initial ground height acquisition model based on the loss function, and the ground height true value is obtained by fitting the ground height of the point cloud training data. This method effectively avoids determining the ground height true value by labeling information, thereby saving the labeling cost and effectively improving the convenience and generalization ability of determining the drivable area.

[0045] In addition, the neural network model can further include an obstacle detection model branch, the ground height prediction model and the obstacle detection model share a main network, and can be obtained by synchronous training, thereby effectively realizing the lightweight of the neural network model and reducing the overall deployment time of the neural network model. BRIEF DESCRIPTION OF DRAWINGS

[0046] The preferred embodiments of the present application will be described below with reference to the accompanying drawings, in which:

[0047] Figure 1 is a flowchart of a drivable area determination method provided by an embodiment of the present application;

[0048] Figure 2 is a structural diagram of a ground height acquisition model provided by an embodiment of the present application;

[0049] Figure 3 is a flowchart of a method for training a ground height acquisition model provided by an embodiment of the present application;

[0050] Figure 4 is a flowchart of a method for determining ground height true values of point cloud training data provided by an embodiment of the present application;

[0051] Figure 5 is a flowchart of another drivable area determination method provided by an embodiment of the present application;

[0052] Figure 6 is a structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0053] Some embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0054] In the prior art, when performing drivable area detection, it is usually necessary to classify and label point cloud data using semantic segmentation, and then determine the point cloud data true value based on the labeled information of the point cloud data. This method needs to be labeled, has high cost, is not convenient, has poor generalization ability, and in addition, deploying the drivable area detection method alone in the central processor takes a long time.

[0055] Therefore, the application provides a drivable area determination method, which comprises the following steps: inputting point cloud data into a neural network model, obtaining ground height based on a ground height acquisition model of the neural network model; and determining a drivable area according to at least the ground height and coordinate data of the point cloud data. The ground height acquisition model is obtained by calculating a loss function of ground predicted height and ground height true value, and training an initial ground height acquisition model based on the loss function. The ground height true value is obtained by fitting the ground height of point cloud training data. The method effectively avoids determining the ground height true value by using labeled information, thereby saving the labeling cost, and effectively improving the convenience and generalization ability of determining the drivable area.

[0056] Referring to Figure 1 as shown, Figure 1 is a flowchart of a drivable area determination method provided by an embodiment of the application, which can comprise the following steps:

[0057] Step S11: collecting point cloud data of a driving environment;

[0058] Step S12: inputting the point cloud data into a trained neural network model, wherein the neural network model at least comprises a ground height acquisition model, and obtaining the ground height of the point cloud data based on the ground height acquisition model;

[0059] Step S13: determining a drivable area according to at least the ground height of the point cloud data and the coordinate data of the point cloud data.

[0060] In some embodiments, step S12 can specifically comprise: dividing the point cloud data into a plurality of grids; and determining the ground height of the point cloud data in each grid by using the ground height acquisition model respectively. One grid corresponds to one ground height.

[0061] In some embodiments, step S13 can specifically comprise:

[0062] determining a coordinate value of the point cloud data projected in a direction perpendicular to the real ground surface according to the coordinate data of the point cloud data;

[0063] screening the point cloud data with a coordinate value greater than the ground height and taking the point cloud data as non-collision point cloud data;

[0064] determining the drivable area based on the non-collision point cloud data.

[0065] In some embodiments, when one grid corresponds to one ground height, screening the point cloud data with a coordinate value greater than the ground height and taking the point cloud data as the non-collision point cloud data can specifically be that, for each grid, comparing the coordinate value of each point cloud data in the grid projected in the direction perpendicular to the real ground with the ground height corresponding to the grid; when the coordinate value of the point cloud data projected in the direction perpendicular to the real ground is less than or equal to the ground height, determining the point cloud data as ground point cloud data and filtering the determined ground point cloud data; when the coordinate value of the point cloud data projected in the direction perpendicular to the real ground is greater than the ground height, determining the point cloud data as non-ground point cloud data and taking the determined non-ground point cloud data as the non-collision point cloud data.

[0066] In the embodiments of the present application, the non-collision point cloud data corresponds to the non-drivable area, and the area other than the non-drivable area can be determined as the drivable area.

[0067] Referring to Figure 2 and Figure 3 , Figure 2 is a structural schematic diagram of a ground height acquisition model provided by the embodiments of the present application, and the ground height acquisition model can include a data input module, a ground height prediction sub-network, a ground height true value generation sub-network and a loss function calculation sub-network. Figure 3 is a method flow schematic diagram of training the ground height acquisition model provided by the embodiments of the present application, and the method can include:

[0068] Step S21: acquiring point cloud training data;

[0069] Step S22: inputting the point cloud training data into an initial ground height acquisition model to obtain a ground predicted height of the point cloud training data;

[0070] and, Step S23: fitting the ground height of the point cloud training data to obtain a ground height true value of the point cloud training data;

[0071] Step S24: calculating a loss function of the ground predicted height and the ground height true value, and training the initial ground height acquisition model based on the loss function to obtain a trained ground height acquisition model.

[0072] In some embodiments, Step S22 can specifically be inputting the point cloud training data into the initial ground height acquisition model, and obtaining the ground predicted height of the point cloud training data based on the ground height prediction sub-network of the initial ground height acquisition model.

[0073] In some embodiments, the point cloud training data can also be rasterized to obtain a plurality of raster point cloud data sets; for each raster point cloud data set, a ground height prediction sub-network is used to obtain a ground prediction height of the raster point cloud data set. One raster point cloud data set corresponds to one ground prediction height.

[0074] In some embodiments, referring to Figure 4 , the ground height of the point cloud training data is determined. Figure 4 A method for determining the ground height of point cloud training data is shown. Step S23 can specifically include:

[0075] Step S231: The point cloud training data is rasterized to obtain a plurality of raster point cloud data sets after rasterization.

[0076] Step S232: The point cloud training data in each raster point cloud data set is respectively fitted with a plane, and the type of the raster point cloud data set is determined according to the fitted plane. The type of the raster point cloud data set includes a ground point raster point cloud data set and / or a non-ground point raster point cloud data set.

[0077] Step S233: The ground height of the raster point cloud data set is determined based on the type of the raster point cloud data set.

[0078] In some embodiments, step S231 can specifically be rasterizing the point cloud training data in a plane parallel to the real ground. After rasterization, a plurality of grids are obtained, and each grid can correspond to a raster point cloud data set. Steps S231 and S22 can use the same way to rasterize the point cloud training data.

[0079] In some embodiments, the principal component analysis method or the RANSAC algorithm can be used to fit the plane of the point cloud training data in each raster point cloud data set.

[0080] In some embodiments, the type of the raster point cloud data set determined according to the fitted plane in step S232 can include:

[0081] The normal of the plane is obtained.

[0082] It is determined whether the included angle between the normal of the plane and the normal of the real ground is less than a preset included angle threshold.

[0083] If yes, the type of the raster point cloud data set is determined as a ground point raster point cloud data set; if no, the type of the raster point cloud data set is determined as a non-ground point raster point cloud data set.

[0084] In other embodiments, the type of the raster point cloud data set determined according to the fitted plane in step S232 can include:

[0085] obtaining a normal of the plane;

[0086] determining whether an included angle between the normal of the plane and the normal of the real ground surface is less than a preset included angle threshold;

[0087] if yes, determining an average value or a variance of the plurality of point cloud training data in the grid point cloud data set projected in a direction perpendicular to the real ground surface;

[0088] when the average value of the plurality of point cloud training data projected in the direction perpendicular to the real ground surface is less than or equal to a first preset threshold, or the variance of the plurality of point cloud training data projected in the direction perpendicular to the real ground surface is less than or equal to a second preset threshold, determining that the type of the grid point cloud data set is a ground point grid point cloud data set;

[0089] when the average value of the plurality of point cloud training data projected in the direction perpendicular to the real ground surface is greater than the first preset threshold, or the variance of the plurality of point cloud training data projected in the direction perpendicular to the real ground surface is greater than the second preset threshold, determining that the type of the grid point cloud data set is a non-ground point grid point cloud data set.

[0090] In some embodiments, step S233 can be specifically:

[0091] when the type of the grid point cloud data set is the ground point grid point cloud data set, determining the ground height true value of the ground point grid point cloud data set according to the coordinate data of the point cloud training data in the ground point grid point cloud data set;

[0092] when the type of the grid point cloud data set is the non-ground point grid point cloud data set, determining the ground height true value of the non-ground point grid point cloud data set based on the ground point grid point cloud data set satisfying the preset distance condition and the preset orientation condition with the non-ground point grid point cloud data set.

[0093] In some embodiments, when the type of the grid point cloud data set is the ground point grid point cloud data set, the coordinate value of the point cloud data projected in the direction perpendicular to the real ground surface can be determined according to the coordinate data of the point cloud training data in the ground point grid point cloud data set; the ground height true value of the corresponding ground point grid point cloud data set is determined based on the average value of the coordinate values corresponding to all point cloud data in the ground point grid point cloud data set.

[0094] In some embodiments, when the type of the grid point cloud data set is the non-ground point grid point cloud data set, an interpolation method can be used to determine the ground height true value of the non-ground point grid point cloud data set based on the ground point grid point cloud data set satisfying the preset distance condition and the preset orientation condition with the non-ground point grid point cloud data set.

[0095] In some embodiments, the ground height true value of the non-ground point grid point cloud dataset can be determined based on the ground point grid point cloud dataset satisfying the preset distance condition and the preset orientation condition with the non-ground point grid point cloud dataset, which can include:

[0096] Selecting, based on the grid corresponding to the grid point cloud dataset, the ground point grid point cloud dataset closest to the non-ground point grid point cloud dataset and located at four different orientations of the non-ground point grid point cloud dataset.

[0097] Obtaining the ground height true value of the ground point grid point cloud dataset at the four different orientations.

[0098] Performing weighted summation and averaging based on the ground height true value of the ground point grid point cloud dataset at the four different orientations and the weight corresponding to each orientation ground point grid point cloud dataset to obtain the ground height true value of the non-ground point grid point cloud dataset, wherein the weight corresponding to each orientation ground point grid point cloud dataset is inversely proportional to the distance between the ground point grid point cloud dataset and the non-ground point grid point cloud dataset.

[0099] In the embodiments of the present application, one grid point cloud dataset corresponds to one grid, and the ground point grid point cloud dataset satisfying the preset distance condition and the preset orientation condition with the non-ground point grid point cloud dataset can be selected by the distance between the grids and the orientation relationship between the grids.

[0100] As an example, when the grid division is performed on the point cloud training data in a plane parallel to the real ground, the four different orientations can be the front, rear, left and right orientations of the grid corresponding to the non-ground point grid point cloud dataset in the plane for the above grid division.

[0101] In some embodiments, step S24 can be specifically calculating the loss function of the ground predicted height and the ground height true value by the sub-network based on the loss function, and training the initial ground height acquisition model based on the loss function.

[0102] In some embodiments, when the grid division is performed on the obtained point cloud training data in steps S22 and S23, the ground predicted height of the grid point cloud dataset and the ground height true value of the grid point cloud dataset can be obtained accordingly, and in step S24, the loss function can be calculated according to the ground predicted height and the ground height true value of each grid point cloud dataset.

[0103] The embodiment of the present application provides a drivable area determination method, point cloud data is input into a neural network model, ground height is obtained based on a branch ground height acquisition model of the neural network model; and the drivable area is determined according to at least the ground height and coordinate data of the point cloud data. The ground height acquisition model is obtained by calculating a loss function of ground predicted height and ground height true value, and training an initial ground height acquisition model based on the loss function, and the ground height true value is obtained by fitting the ground height of point cloud training data. This method effectively avoids determining the ground height true value by labeling information, thereby saving the labeling cost, and effectively improving the convenience and generalization ability of determining the drivable area.

[0104] In other embodiments, the neural network model can also include an obstacle detection model, in order to reduce the deployment time of the overall neural network model and realize the lightweight of the neural network model, the obstacle detection model and the ground height acquisition model share a backbone network, and the obstacle detection model and the ground height acquisition model are obtained by synchronous training.

[0105] When the neural network model includes the obstacle detection model and the ground height acquisition model sharing the backbone network, the training of the neural network model can be: inputting the point cloud training data into the neural network model; performing feature extraction based on the backbone network to obtain feature data; training the obstacle detection model based on the feature data and training the ground height acquisition model based on the feature data.

[0106] Correspondingly, the embodiment of the present application also provides another drivable area determination method, as shown in Figure 5 which can include:

[0107] Step S31: collecting point cloud data of a driving environment;

[0108] Step S32: inputting the point cloud data into a trained neural network model, the neural network model including a ground height acquisition model and an obstacle detection model sharing a backbone network, obtaining ground height of the point cloud data based on the ground height acquisition model and obtaining obstacle detection results based on the obstacle detection model;

[0109] Step S33: determining the drivable area according to the obstacle detection results and the ground height and coordinate data of the point cloud data.

[0110] In step S32, the ground height of the point cloud data can be obtained based on the ground height acquisition model in the same way as step S12.

[0111] In some embodiments, step S33 can be specifically: determining the non-collision point cloud data according to the ground height of the point cloud data and the coordinate data of the point cloud data; fusing the determined non-collision point cloud data and the obstacle detection result to determine the non-travelable area; determining the travelable area based on the non-travelable area, that is, determining the area except the non-travelable area as the travelable area.

[0112] The above is another travelable area determination method provided by the embodiments of the present application, which can achieve the same beneficial effects as the corresponding embodiments, and in addition, determining the travelable area based on the obstacle detection result and the ground height of the point cloud data and the coordinate data of the point cloud data can also improve the accuracy of determining the travelable area. Figure 1

[0113] Another aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the travelable area determination method in any of the above embodiments. The computer readable storage medium can be a storage device formed by various electronic devices, and optionally, the computer readable storage medium in the embodiments of the present application is a non-transitory computer readable storage medium.

[0114] Another aspect of the present application also provides an electronic device, referring to Figure 6 , and Figure 6 The electronic device provided by the embodiments of the present application is shown in the structural schematic diagram, which can include a memory 61 and a processor 62, the memory 61 stores a computer program, and the computer program is executed by the processor 62 to implement the travelable area determination method in any of the above embodiments.

[0115] Another aspect of the present application also provides a driving device, which includes a driving device body, a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to implement the travelable area determination method in any of the above embodiments.

[0116] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will fall within the protection scope of the present application.​

Claims

1. A travelable area determination method characterized by comprising: The method comprises: collecting point cloud data of a driving environment; inputting the point cloud data into a trained neural network model, the neural network model comprising at least a ground height acquisition model, and obtaining a ground height of the point cloud data based on the ground height acquisition model; determining a drivable area according to at least the ground height of the point cloud data and coordinate data of the point cloud data; wherein the ground height acquisition model is trained by: obtaining point cloud training data; inputting the point cloud training data into an initial ground height acquisition model to obtain a ground predicted height of the point cloud training data; ground height fitting is performed on the point cloud training data to obtain ground truth of the point cloud training data; calculating a loss function of the ground predicted height and the ground truth, and training the initial ground height acquisition model based on the loss function to obtain the trained ground height acquisition model.

2. The method of claim 1, wherein, The ground height fitting on the point cloud training data to obtain the ground truth of the point cloud training data comprises: grid division is performed on the point cloud training data to obtain a plurality of divided grid point cloud data sets; plane fitting is performed on the point cloud training data in each grid point cloud data set, and the type of the grid point cloud data set is determined according to the fitted plane, the type of the grid point cloud data set comprising a ground point grid point cloud data set and / or a non-ground point grid point cloud data set; the ground truth of the grid point cloud data set is determined based on the type of the grid point cloud data set.

3. The method of claim 2, wherein, The determination of the ground truth of the grid point cloud data set based on the type of the grid point cloud data set comprises: when the type of the grid point cloud data set is the ground point grid point cloud data set, the ground truth of the ground point grid point cloud data set is determined according to the coordinate data of the point cloud training data in the ground point grid point cloud data set; when the type of the grid point cloud data set is the non-ground point grid point cloud data set, the ground truth of the non-ground point grid point cloud data set is determined based on the ground point grid point cloud data set satisfying a preset distance condition and a preset orientation condition with the non-ground point grid point cloud data set.

4. The method of claim 3, wherein, The determination of the ground truth of the non-ground point grid point cloud data set based on the ground point grid point cloud data set satisfying the preset distance condition and the preset orientation condition with the non-ground point grid point cloud data set comprises: selecting the ground point grid point cloud data set closest to the non-ground point grid point cloud data set and located at four different orientations of the non-ground point grid point cloud data set; obtaining the ground truth of the ground point grid point cloud data set at the four different orientations; The ground height true value of the non-ground point grid point cloud data set is obtained by weighted summation and averaging of the ground height true values of the ground point grid point cloud data sets in four different orientations and the weights corresponding to the ground point grid point cloud data sets in each orientation, wherein the weight corresponding to the ground point grid point cloud data set in each orientation is inversely proportional to the distance between the ground point grid point cloud data set and the non-ground point grid point cloud data set.

5. The method of claim 2, wherein, The type of the grid point cloud data set is determined according to the fitted plane, including: obtaining the normal of the plane; determining whether the angle between the normal of the plane and the normal of the real ground is less than a preset angle threshold; if yes, determining that the type of the grid point cloud data set is a ground point grid point cloud data set; and if no, determining that the type of the grid point cloud data set is a non-ground point grid point cloud data set.

6. The method of claim 2, wherein, The type of the grid point cloud data set is determined according to the fitted plane, including: obtaining the normal of the plane; determining whether the angle between the normal of the plane and the normal of the real ground is less than a preset angle threshold; if yes, determining the average value or variance of the plurality of point cloud training data in the grid point cloud data set projected in the direction perpendicular to the real ground; when the average value of the plurality of point cloud training data projected in the direction perpendicular to the real ground is less than or equal to a first preset threshold, or the variance of the plurality of point cloud training data projected in the direction perpendicular to the real ground is less than or equal to a second preset threshold, determining that the type of the grid point cloud data set is a ground point grid point cloud data set; when the average value of the plurality of point cloud training data projected in the direction perpendicular to the real ground is greater than the first preset threshold, or the variance of the plurality of point cloud training data projected in the direction perpendicular to the real ground is greater than the second preset threshold, determining that the type of the grid point cloud data set is a non-ground point grid point cloud data set.

7. The method of claim 1, wherein, The drivable area is determined at least according to the ground height of the point cloud data and the coordinate data of the point cloud data, including: determining the coordinate value of the point cloud data projected in the direction perpendicular to the real ground according to the coordinate data of the point cloud data; screening the point cloud data with a coordinate value greater than the ground height and taking the point cloud data as non-collision point cloud data; determining a drivable area based on the non-collision point cloud data.

8. The method of claim 1, wherein, The neural network model further includes an obstacle detection model, the obstacle detection model and the ground height acquisition model share a main network, and the obstacle detection model and the ground height acquisition model are obtained by synchronous training; after inputting the point cloud data into the neural network model, the method further includes: obtaining an obstacle detection result based on the obstacle detection model; The drivable area is determined at least according to the ground height of the point cloud data and the coordinate data of the point cloud data, including: determining a drivable area according to the obstacle detection result and the ground height and the coordinate data of the point cloud data.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program, when executed by a processor, implements the drivable area determination method in any one of claims 1 to 8.

10. A driving apparatus characterized by comprising: The driving device comprises a driving device body, a memory, and a processor, the memory stores a computer program, and the computer program, when executed by the processor, implements the drivable area determination method in any one of claims 1 to 8.

Images