Two-dimensional grid environment modeling method based on depth estimation method

Abstract

The invention discloses a two-dimensional grid environment modeling method based on a depth estimation method, and the method comprises the steps: S1, environment perception and data collection: scanning the surrounding environment of a vehicle through a laser radar, and carrying out the repeated scanning, and collecting a large amount of point cloud data, thereby forming an environment point cloud; s2, data processing: cleaning the original point cloud data by using a digital filter, removing noise points and abnormal values, extracting useful information from the point cloud, and reconstructing a three-dimensional model based on a surface reconstruction algorithm by using the point cloud data; s3, model conversion: converting geometric attributes of the three-dimensional model from an overlook angle by adopting a mathematical mapping principle, establishing a navigation point grid according to a navigation point theory, performing accurate modeling on a navigation point model on a two-dimensional plane by applying a computational geometry method, and adjusting model parameters through a particle swarm optimization algorithm; and S4, generating a depth map. According to the method, the calculation amount is reduced, and the calculation time is effectively saved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned driving technology, and in particular to a two-dimensional grid environment modeling method based on a depth estimation method. Background Art

[0002] With the rapid advancement of science and technology, driverless vehicle technology has become a global research focus. One of the core technologies of driverless vehicles is path planning, which is directly related to whether the vehicle can drive safely and efficiently under various road conditions. Path planning technology needs to solve how to plan an optimal or feasible driving path for driverless vehicles in complex traffic environments and uncertain dynamic factors.

[0003] At present, there are many different methods for path planning of unmanned vehicles, among which the more common ones are graph search-based methods, such as A* algorithm, Dijkstra algorithm, etc. These methods search for the optimal path from the starting point to the end point by constructing a graph model of the environment map; sampling-based methods, such as RRT (rapid random tree) algorithm, explore feasible paths by randomly sampling in the configuration space; and optimization-based methods, such as using genetic algorithms, particle swarm optimization and other optimization techniques to find solutions for path planning. Although these methods are theoretically feasible, in practical applications, especially when facing complex and changing environments and high-dimensional planning spaces, they often show some shortcomings. For example, graph search-based methods may cause a sharp increase in the amount of calculation due to the large state space, and the real-time performance is affected; sampling-based methods may find it difficult to find an effective path in a complex environment, and the quality of the path is often difficult to guarantee; and optimization-based methods may be difficult to obtain a satisfactory solution within a limited time due to the complexity of the optimization process.

[0004] As an innovative path planning strategy, the core of the grid method is to decompose the entire path planning problem into multiple relatively simple sub-problems. Specifically, this method first divides the entire planning space where the vehicle is located into several subspaces, which can be divided based on environmental characteristics, road structure or other vehicle behavior patterns. In each subspace, the unmanned vehicle can independently perform local path planning. Such a planning process is more centralized, the computational complexity is relatively low, and it is easier to consider the special constraints in a specific subspace. After completing the local path planning, these local paths are merged into a complete optimal or feasible global path through a certain splicing strategy. In this way, the grid method not only effectively reduces the computational complexity of the overall path planning, but also greatly improves the quality of path planning. However, the grid method also has obvious disadvantages. A large amount of edge data is included in the modeling process of the grid method. The impact of these environmental data on path planning can be ignored, but it will greatly increase the amount of calculation.

[0005] Therefore, in order to solve the above problems, a two-dimensional grid environment modeling method based on depth estimation method is provided. Summary of the invention

[0006] The purpose of the present invention is to overcome the existing defects and provide a two-dimensional grid environment modeling method based on a depth estimation method, which reduces the amount of calculation and effectively saves calculation time.

[0007] The technical solution to achieve the above purpose is:

[0008] A two-dimensional grid environment modeling method based on a depth estimation method, comprising:

[0009] Step S1, environmental perception and data collection: using laser radar to scan the environment around the vehicle, repeatedly scanning and collecting a large amount of point cloud data to form an environmental point cloud;

[0010] Step S2, data processing: cleaning the original point cloud data with a digital filter, removing noise points and outliers, extracting useful information from the point cloud, and reconstructing the three-dimensional model based on the point cloud data based on the surface reconstruction algorithm;

[0011] Step S3, model conversion: the geometric attributes of the three-dimensional model are converted from a bird's-eye view using the mathematical mapping principle, a navigation point grid is established based on the navigation point theory, the navigation point model on the two-dimensional plane is accurately modeled using computational geometry methods, and the model parameters are adjusted using a particle swarm optimization algorithm;

[0012] Step S4, generating a depth map: the depth of each point is calculated according to the flight time and reflection intensity of the laser pulse to convert the two-dimensional model into a depth map;

[0013] Step S5, data screening: using the threshold segmentation method, setting a color depth threshold, and removing the part of the depth map with a color depth lower than the threshold;

[0014] Step S6, two-dimensional grid environment modeling: finely divide the processed depth map and convert it into a series of grid units of equal size;

[0015] Step S7, identifying traversable areas: identifying obstacles in the processed depth map and marking these cells as impassable in the two-dimensional grid model accordingly, thereby determining traversable cells, i.e., areas where the driverless vehicle can safely travel.

[0016] Preferably, in step S1, a laser radar is installed in front of the unmanned vehicle to perform environment perception and data collection, including:

[0017] Step S11, the laser radar system emits a series of laser pulses to the target area;

[0018] Step S12, when the laser pulse encounters surrounding objects, it will be reflected, scattered or absorbed;

[0019] Step S13, the receiver in the laser radar system captures the laser pulse reflected from the object and records the arrival time and intensity of the reflected photon;

[0020] Step S14, based on the time interval between transmitting and receiving the reflected light and the known propagation speed of light in the air, the laser radar calculates the time required for the pulse to go back and forth, thereby determining the distance of the object;

[0021] In step S15, the laser radar system scans the surrounding environment by quickly repeating the above process, records each reflection point, and corresponds it to the three-dimensional coordinates. These coordinate points are collected together to finally form point cloud data.

[0022] Preferably, the step S2 comprises:

[0023] Step S21, firstly, the point cloud data collected by the laser radar is processed by a digital filter to eliminate random noise and abnormal fluctuations caused by environmental factors or the sensor itself, while maintaining the overall shape characteristics of the point cloud data;

[0024] Step S22, then extracting surface normal, curvature, and shape descriptor features from the point cloud to extract useful information for subsequent recognition and classification;

[0025] Step S23, finally reconstructing the three-dimensional model using the point cloud data through a surface reconstruction algorithm.

[0026] Preferably, step S3 comprises:

[0027] Step S31, firstly, the geometric attributes of the three-dimensional model are converted at a top-down angle using a mathematical mapping principle;

[0028] Step S32, using a feature recognition algorithm to ensure important geometric and topological features of the three-dimensional model during the dimensionality reduction process;

[0029] Step S33, then establishing a navigation point grid on the two-dimensional plane according to the navigation point theory;

[0030] Step S34, finally, the navigation point model on the two-dimensional plane is accurately modeled using computational geometry methods, and the model parameters are adjusted using a particle swarm optimization algorithm.

[0031] Preferably, in step S4, when the laser radar emits a laser pulse, the emission time t1 of the laser pulse and the time t2 when the laser pulse returns and is received by the sensor are recorded, and then the time difference Δt=t1-t2 is calculated, and then the formula: distance=(speed of light*Δt) / 2 is used to calculate the distance of each point, and the laser radar is used to scan each point to generate a depth map.

[0032] Preferably, in step S5, a threshold segmentation technique is used to screen the data in the depth map, eliminating darker point cloud data, i.e., point cloud data closer to the vehicle, and retaining lighter point cloud data, i.e., point cloud data farther from the vehicle.

[0033] Preferably, in step S6, the depth map after data elimination is subjected to two-dimensional grid modeling. First, the modeling boundary is determined, and then the grid size is determined according to the size of the vehicle, the sensor accuracy and the accuracy requirements of the path planning. Finally, the depth map is divided into grid units of the same size.

[0034] Preferably, the step S7 comprises:

[0035] Step S71, extracting features of obstacles from the depth map, and classifying obstacles using a convolutional neural network to distinguish between static and dynamic obstacles;

[0036] Step S72, marking the cells where the static obstacles are located as inaccessible in the two-dimensional grid model, indicating that vehicles are prohibited from entering these areas;

[0037] Step S73, using a motion prediction algorithm to predict the future position of the dynamic obstacle, and marking it accordingly in the grid model, thereby determining the navigable cells, that is, the area where the driverless vehicle can drive safely.

[0038] Preferably, in step S71, the characteristics of the obstacle include but are not limited to shape, size, position and speed.

[0039] The beneficial effects of the present invention are:

[0040] The present invention scans the surrounding environment of the vehicle through a laser radar, can accurately record each reflection point, and form a point cloud map corresponding to the three-dimensional coordinates;

[0041] During data processing, the present invention performs a series of cleaning and optimization operations on the collected raw data, and uses a digital filter for processing to eliminate random noise and abnormal fluctuations introduced by environmental factors or the sensor itself. In the filtering process, a Gaussian filtering algorithm is selected to ensure that the original structure and overall shape characteristics of the point cloud data are kept to the maximum extent while removing noise. On the basis of the filtered point cloud data, geometric features such as surface normal, curvature, and shape descriptors are extracted to perform feature extraction, providing a rich information basis for subsequent point cloud recognition and classification. Through a surface reconstruction algorithm, the three-dimensional model is reconstructed using the filtered and feature-extracted point cloud data. According to the points in the point cloud and their feature information, a continuous and smooth surface model is constructed to obtain an accurate three-dimensional model, which retains the basic shape and structure of the original point cloud data and reflects the surface details of the object to a certain extent.

[0042] The present invention adopts the mathematical mapping principle to transform the geometric properties of the three-dimensional model, ensuring that the key geometric features in the three-dimensional model can be accurately mapped on the two-dimensional plane. In the process of dimensionality reduction, the feature recognition algorithm is used to ensure that the important geometric and topological features in the three-dimensional model can be retained on the two-dimensional plane. The feature recognition algorithm is used to conduct in-depth analysis of the three-dimensional model, identify important feature points, and process them during the conversion process, ensuring that these features will not be lost or deformed in the two-dimensional mapping. According to the navigation point theory, the computational geometry method is used to accurately model the navigation point model on the two-dimensional plane, achieving the best spatial representation effect, and further using the particle swarm optimization algorithm to adjust the model parameters. Through continuous iteration and optimization, a precise and efficient two-dimensional model is finally obtained.

[0043] The present invention uses a threshold segmentation technique to filter the data in the generated depth map, which reduces the amount of data processing and focuses on close obstacles that are more critical to vehicle driving safety;

[0044] In the two-dimensional grid environment modeling stage, the present invention determines the modeling boundary, which defines the area covered by the grid model. When determining the boundary, the actual driving requirements of the vehicle, the detection range of the sensor and the expected driving scene are considered to ensure that the grid model can cover all relevant environmental information. A convolutional neural network is used to classify the extracted obstacle features. CNN can accurately output the category label of the obstacle by learning a large amount of labeled data. After the obstacle classification result is obtained, it is applied to the two-dimensional grid model, and the cell where the identified static obstacle is located is marked as impassable. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a two-dimensional grid environment modeling method based on a depth estimation method of the present invention;

[0046] Figure 2 It is a specific flow chart of installing a laser radar in front of an unmanned vehicle for environment perception and data collection in the present invention;

[0047] Figure 3 It is a specific flow chart of the present invention for cleaning the original point cloud data with a digital filter, removing noise points and outliers, extracting useful information from the point cloud, and reconstructing a three-dimensional model based on a surface reconstruction algorithm using the point cloud data;

[0048] Figure 4 It is a specific flow chart of the present invention for converting the geometric properties of a three-dimensional model from a top-down perspective using the mathematical mapping principle to construct a two-dimensional model;

[0049] Figure 5 It is a specific flow chart of identifying obstacles in the processed depth map and marking these cells as impassable in the two-dimensional grid model accordingly, and then determining the impassable cells in the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0051] The present invention will be further described below in conjunction with the accompanying drawings.

[0052] like Figure 1 As shown, a two-dimensional grid environment modeling method based on depth estimation method includes:

[0053] Step S1, environmental perception and data collection: Use LiDAR to scan the environment around the vehicle, and repeatedly scan to collect a large amount of point cloud data to form an environmental point cloud.

[0054] like Figure 2 As shown, a laser radar is installed in front of the unmanned vehicle for environmental perception and data collection, including:

[0055] In step S11 , the laser radar system emits a series of laser pulses toward the target area.

[0056] In the embodiment, these laser pulses have a specific wavelength to ensure their propagation efficiency and detection accuracy in the environment; at the same time, the duration of these pulses is extremely short, usually at the nanosecond level, so a large amount of environmental information can be obtained in a very short time.

[0057] Step S12: When the laser pulse encounters surrounding objects, it will be reflected, scattered or absorbed.

[0058] In an embodiment, when these laser pulses encounter surrounding obstacles, such as buildings, roadside trees, other vehicles or pedestrians, a series of physical phenomena occur, including reflection, scattering and absorption.

[0059] In step S13, the receiver in the laser radar system, i.e., the photodetector, will keenly capture the laser pulse reflected from the object and record the arrival time and intensity of the reflected photon.

[0060] In step S14, based on the time interval between transmitting and receiving the reflected light and the known propagation speed of light in the air, the laser radar calculates the time required for the pulse to travel back and forth, thereby determining the distance of the object.

[0061] In step S15, the laser radar system scans the surrounding environment by quickly repeating the above process, records each reflection point, and corresponds it to the three-dimensional coordinates. These coordinate points are collected together to finally form point cloud data.

[0062] Step S2, data processing: clean the original point cloud data with a digital filter, remove noise points and outliers, extract useful information from the point cloud, and use the point cloud data to reconstruct the three-dimensional model based on the surface reconstruction algorithm.

[0063] like Figure 3 As shown, step S2 includes:

[0064] In step S21, the point cloud data collected by the laser radar is first processed using a digital filter to eliminate random noise and abnormal fluctuations caused by environmental factors or the sensor itself, while maintaining the overall shape characteristics of the point cloud data.

[0065] In step S22, surface normals, curvatures, and shape descriptor features are then extracted from the point cloud to extract useful information for subsequent recognition and classification.

[0066] Step S23, finally, through the surface reconstruction algorithm, the three-dimensional model is reconstructed using the point cloud data, which retains the basic shape and structure of the original point cloud data and reflects the surface details of the object to a certain extent.

[0067] Step S3, model conversion: the geometric properties of the three-dimensional model are converted from a bird's-eye view using the mathematical mapping principle, a navigation point grid is established based on the navigation point theory, the navigation point model on the two-dimensional plane is accurately modeled using computational geometry methods, and the model parameters are adjusted using the particle swarm optimization algorithm.

[0068] like Figure 4 As shown, step S3 includes:

[0069] In step S31, the geometric properties of the three-dimensional model are first converted at a top-down angle using the mathematical mapping principle, thereby ensuring that key geometric features in the three-dimensional model can be accurately mapped on a two-dimensional plane.

[0070] Step S32, in the process of dimensionality reduction, feature recognition algorithms are used to ensure important geometric and topological features of the three-dimensional model. These features include but are not limited to the edges, corners, surface continuity, etc. of the model. The three-dimensional model is deeply analyzed using feature recognition algorithms to identify important feature points, which are processed during the conversion process to ensure that these features are not lost or deformed in the two-dimensional mapping.

[0071] Step S33, then establishing a navigation point grid on the two-dimensional plane according to the navigation point theory.

[0072] Step S34, finally, the navigation point model on the two-dimensional plane is accurately modeled using computational geometry methods, and the model parameters are adjusted using a particle swarm optimization algorithm, ultimately obtaining a two-dimensional model that is both accurate and efficient.

[0073] Step S4, generate a depth map: calculate the depth of each point according to the flight time and reflection intensity of the laser pulse to convert the two-dimensional model into a depth map.

[0074] In an embodiment, when the laser radar emits a laser pulse, the emission time t1 of the laser pulse and the time t2 when the laser pulse returns and is received by the sensor are recorded, and then the time difference Δt=t1-t2 is calculated, and then the formula: distance=(speed of light*Δt) / 2 is used to calculate the distance of each point, and the laser radar is used to scan each point to generate a depth map.

[0075] Step S5, data screening: adopting the threshold segmentation method, setting the color depth threshold, and removing the part of the depth map whose color depth is lower than the threshold.

[0076] In an embodiment, a threshold segmentation technique is used to filter the data in the depth map, remove darker point cloud data, and retain lighter point cloud data, that is, the color depth corresponds to the distance information in the depth map, wherein the darker area represents an obstacle closer to the vehicle, and the lighter area represents an obstacle farther from the vehicle; specifically, each pixel in the depth map is traversed to check whether its color depth meets the set threshold condition, and the lighter pixel, that is, the point cloud data representing the point cloud data farther from the vehicle, is removed from the data set, in order to reduce the amount of data processing and focus on the close obstacles that are more critical to the vehicle's driving safety, and the darker pixel, that is, the point cloud data representing the point cloud data closer to the vehicle, is retained in the data set.

[0077] Step S6, two-dimensional grid environment modeling: finely divide the processed depth map and convert it into a series of grid units of equal size.

[0078] In the embodiment, the depth map after data elimination is subjected to two-dimensional grid modeling. First, the boundary of the modeling is determined. The boundary defines the area covered by the grid model. This area includes all possible driving spaces around the vehicle and the environment within a certain distance in front of the vehicle. When determining the boundary, it is necessary to consider the actual driving needs of the vehicle, the detection range of the sensor, and the expected driving scenarios to ensure that the grid model can cover all relevant environmental information; secondly, the grid size is determined according to the size of the vehicle, the accuracy of the sensor, and the accuracy requirements of the path planning, and finally the depth map is divided into grid units of the same size.

[0079] Step S7, identifying traversable areas: identifying obstacles in the processed depth map and marking these cells as impassable in the two-dimensional grid model accordingly, thereby determining traversable cells, i.e., areas where the driverless vehicle can safely travel.

[0080] like Figure 5 As shown, step S7 includes:

[0081] Step S71, extracting the features of obstacles from the depth map, and using a convolutional neural network to classify the obstacles and distinguish between static and dynamic obstacles. The convolutional neural network can accurately output the category labels of the obstacles by learning a large amount of labeled data.

[0082] In an embodiment, characteristics of an obstacle include, but are not limited to, shape, size, position, and speed.

[0083] Step S72, marking the cells where the static obstacles are located as impassable in the two-dimensional grid model, indicating that vehicles are prohibited from entering these areas.

[0084] Step S73, using a motion prediction algorithm to predict the future position of the dynamic obstacle, and marking it accordingly in the grid model, thereby determining the navigable cells, that is, the area where the driverless vehicle can drive safely.

[0085] In the embodiment, for dynamic obstacles, their positions need to be predicted, which is achieved through a motion prediction algorithm. The algorithm takes into account factors such as the current speed, acceleration, direction, etc. of the obstacles to estimate their positions in the future. The predicted future positions of the dynamic obstacles are marked in a two-dimensional grid model, and the remaining grids are the traversable areas for unmanned vehicles.

[0086] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some or all of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A two-dimensional grid environment modeling method based on depth estimation method, characterized in that: include: Step S1, environmental perception and data collection: using laser radar to scan the environment around the vehicle, repeatedly scanning and collecting a large amount of point cloud data to form an environmental point cloud; Step S2, data processing: cleaning the original point cloud data with a digital filter, removing noise points and outliers, extracting useful information from the point cloud, and reconstructing the three-dimensional model based on the point cloud data based on the surface reconstruction algorithm; Step S3, model conversion: the geometric attributes of the three-dimensional model are converted from a bird's-eye view using the mathematical mapping principle, a navigation point grid is established based on the navigation point theory, the navigation point model on the two-dimensional plane is accurately modeled using computational geometry methods, and the model parameters are adjusted using a particle swarm optimization algorithm; Step S4, generating a depth map: the depth of each point is calculated according to the flight time and reflection intensity of the laser pulse to convert the two-dimensional model into a depth map; Step S5, data screening: using the threshold segmentation method, setting a color depth threshold, and removing the part of the depth map with a color depth lower than the threshold; Step S6, two-dimensional grid environment modeling: finely divide the processed depth map and convert it into a series of grid units of equal size; Step S7, identifying traversable areas: identifying obstacles in the processed depth map and marking these cells as impassable in the two-dimensional grid model accordingly, thereby determining traversable cells, i.e., areas where the driverless vehicle can safely travel.

2. A two-dimensional grid environment modeling method based on depth estimation method according to claim 1, characterized in that: In step S1, a laser radar is installed in front of the unmanned vehicle for environment perception and data collection, including: Step S11, the laser radar system emits a series of laser pulses to the target area; Step S12, when the laser pulse encounters surrounding objects, it will be reflected, scattered or absorbed; Step S13, the receiver in the laser radar system captures the laser pulse reflected from the object and records the arrival time and intensity of the reflected photon; Step S14, based on the time interval between the emission and reception of the reflected light and the known propagation speed of light in the air, the laser radar calculates the time required for the pulse to travel back and forth, thereby determining the distance of the object; In step S15, the laser radar system scans the surrounding environment by quickly repeating the above process, records each reflection point, and corresponds it to the three-dimensional coordinates. These coordinate points are collected together to finally form point cloud data.

3. A two-dimensional grid environment modeling method based on depth estimation method according to claim 2, characterized in that: The step S2 comprises: Step S21, firstly, the point cloud data collected by the laser radar is processed by a digital filter to eliminate random noise and abnormal fluctuations caused by environmental factors or the sensor itself, while maintaining the overall shape characteristics of the point cloud data; Step S22, then extracting surface normal, curvature, and shape descriptor features from the point cloud to extract useful information for subsequent recognition and classification; Step S23, finally reconstructing the three-dimensional model using the point cloud data through a surface reconstruction algorithm.

4. A two-dimensional grid environment modeling method based on depth estimation method according to claim 3, characterized in that: The step S3 comprises: Step S31, firstly, the geometric attributes of the three-dimensional model are converted at a top-down angle using a mathematical mapping principle; Step S32, using a feature recognition algorithm to ensure important geometric and topological features of the three-dimensional model during the dimensionality reduction process; Step S33, then establishing a navigation point grid on the two-dimensional plane according to the navigation point theory; Step S34, finally, the navigation point model on the two-dimensional plane is accurately modeled using computational geometry methods, and the model parameters are adjusted using a particle swarm optimization algorithm.

5. A two-dimensional grid environment modeling method based on depth estimation method according to claim 4, characterized in that: In step S4, when the laser radar emits a laser pulse, the emission time t1 of the laser pulse and the time t2 when the laser pulse returns and is received by the sensor are recorded, and then the time difference Δt=t1-t2 is calculated, and then the formula: distance=(speed of light*Δt) / 2 is used to calculate the distance of each point, and the laser radar is used to scan each point to generate a depth map.

6. A two-dimensional grid environment modeling method based on depth estimation method according to claim 5, characterized in that: In step S5, a threshold segmentation technique is used to screen the data in the depth map, remove darker point cloud data, i.e., point cloud data closer to the vehicle, and retain lighter point cloud data, i.e., point cloud data farther from the vehicle.

7. A two-dimensional grid environment modeling method based on depth estimation method according to claim 6, characterized in that: In step S6, the depth map after data elimination is subjected to two-dimensional grid modeling. First, the modeling boundary is determined, and then the grid size is determined according to the size of the vehicle, the sensor accuracy and the accuracy requirements of the path planning. Finally, the depth map is divided into grid units of the same size.

8. A two-dimensional grid environment modeling method based on depth estimation method according to claim 7, characterized in that: The step S7 comprises: Step S71, extracting features of obstacles from the depth map, and classifying obstacles using a convolutional neural network to distinguish between static and dynamic obstacles; Step S72, marking the cells where the static obstacles are located as inaccessible in the two-dimensional grid model, indicating that vehicles are prohibited from entering these areas; Step S73, using a motion prediction algorithm to predict the future position of the dynamic obstacle, and marking it accordingly in the grid model, thereby determining the navigable cells, that is, the area where the driverless vehicle can drive safely.

9. A two-dimensional grid environment modeling method based on depth estimation method according to claim 8, characterized in that: In step S71, the characteristics of the obstacle include but are not limited to shape, size, position and speed.

Images