Negative obstacle joint detection and depth estimation method in cross-country environment

Abstract

The invention relates to the technical field of terrain surface modeling based on multi-source laser radar fusion, and particularly discloses a negative obstacle joint detection and depth estimation method in a cross-country environment, comprising the following steps: establishing a ground model based on a polar coordinate grid map; establishing a terrain segmentation method based on Gaussian process regression fitting; establishing a negative obstacle detection algorithm based on the ground reference height; establishing a negative obstacle point cloud clustering algorithm based on a radial bounded nearest neighbor graph; establishing a false alarm reducing method based on the trench crossing width and the trench crossing depth of the vehicle; according to the method, the modeling complexity and the computational complexity are reduced, and the segmentation precision of the cross-country undulating pavement is improved by utilizing the characteristics of Gaussian distribution; then, a negative obstacle point cloud detection method based on the grid reference height is provided, and a negative obstacle is detected in real time; finally, a negative obstacle false alarm reducing method is provided in combination with the moat crossing height and depth of the vehicle, and the trafficability of the vehicle is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of terrain and surface modeling based on multi-source lidar fusion, and specifically relates to a method for joint detection and depth estimation of negative obstacles in an off-road environment. Background Art

[0002] With the development of sensor technology, information fusion technology, and artificial intelligence technology, the mode of war has begun to transform from information-based war to intelligent war. The typical feature of the intelligence of land equipment is the unmanned operation of vehicles, mainly because land military operations involve a large number of personnel and complex environments, and often face a high risk of personnel casualties. Therefore, realizing the unmanned autonomous driving of vehicles in high-risk areas during wartime is one of the important ways to reduce personnel casualties and is also an important field that various military powers are competing to research.

[0003] There are the following difficulties in the joint detection and depth estimation of negative obstacles based on the centralized fusion of multi-source lidar: (1) The detection distance of negative obstacles is proportional to the installation height of the lidar and the longitudinal width of the negative obstacle, and inversely proportional to the depth of the pit. It is necessary to develop corresponding algorithms in combination with the ability of the unmanned vehicle to climb negative obstacles. Therefore, the detection threshold of negative obstacles is adaptively adjustable. (2) The types of negative obstacles on off-road surfaces are diverse and of different sizes, so the stability requirements for the algorithm are very high. (3) The heights of negative obstacles are inconsistent: for the same negative obstacle, the flatness of the bottom is inconsistent, and it is necessary to estimate its depth in real time.

[0004] The above difficulties can be summarized as the problem of joint detection and estimation of negative obstacles based on the centralized fusion of multi-source lidar in off-road complex environments, that is, it is necessary to detect multiple possible negative obstacles in real time and estimate the depth of the negative obstacles at the same time. However, the traditional feature-based negative obstacle detection algorithm does not consider the off-road complex road conditions. The traditional method is a negative obstacle detection algorithm based on the features of point cloud data, which extracts the scan line features of the lidar and then detects negative obstacles according to the features. The following problems exist:

[0005] 1. Without combining the prior information of the trench-crossing width and depth of the vehicle itself, the drivable road surface of the negative obstacle that can be passed is segmented into non-drivable road surfaces, affecting the normal driving of the vehicle.

[0006] 2. The lidar point cloud becomes sparser as the distance increases. Calculating reasonable ground features from the sparse scan lines consumes a huge amount of computing power and thus does not meet the requirements of engineering practice; on the other hand, finding ground features from the already sparse scan lines is a very challenging topic.

[0007] Therefore, in combination with the characteristics of the vehicle itself and using the point cloud information of lidar, this patent proposes a joint detection and estimation algorithm for negative obstacles to solve the problems that traditional methods cannot solve the joint detection and estimation of negative obstacles based on centralized fusion of multi-source lidar and have a high false alarm rate. Summary of the Invention

[0008] Aiming at the problems existing in the prior art, the purpose of the present invention is to provide a method for joint detection and depth estimation of negative obstacles in an off-road environment, which improves the stability of modeling; by projecting point cloud data into a two-dimensional polar coordinate grid map, the data processing process is simplified and the real-time performance of the algorithm is improved. Combining the trench-crossing width and depth of the vehicle itself, a method for reducing false alarms of negative obstacles is proposed to improve the passability of the road surface.

[0009] The technical solution adopted by the present invention to solve its technical problems is: a method for joint detection and depth estimation of negative obstacles in an off-road environment, including the following steps:

[0010] S1. Establish a ground model based on a polar coordinate grid map: project three-dimensional point cloud data into a two-dimensional grid map;

[0011] S2. Establish a terrain segmentation method based on Gaussian process regression fitting: select the lidar point cloud that is relatively close to the vehicle body and has the minimum height as the seed point, and continuously iteratively learn the Gaussian model using the seed point to form a Gaussian ground segmentation model;

[0012] S3. Establish a negative obstacle detection algorithm based on the ground reference height; combine the trained Gaussian ground segmentation model, predict the ground height of each segmented grid, and then judge whether the point cloud is a ground point cloud based on the difference between the measured height and the predicted height of the lidar point cloud;

[0013] S4. Establish a negative obstacle point cloud clustering algorithm based on a radially bounded nearest neighbor graph; cluster the grids with the attribute of negative obstacles, and adopt a clustering method based on a radially bounded nearest neighbor graph;

[0014] S5. Establish a false alarm reduction method based on the trench-crossing width and depth of the vehicle itself; the purpose of negative obstacle detection is to ensure the safe driving of the unmanned vehicle and improve the passability of the road surface, but there is a certain contradiction between the two: if the detection sensitivity of the algorithm is very high, the negative obstacle road surface that the vehicle can pass through may be set as an impassable road surface.

[0015] Preferably, the specific steps of establishing the ground model based on the polar coordinate grid map in step S1 are as follows:

[0016] 1) Non-uniform grid division method:

[0017] The terrain modeling of a larger area does not meet the real-time requirements of an unmanned vehicle. Therefore, the range of the vehicle body 360 is unevenly divided. With the positive direction of the Y-axis, when the angle θ satisfies 0 ≤ θ ≤ 90 and -90 ≤ θ < 0, the division angle is 0.5; when θ satisfies 90 < θ < 270, the division angle is 2. At this time, the total number of sectors is

[0018]

[0019] 2) After the above step 1), the rectangle is divided into 450 uneven sectors. To further reduce the computational complexity of terrain modeling, each sector is further unevenly divided into multiple grids. For sector j and lidar point cloud i, if

[0020]

[0021] where is the radial distance of lidar point cloud i in sector j; k = 1, …, C j ; is the lower bound of grid k, while is the upper bound of grid k; according to the uneven division method, if when, then if when, then

[0022] 3) After steps 1) and 2), the point clouds within the region of interest have all been mapped into the polar coordinate grid map. To further reduce the computational complexity, the three-dimensional data is mapped into two-dimensional data:

[0023]

[0024] Preferably, the specific steps of establishing the terrain segmentation method based on Gaussian process regression fitting in step S2 are as follows:

[0025] 1) Selection of initial seed points: For sector j, assuming the total number of grids it has is C j , for the grid within ρ ≤ 10m, the initial point cloud set composed of the point cloud with the minimum height is

[0026] G j ={p i , i = 0…C j};

[0027] The set composed of all the points with the minimum height in sector j is

[0028]

[0029] 2) Gaussian regression model: For the case where the Bayesian linear regression model is a Gaussian random process and its measurement model with observation noise is \(z = f(x)+v\), where its posterior probability distribution is as follows:

[0030]

[0031] where \(K(\cdot)\) represents the covariance function; \(X\) is the mean of the Gaussian density function; \(\Sigma\) is the covariance of the Gaussian probability density function.

[0032] 3) Iterative training of the Gaussian model: Use the initial seed point set as the training set \(G\) j ; Use all possible seed point sets as the test point set to perform iterative Gaussian model training. The trained model is as follows

[0033]

[0034] For each make a judgment according to the following threshold

[0035]

[0036] If the point satisfies the above threshold conditions, then add the point to the set \(G\) j ;

[0037] 4) Ground model based on Gaussian process regression: Use the trained Gaussian model. For sector \(j\), the total number of grids is \(C\) j .

[0038] The radial distance of grid \(k\) from the ego-vehicle is

[0039]

[0040] Substitute the parameter into the trained Gaussian model, then the predicted height of grid \(k\) can be obtained

[0041]

[0042] According to the pre-set parameter \(T\) Ground_Th , for the point cloud \(p\) i = [x i , y i , z i in this grid, if the absolute value of its difference from the reference height is less than the following threshold

[0043]

[0044] If this point is considered a ground point, otherwise it is considered an obstacle point.

[0045] Preferably, the specific steps of establishing the negative obstacle detection algorithm based on the ground reference height in step S3 are as follows:

[0046] 1) Model and analyze the main distribution of negative obstacles on off-road surfaces. Through the analysis of the negative obstacle point cloud, the relationship between the installation height of the lidar, the detection distance D, and the depth d of the negative obstacles that can be detected is

[0047]

[0048] 2) Based on the constructed grid map, a method for representing the negative obstacle point cloud based on the point cloud reference height is proposed. The operation process is as follows:

[0049] (1) For any grid in the fan-shaped block j (0 ≤ j < M) of the polar coordinate grid map As long as there is a three-dimensional lidar point With the relative height to the reference ground Then this grid is marked as a negative obstacle grid;

[0050] (2) When it is determined that the grid Is a negative obstacle grid, then in the current fan-shaped block, starting from the grid Search forward to find the grid closest to it and marked as the ground All grids in between are marked as negative obstacle grids;

[0051] (3) For the negative obstacle grid, calculate all grids within the 5-grid map neighborhood that have no point cloud data, and they are all marked as negative obstacle grids.

[0052] Preferably, the specific steps of establishing the negative obstacle point cloud clustering algorithm based on the radially bounded nearest neighbor graph in step S4 are as follows:

[0053] For the negative obstacle grid, use the method based on the radially bounded nearest neighbor graph to cluster the grids of the same class. The steps are as follows:

[0054] 1) Traverse all negative obstacle grids;

[0055] 2) If the current negative obstacle grid already belongs to a certain negative obstacle class, skip to the next grid;

[0056] 3) For the currently traversed negative obstacle grid;

[0057] (1) Find all grids in the r-neighborhood;

[0058] (2) If the neighboring point already belongs to a certain negative obstacle class, then classify the current point into the current class;

[0059] (3) If the current point already belongs to a certain negative obstacle class while the negative obstacle class of the neighboring point is different from that of the current point, then class fusion is required.

[0060] Preferably, the specific steps of establishing the false alarm reduction method based on the vehicle's trench crossing width and depth in step S5 are as follows:

[0061] After obtaining the negative obstacle grid, calculate the width of the grid in the longitudinal direction, and calculate the central distance ρ of the negative obstacle grid farthest from the vehicle max and the central point distance ρ of the negative obstacle grid closest to the vehicle body min . If the size of the negative obstacle meets the following two conditions:

[0062]

[0063] then this negative obstacle is real; otherwise, it is false; design the corresponding false alarm reduction method in combination with the vehicle's trench crossing width and depth.

[0064] The present invention has the following beneficial effects:

[0065] The negative obstacle joint detection and depth estimation method in the off-road environment designed by the present invention establishes a terrain surface modeling method based on polar coordinate grids, reducing the complexity of modeling and the computational complexity; first constructs a point cloud segmentation method based on Gaussian process regression, utilizes the characteristics of the Gaussian distribution, and improves the segmentation accuracy of off-road undulating roads; then proposes a negative obstacle point cloud detection method based on the grid reference height to detect negative obstacles in real time; finally, in combination with the vehicle's trench crossing height and depth, proposes a negative obstacle false alarm reduction method to further improve the vehicle's passability. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a structural block diagram based on the centralized fusion of multi-source lidars.

[0067] Figure 2 is a grid segmentation diagram.

[0068] Figure 3 is a point cloud distribution diagram of different negative obstacle types.

[0069] Figure 4 is a negative obstacle detection model diagram.

[0070] Figure 5 is a schematic diagram of the change of the negative obstacle with distance.

[0071] Figure 6 is a schematic diagram of the negative obstacle from a bottom-up perspective.

[0072] Figure 7 It is a schematic diagram of a negative obstacle from a top-down perspective.

[0073] Figure 8 It is a schematic diagram of a negative obstacle entering the detectable area.

[0074] Figure 9 It is a schematic diagram of negative obstacle detection. Specific implementation manners

[0075] The following will further clearly and completely describe in detail the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Based on the embodiments of 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.

[0076] As Figures 1-9 shown, a method for joint detection and depth estimation of negative obstacles based on centralized fusion of multi-source lidar specifically includes the following steps:

[0077] S001: Establish a ground model based on a polar coordinate grid map: project the three-dimensional point cloud data into a two-dimensional grid map to reduce the calculation amount;

[0078] S002: Establish a terrain segmentation algorithm based on Gaussian process regression fitting: select the lidar point cloud that is relatively close to the vehicle body and has the minimum height as the seed point, and continuously iteratively learn the Gaussian model using the seed point to increase the accuracy of point cloud segmentation;

[0079] S003: Establish a negative obstacle detection algorithm based on the ground reference height; combine the trained Gaussian ground segmentation model, predict the ground height of each segmented grid, and then judge whether the point cloud is ground point cloud based on the difference between the measured height and the predicted height of the lidar point cloud;

[0080] S004: Establish a negative obstacle point cloud clustering algorithm based on a radially bounded nearest neighbor graph; cluster the grids with the attribute of negative obstacles, and adopt a clustering method based on a radially bounded nearest neighbor graph, which can effectively suppress the problems of over-segmentation and under-segmentation

[0081] S005: Establish a false alarm reduction method based on the trench crossing width and depth of the vehicle itself; the purpose of negative obstacle detection is to ensure the safe driving of unmanned vehicles and improve the passability of the road surface, but there is a certain contradiction between the two: if the detection sensitivity of the algorithm is very high, it may set the road surface with negative obstacles that the vehicle can pass as an impassable road surface, thus reducing the passability of the road surface. Therefore, the present invention combines the trench crossing depth and width of the vehicle to reduce false alarms to increase the passability of the road surface.

[0082] AsFigure 1 As shown in the figure, an embodiment of the present invention provides a negative obstacle joint detection and estimation algorithm based on centralized fusion of multi-source lidar, which specifically includes the following steps:

[0083] S001: Establish a terrain modeling method based on polar coordinate grids. Specifically, it includes the following steps:

[0084] (1) Project the point cloud data onto a two-dimensional grid coordinate map: As Figure 2 shown, the range that the vehicle needs to perceive is set as a rectangle or a circle (the rectangle is set according to the characteristics of the lidar and the task requirements). For any lidar measurement data p = [x, y, z], its radial distance and azimuth angle can be calculated as follows:

[0085]

[0086] The terrain modeling of a larger area does not meet the real-time requirements of unmanned vehicles. Therefore, the range of the vehicle body 360 is unevenly divided. Combining Figure 2 , with the Y-axis as the positive direction, when 0 ≤ θ ≤ 90 and -90 ≤ θ < 0, the division angle is 0.5; when 90 < θ < 270, the division angle is 2. At this time, the total number of sectors is

[0087]

[0088] (2) After step (1), the rectangle is divided into 450 uneven sectors. To further reduce the computational complexity of terrain modeling, each sector is further unevenly divided into multiple grids (cells). For sector j and lidar point cloud i, if

[0089]

[0090] where, is the radial distance of lidar point cloud i in sector j. If when, then If when, then Performing uneven division conforms to the physical characteristics of the lidar: the number of point clouds falling within a unit area becomes sparser as the distance increases from near to far.

[0091] (3) After steps (1) and (2), the point clouds within the region of interest have been mapped onto the polar coordinate grid map. To further reduce the computational complexity of ground segmentation, the three-dimensional data is mapped into two-dimensional data:

[0092]

[0093] S002: Establish a terrain segmentation algorithm based on Gaussian process regression. The specific steps are as follows:

[0094] (1) Selection of initial seed points: The selection of seed points is crucial for the accuracy of ground segmentation. The grids closer to the vehicle body (ρ ≤ 10m) are more likely to be the ground. Therefore, select the point with the smallest absolute value |z| among such grids as the initial seed point. For sector j, assume the total number of grids is C j . The initial point cloud set composed of the point clouds with the smallest height within the grids where ρ ≤ 10m is

[0095] G j = {p i , i = 0…C j};

[0096] Among them, the |z| corresponding to point p i is the smallest in grid i.

[0097] The set composed of all points with the smallest height in sector j is

[0098]

[0099] Among them,

[0100] (2) Gaussian regression model: For the case where the Bayesian linear regression model is a Gaussian random process:

[0101]

[0102] where ω ∼ N(0, Σ); its mean is:

[0103]

[0104] The covariance is:

[0105]

[0106] The covariance function of the Gaussian process is required to be positive definite on a finite input set and is a symmetric function satisfying the Mercer condition. Therefore, the covariance function is equivalent to the kernel function. In the present invention, the squared exponential covariance function is adopted:

[0107]

[0108] According to the above formula, the joint probability distribution of any two random variables can be given as follows:

[0109]

[0110] where X *is the training sample set (the set of initial seed points); X is the set of test sample points (the set of all possible ground points); K(X,X) is the covariance matrix of the input test points; K(X * ,X * ) is the covariance matrix of the input training sample points; K(X,X * ) is the cross-covariance of the input test points and the training sample points; X * is the set of training sample points, X is the set of test points; f is the output of the test points; f * is the output of the training samples.

[0111] Combined with the joint probability density function, the corresponding posterior probability density function can be calculated

[0112]

[0113] For the measurement model with observation noise z = f(x) + v, where its posterior probability distribution is as follows:

[0114]

[0115] (3) Train the Gaussian regression model: Use the set of initial seed points as the training set G j ; Use the set of all possible seed points as the test point set and perform iterative Gaussian model training. For each point in the test set Calculate the mean and variance predicted by the Gaussian model, and the calculation method is as follows

[0116]

[0117] For each Make a judgment according to the following threshold

[0118]

[0119] If the point meets the above threshold conditions, then add the point to the set G j .

[0120] (4) Ground model based on Gaussian process regression: The set G j contains the point cloud in sector j that is most likely to be the ground. Consider all the point clouds in G j as the training set and train the Gaussian process regression model to improve the accuracy and stability of segmentation; then use the trained ground segmentation model to segment the point cloud in sector j. Ground model based on Gaussian process regression: Use the trained Gaussian model. For sector j, the radial distance of grid k from the ego vehicle is

[0121]

[0122] Substitute the parameter into the trained Gaussian model, and the predicted height of grid k can be obtained

[0123]

[0124] According to the pre-set parameter T Ground_Th , for the point cloud p i = [x i , y i , z i in this grid, if the absolute value of its difference from the reference height is less than the following threshold

[0125]

[0126] then this point is considered a ground point, otherwise it is considered an obstacle point

[0127] S003: Establish a negative obstacle detection algorithm based on the ground reference height. Specifically, it includes the following steps

[0128] (1) Conduct modeling and analysis on the main distribution of negative obstacles on off-road surfaces. The main possible distribution of negative obstacle point clouds is as Figure 3 shown. As can be seen from Figure 3 , the distribution characteristics of negative obstacles are diverse and their sizes vary. For such negative obstacles, a detection model as Figure 4 is established

[0129] As can be seen from Figure 4 , triangles HDO and dWO are similar triangles. Therefore, the relationship between the installation height of the lidar, the detection distance D, and the depth d of the negative obstacle that can be detected is

[0130]

[0131] From this model, we can find that the detection distance is inversely proportional to the depth of the negative obstacle and directly proportional to the installation height and the width of the negative obstacle, which conforms to subjective understanding. This geometric relationship is the basis for us to judge the depth of negative obstacles and the safe driving of vehicles, and is also the key to evaluating the performance of the negative obstacle perception system

[0132] (2) Based on the constructed grid map, a negative obstacle grid method based on the point cloud reference height is proposed, and the operation process is as follows

[0133] S1: For any grid in the fan-shaped block j (0 ≤ j < M) of the polar coordinate grid map as long as there is a 3D lidar point Relative height to the reference ground (depth feature), then this grid is marked as a negative obstacle grid.

[0134] S2: When the grid is determined If it is a negative obstacle grid, then in the current sector block, from the grid Start by looking forward for the grid closest to it that is marked as ground. All grids here are marked as negative obstacle grids.

[0135] S3: For the negative obstacle grid, all grids without point cloud data in its 5-grid map neighborhood are calculated and marked as negative obstacle grids.

[0136] S004: Establish a negative obstacle grid clustering algorithm based on radial bounded nearest neighbor graph. Specifically, it includes the following steps:

[0137] For negative obstacle grids, the method based on radial bounded nearest neighbor graph is used to cluster the grids of the same type, which can be simply described as follows:

[0138] (1) Traverse all negative obstacle grids.

[0139] (2) If the current negative obstacle grid already belongs to a certain negative obstacle class, jump to the next grid.

[0140] (3) For the currently traversed negative obstacle grid.

[0141] a. Find all negative obstacle grids in the neighborhood of r.

[0142] b. If the neighborhood grid already belongs to a negative obstacle class, the current point is classified into the current class.

[0143] c. If the current point already belongs to a negative obstacle class and the negative obstacle class of the neighboring point is different from the negative obstacle class of the current point, raster merging is required.

[0144] S005: Establish a false alarm reduction algorithm based on the width and depth of the trench crossed by the vehicle;

[0145] From the perspective of the sensor, the depth of the negative obstacle that can be detected also changes in real time with the distance between the vehicle body and the negative obstacle. Figure 5 shown.

[0146] According to the change of negative obstacles in the perspective, we can model the obstacle depth calculation under different perspectives as a set-valued optimization problem, where a point cloud in the grid m Depth of negative obstacles Reference height to grid m The mapping relationship between them is

[0147]

[0148] Based on this mapping relationship, the depth of negative obstacles distributed in the fan-shaped block s is estimated as follows:

[0149]

[0150] At the same time, considering that negative obstacles may occupy multiple sectors,

[0151]

[0152] Finally, the depth is judged, and as the vehicle moves, the negative obstacles directly in front of the vehicle are estimated in real time to ensure the safety of the vehicle's autonomous driving.

[0153] After obtaining the negative obstacle grid, calculate the width of the grid in the longitudinal direction and calculate the center distance ρ of the negative obstacle grid farthest from the vehicle. max The distance ρ from the center point of the negative obstacle grid closest to the vehicle body min , so the negative obstacle satisfies the following two conditions

[0154]

[0155] Then the negative obstacle is true; otherwise, it is false. According to the width and depth of the trench crossing of the vehicle, a corresponding false alarm reduction method is designed to increase the vehicle's passability.

[0156] Figures 6-9 The point cloud data of the real-time scene was used to complete the simulation test of the algorithm. The sampling interval of the laser radar is T = 0.1s. Under the Ubuntu system, a negative obstacle perception system based on the centralized fusion of multi-source laser radars was developed based on C++, using ROS as the communication middleware and developed through QT. The main laser radar is the Tudatong Falcon Smart Edition, the forward blind radar is the Hesai QT128, and the left and right and rear blind radars are the Sagitar RSBpearl.

[0157] Figures 6-7 The distribution of negative obstacles at different viewing angles is given. From the bottom view, it can be seen that the point cloud of negative obstacles is distributed below the ground; from the top view, it can be seen that the negative obstacles appear as holes on the ground and are only sparsely distributed on the back wall of the negative obstacles. Figures 8-9 It can be seen that when a negative obstacle enters the detectable range, the detection scheme provided by the present invention can quickly detect the obstacle, and conducts ground modeling analysis based on the negative obstacle. Through the multi-sensor fusion scheme, the negative obstacle can be detected at a maximum distance of 10 meters, which meets the requirements of low-speed unmanned driving of the vehicle in off-road environments.

[0158] The above experimental results verify that the present invention improves the detection performance of negative obstacles, which has great theoretical and practical value for driverless vehicles in off-road complex environments.

[0159] The present invention is not limited to the above embodiments. Anyone should know that structural changes made under the inspiration of the present invention, as long as they have the same or similar technical solutions as the present invention, fall within the protection scope of the present invention.

[0160] The technologies, shapes, and structures not described in detail in the present invention are all well-known technologies.

Claims

1. A method for joint detection and depth estimation of negative obstacles in off-road environments, characterized in that: The following steps are involved: S1. Establishing a ground model based on a polar coordinate grid map: projecting the three-dimensional point cloud data into a two-dimensional grid map; S2. Establish a terrain segmentation method based on Gaussian process regression fitting: select the lidar point cloud that is close to the vehicle body and has the smallest height as the seed point, and use the seed point to iteratively learn the Gaussian model to form a Gaussian ground segmentation model; S3. Establish a negative obstacle detection algorithm based on the ground reference height; combine the trained Gaussian ground segmentation model to predict the ground height of each segmented grid, and then use the difference between the measured height and the predicted height of the lidar point cloud to determine whether the point cloud is a ground point cloud; S4. Establish a negative obstacle point cloud clustering algorithm based on radial bounded nearest neighbor graph; The rasters with negative barrier attributes are clustered using a clustering method based on radial bounded nearest neighbor graph; S5. Establish a false alarm reduction method based on the width and depth of the trench crossed by the vehicle. The purpose of negative obstacle detection is to ensure the safe driving of unmanned vehicles and improve the passability of the road surface. However, there is a certain contradiction between the two: if the detection sensitivity of the algorithm is very high, the negative obstacle road surface that the vehicle can pass through may be set as an impassable road surface.

2. The method for joint detection and depth estimation of negative obstacles in off-road environments according to claim 1, characterized in that: The specific steps of establishing the ground model based on the polar coordinate grid map in S1 are as follows: 1) Non-uniform grid division method: The terrain modeling of a larger area does not meet the real-time requirements of the unmanned vehicle. Therefore, the 360-degree range of the vehicle body is divided unevenly. Taking the Y axis as the positive direction, when the angle θ satisfies 0≤θ≤90 and -90≤θ<0, the division angle is 0.5; when θ satisfies 90<θ<270, the division angle is 2. At this time, the total number of sectors is 2) After step 1), the rectangle is divided into 450 non-uniform sectors. In order to further reduce the amount of calculation for terrain modeling, each sector is further divided non-uniformly into multiple grids. The grid division also adopts a non-uniform division method. For sector j, if the distance to the vehicle body is less than 15m, the division interval is 0.1m; if the distance to the vehicle body is greater than 15m, the division interval is 0.2m. Assuming that according to the non-uniform division method, the number of grids in sector j is C j ; For sector j and lidar point cloud i, if in is the radial distance of the lidar point cloud i in sector j; k = 1,…,C j ; is the lower bound of grid k, and is the upper bound of grid k; according to the non-uniform partitioning method, if When like When 3) After steps 1) and 2), the point clouds within the range of interest have been mapped to the polar coordinate grid map; in order to further reduce the amount of calculation, the three-dimensional data is mapped to two-dimensional data: Where p is the measurement data of the laser radar; ρ is the radial distance from this point cloud to the vehicle body.

3. The method for joint detection and depth estimation of negative obstacles in off-road environment according to claim 1, characterized in that: The specific steps of establishing the terrain segmentation method based on Gaussian process regression fitting in S2 are as follows: 1) Selection of initial seed points: For sector j, the initial point cloud set consisting of the point cloud with the smallest height in each grid within ρ≤10m is where p i is the point with the smallest height in grid i, is the number of grids within 10 meters of the vehicle body in sector j; The set of all point combinations with the smallest height in sector j is 2) Gaussian regression model: For the case where the Bayesian linear regression model is a Gaussian random process and has a measurement model z=f(x)+v with observation noise, where the observation noise z is the measured data, f(x) is the mapping model, and its posterior probability distribution is as follows: Where K(·) represents the covariance function; X is the mean of the Gaussian density function; Σ is the covariance of the Gaussian probability density function; 3) Iterative training of Gaussian model: The initial seed point set is used as the training set G j ; Take all possible seed points as test points Perform iterative Gaussian model training. The trained model is as follows For each Judge based on the following thresholds If point If the above threshold conditions are met, the point Add to collection G j middle; 4) Ground model based on Gaussian process regression: Using the trained Gaussian model, for sector j, the total number of grids is C j , the radial distance between grid k and the vehicle is The parameters Substituting into the trained Gaussian model, the predicted height of grid k can be obtained. According to the pre-set parameter T Ground_Th , for the point cloud p in this grid i =[x i ,y i ,z i ], if the absolute value of the difference between it and the reference height is less than the following threshold If the point is positive, it is considered to be a ground point; otherwise, it is considered to be an obstacle point.

4. The method for joint detection and depth estimation of negative obstacles in off-road environment according to claim 1, characterized in that: The specific steps of establishing the negative obstacle detection algorithm based on the ground reference height in S3 are as follows: 1) Modeling and analysis of the main distribution of negative obstacles on off-road roads. Through the analysis of the negative obstacle point cloud, the relationship between the installation height of the lidar, the detection distance D, and the detectable negative obstacle depth d is: 2) Based on the constructed grid map, a negative obstacle point cloud representation method based on point cloud reference height is proposed. The operation process is as follows: (1) For any grid n (0≤n<C) in the sector j (0≤j<M) of the polar coordinate grid map j ), as long as there is a 3D LiDAR point Relative height to the reference ground meters, then this grid is marked as a negative obstacle grid; (2) When determining the grid If it is a negative obstacle grid, then in the current sector block, from the grid Start by looking forward for the grid closest to it that is marked as ground. All grids here are marked as negative obstacle grids; (3) For the negative obstacle grid, all grids without point cloud data in its 5-grid map neighborhood are calculated and marked as negative obstacle grids.

5. The method for joint detection and depth estimation of negative obstacles in off-road environment according to claim 1, characterized in that: The specific steps of establishing the negative obstacle point cloud clustering algorithm based on the radial bounded nearest neighbor graph in S4 are as follows: For negative obstacle grids, a method based on radial bounded nearest neighbor graph is used to cluster grids of the same type. The steps are as follows: 1) Traverse all negative obstacle grids; 2) If the current negative obstacle grid already belongs to a certain negative obstacle class, jump to the next grid; 3) For the currently traversed negative obstacle grid; (1) Find all grids in the neighborhood of r; (2) If the neighboring point already belongs to a negative obstacle class, the current point is classified into the current class; (3) If the current point already belongs to a certain negative obstacle class and the negative obstacle class of the neighboring point is different from the negative obstacle class of the current point, class fusion is required.

6. The method for joint detection and depth estimation of negative obstacles in off-road environment according to claim 1, characterized in that: The specific steps of establishing the false alarm reduction method based on the trench width and trench depth of the vehicle in S5 are as follows: After obtaining the negative obstacle grid, calculate the width of the grid in the longitudinal direction and calculate the center distance ρ of the negative obstacle grid farthest from the vehicle. max The distance ρ from the center point of the negative obstacle grid closest to the vehicle body min , if the negative obstacle size satisfies the following two conditions: in is the deepest depth of negative obstacle estimation; Then this negative obstacle is true; otherwise, it is false; design a corresponding false alarm reduction method based on the trench width and trench depth of the vehicle.

Images