A method for detecting roadway obstacles based on 3D point clouds

Abstract

This invention relates to the field of image processing technology, and more particularly to a method for detecting roadway obstacles based on 3D point clouds. The method includes acquiring a 3D point cloud image and obtaining a point cloud set therein; processing the point cloud set using an obstacle detection model to identify false obstacle point cloud sets; and removing the point clouds containing false obstacle point cloud sets from the 3D point cloud image. The obstacle detection model includes a motion feature classification layer, comprising: acquiring the point cloud set and calculating the velocity of its points; identifying whether the point cloud set belongs to a fast-moving obstacle or a suspected obstacle based on the point velocity; calculating the motion direction of points in the suspected obstacle; and identifying whether the point cloud set belongs to a false obstacle or a slow-moving obstacle based on whether the motion directions of the points are consistent. The motion feature classification layer of this invention identifies false obstacles through velocity thresholds and motion directions, effectively filtering out false obstacles and avoiding misclassification as real obstacles, thereby reducing the false detection rate.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for detecting roadway obstacles based on three-dimensional point clouds. Background Technology

[0002] Currently, the field of artificial intelligence is booming, and autonomous driving technology, as one of its important applications, is gradually moving from the laboratory to real life. Autonomous driving technology relies on the vehicle's perception and detection of its environment, and obstacle detection is a crucial part of this perception, and a key factor in the safe and efficient operation of autonomous vehicles. Accurately identifying and distinguishing between real and false obstacles plays a decisive role in the stability and safety of autonomous driving systems.

[0003] In practical applications, autonomous vehicles face complex and ever-changing environmental conditions, especially when encountering natural phenomena such as rain, snow, and dust storms. They are prone to misidentifying these as obstacles. For example, in heavy rain, the splashing raindrops interfere with the sensors of autonomous vehicles, causing them to misidentify the dense rain as an obstacle, thus triggering braking or automatic lane-changing avoidance measures. This not only reduces driving efficiency but may also cause traffic chaos or even rear-end collisions.

[0004] In existing technologies, the confidence values ​​of each point in a 3D point cloud image are typically initialized, and obstacle and non-obstacle regions are divided based on manually set spatial spacing between imaging points and target size thresholds. Subsequently, the confidence values ​​are further adjusted based on the size and density information of the regional point clouds to eliminate false obstacles.

[0005] However, these methods have significant limitations. Because they rely on manually preset size thresholds, density standards, and other fixed rules, they are ill-suited to the diversity of pseudo-obstacles in complex and ever-changing scenarios. Under conditions of varying concentrations of blowing sand and diverse forms of rain and snow, the point cloud characteristics of pseudo-obstacles differ significantly, and existing methods cannot flexibly address these differences, easily leading to misjudgments or missed detections. This results in reduced safety for autonomous driving and insufficient reliability of intelligent detection systems. Summary of the Invention

[0006] To address the technical problem of misjudging or missing false obstacles in complex environments during autonomous driving scenarios, this invention provides a method for alleyway obstacle detection based on 3D point clouds. The method includes: acquiring a 3D point cloud image and obtaining multiple point cloud sets from the 3D point cloud image; processing any point cloud set using a constructed obstacle detection model to identify the category information of the obstacle to which the point cloud set belongs; and removing the point cloud of the false obstacle from the 3D point cloud image when the point cloud set belongs to a false obstacle.

[0007] The obstacle detection model includes a motion feature classification layer, which is used to obtain the category of obstacles in the point cloud set, output the velocity and motion direction of the point cloud set, including: acquiring the point cloud set, calculating the velocity of the points in the point cloud set; if the velocity of the points in the point cloud set is above the velocity threshold, the point cloud set is determined to be a fast-moving obstacle, otherwise it is a suspected obstacle; calculating the motion direction of the points in the suspected obstacle; if the motion direction of the points in the suspected obstacle is inconsistent with the motion direction of other points in the suspected obstacle, the point cloud set is determined to be a false obstacle.

[0008] In the motion feature classification layer of this invention, obstacle categories are initially identified through speed thresholds. This distinguishes between high-speed vehicles and slow-moving pedestrians, as well as different types of obstacles such as wind, sand, rain, and snow, in complex environments. This step significantly narrows the scope of subsequent analysis, reduces unnecessary computation and processing time, and improves overall detection efficiency. Subsequently, by analyzing the motion direction to capture the differences between obstacles such as wind, sand, rain, snow, and pedestrians, false obstacles are identified, effectively preventing them from being misclassified as real obstacles and reducing the false detection rate. Finally, false obstacles are excluded from the image, making the final detection results more reliable.

[0009] As a further improvement to the method of the present invention, calculating the velocity of a point in a point cloud set includes: obtaining the current position information of a point in the point cloud set; predicting the next position information of a point in the point cloud set using Kalman filtering; and calculating the velocity of a point in the point cloud set based on the current position information and the next position information.

[0010] The Kalman filter algorithm can effectively integrate historical data with current observations and has good robustness to noise. Even in complex environments, such as poor lighting in tunnels or electromagnetic interference that causes noise in the measurement data, it can accurately predict the position and velocity of points, providing a reliable basis for subsequent detection.

[0011] As a further improvement to the method of the present invention, the motion direction of the midpoint of the suspected obstacle is calculated, including: obtaining a velocity vector based on the position information of the midpoint of the suspected obstacle; calculating the motion direction vector of the point based on the velocity vector; and calculating the azimuth and pitch angles of the midpoint of the suspected obstacle based on the motion direction vector.

[0012] By obtaining the velocity vector from the position information, and then deriving the motion direction vector based on the velocity vector, the azimuth and pitch angles are finally calculated, enabling a comprehensive and highly accurate description of the motion direction of the midpoint of a suspected obstacle. In complex alleyways or autonomous driving scenarios, the differences in motion direction characteristics between real and false obstacles are subtle. This refined calculation method can effectively capture these differences, providing crucial information for subsequent false obstacle identification.

[0013] As a further improvement to the method of the present invention, the obstacle detection model also includes loss calculation, which includes: calculating the classification loss of each obstacle. Location loss due to obstacles Loss of motion vectors of obstacles ; Calculate the total loss due to obstacles ,in, and These are preset weighting coefficients.

[0014] By calculating classification loss, localization loss, and motion vector loss, the accuracy of classification is improved, the accuracy of localization is enhanced, and the motion vector estimation is optimized. The total loss function, through preset weighting coefficients, allows for flexible adjustment of the proportions of classification loss, localization loss, and motion vector loss in the total loss, improving the model's adaptability and performance in various complex scenarios.

[0015] As a further improvement to the method of the present invention, the obstacle detection model also includes an input layer, which includes: acquiring the coordinate data of any point in the point cloud set; constructing its coordinate data matrix based on the coordinate data; and constructing a point cloud data set of all points in the point cloud set based on the coordinate data matrix.

[0016] As another improvement to the method of the present invention, the obstacle detection model further includes a feature extraction layer, which includes: receiving motion features of a point cloud set output by the motion feature classification layer, the motion features including speed and direction of motion; and performing three-layer convolution on the motion features to obtain a three-layer convolution feature map.

[0017] As another improvement to the method of the present invention, the obstacle detection model further includes a dense classification layer, which includes: receiving the feature map output by the feature extraction layer; calculating the probability of the feature map belonging to the obstacle category through an activation function, wherein the obstacle categories include: fast-moving obstacles, slow-moving obstacles, and pseudo-obstacles; extracting the nonlinear relationship between motion features and obstacle categories using a multilayer perceptron (MLP), and calculating the motion state probability of the midpoint of the suspected obstacle; and weightedly fusing the motion state probability and the obstacle category probability to obtain the obstacle category information of the midpoint of the suspected obstacle.

[0018] Dense classification layers can quickly transform data in feature maps into probability distributions for each obstacle category. Simultaneously, the Multilayer Perceptron (MLP) delves into the nonlinear relationship between motion features and obstacle categories, calculating the motion state probability of points within suspected obstacles. These two layers work synergistically to perform deep analysis of feature maps from multiple perspectives, fully utilizing hidden information in the data and enhancing the ability to express obstacle features. This enables more accurate capture of subtle differences between different types of obstacles, laying the foundation for accurate classification.

[0019] As another improvement to the method of the present invention, acquiring a three-dimensional point cloud image includes: acquiring a three-dimensional point cloud image; and performing noise reduction processing on the three-dimensional point cloud image.

[0020] As another improvement to the method of the present invention, obtaining multiple point cloud sets in a three-dimensional point cloud image includes: applying mean offset clustering to the point clouds in the three-dimensional point cloud image to obtain multiple point cloud sets in the three-dimensional point cloud image.

[0021] As another improvement to the method of the present invention, when the direction of motion of the point in the suspected obstacle is inconsistent with the direction of motion of other points in the suspected obstacle, the angle difference between the point and any other point is greater than a preset difference threshold.

[0022] The beneficial effects of this invention are as follows: The motion feature classification layer of this invention first distinguishes different types of obstacles based on velocity thresholds, narrowing the scope of subsequent analysis and improving detection efficiency; then, it analyzes the motion direction to identify false obstacles, reducing the false detection rate and eliminating false obstacles to make the detection results more reliable. The Kalman filter algorithm integrates historical and current data to accurately predict the position and velocity of points in complex environments, providing a foundation for detection. By calculating classification, localization, and motion vector losses, it improves classification accuracy, localization precision, and motion vector estimation. The total loss function can flexibly adjust the weighting of each loss through preset weighting coefficients, enhancing the model's adaptability and performance in complex scenarios. Attached Figure Description

[0023] Figure 1 A flowchart of a method for detecting roadway obstacles based on three-dimensional point clouds, provided for an embodiment of the present invention;

[0024] Figure 2 This is a flowchart of the obstacle detection model construction in this embodiment;

[0025] Figure 3 This is a flowchart of motion detection in this embodiment;

[0026] Figure 4 This is a schematic diagram of the 3D point cloud identified by the obstacle detection model in this embodiment;

[0027] Figure 5 This is a schematic diagram of a three-dimensional point cloud image obtained after processing by the method according to an embodiment of the present invention.

[0028] In the picture, 31 is a building; 32 is blowing sand; and 33 is the ground. Detailed Implementation

[0029] This embodiment provides a method for detecting roadway obstacles based on three-dimensional point clouds, such as... Figure 1 As shown, the method includes steps S100-S300:

[0030] Step S100: Acquire a 3D point cloud image and obtain multiple point cloud sets from the 3D point cloud image.

[0031] To elaborate, three-dimensional point cloud images can be acquired using a 3D laser scanner or a depth camera.

[0032] After acquiring a 3D point cloud image, the points need to be clustered. This clustering process yields several clusters, each representing a set of points that share spatial similarities. Each cluster can be viewed as a point cloud representation of an obstacle or a region with similar features. For example, in an autonomous driving scenario, one cluster might represent a vehicle, another a pedestrian, and still others might represent dust storms, rain, or snow.

[0033] Point cloud clustering can be achieved through clustering algorithms, such as mean shift clustering, K-means clustering, Gaussian mixture model clustering, and so on.

[0034] In addition, the 3D point cloud image needs to be denoised. Gaussian filtering can be used to remove noise points from the point cloud and for data augmentation.

[0035] Step S200: Use the constructed obstacle detection model to process the point cloud set and identify the category information of the obstacles to which the point cloud set belongs.

[0036] To elaborate, such as Figure 2 As shown, the obstacle detection model refers to the model used in this invention to detect fast-moving obstacles and false obstacles. This model is built upon the existing PointNet framework and consists of four layers: an input layer, a motion feature classification layer, a feature extraction layer, and a dense classification layer. The main addition is a motion feature classification layer, which identifies the category information of the obstacle. The construction of these four layers is described below.

[0037] Step S210: Construct the input layer.

[0038] The input layer, as the entry point of the model, receives 3D point cloud data and transforms it to facilitate processing by the subsequent motion feature layer. This includes: receiving the point cloud set and obtaining the coordinate data of any point within it; constructing a coordinate data matrix based on the coordinate data; and constructing a point cloud dataset containing all points in the point cloud set based on the coordinate data matrix.

[0039] Specifically, typically, each point in a 3D point cloud set has coordinates. Since the obstacles are moving, the coordinates of each point are different at different times. Therefore, each point in the point cloud set forms a point cloud coordinate matrix at each time step:

[0040] ;

[0041] in, Indicates the first The point is at The point cloud coordinate data matrix at time step, It is the first The one in the first The coordinates corresponding to the time.

[0042] After obtaining the point cloud coordinate matrices of all points in the point cloud set at different times, construct a point cloud data set from the point cloud coordinate matrices of all points at the same time:

[0043] ;

[0044] in ， Represents a point cloud data set. Indicates the first The point is at Point cloud coordinate data matrix at any given time.

[0045] For example, a point cloud set contains two points with coordinates of... , , its first k The point cloud coordinate matrices corresponding to the time points are as follows:

[0046] ;

[0047] ;

[0048] The corresponding point cloud dataset is:

[0049] .

[0050] After the input layer constructs the point cloud data, it outputs it to the motion feature classification layer.

[0051] Step S220: Construct a motion feature classification layer.

[0052] In detail, the motion feature classification layer is used to obtain the category of obstacles in the point cloud set, including: classifying the point cloud set into fast-moving obstacles, slow-moving obstacles, or pseudo-obstacles based on motion detection, and outputting its motion features, such as speed and direction of motion, to the dense classification layer.

[0053] like Figure 3 As shown, motion detection classification includes steps S221-S224:

[0054] Step S221: Obtain the point cloud set and calculate the velocity of the points in the point cloud set.

[0055] In detail, the motion feature classification layer first receives the point cloud data set output by the input layer. Then, it predicts the position (coordinates) for the next moment using a Kalman filter, and calculates the velocity of the points in the point cloud set. If the velocity of a point in the point cloud set is above a velocity threshold, the point cloud set is determined to be a fast-moving obstacle; otherwise, it is a suspected obstacle. The motion direction of the points in the suspected obstacle is calculated. If the motion direction of a point in the suspected obstacle is inconsistent with the motion direction of other points in the suspected obstacle, the point cloud set is determined to be a false obstacle.

[0056] It should be noted that the point cloud data set output by the input layer only contains coordinate information for each point, without velocity information. Therefore, it is necessary to reconstruct the point cloud coordinate matrix for each point, adding velocity coordinates. This velocity value can be initialized to 0 at only one specific moment. The reconstructed point cloud velocity coordinate matrix is ​​as follows:

[0057] ;

[0058] in, Indicates the first The point is at The point cloud coordinate data matrix at time step, It is the first The one in the first The coordinates corresponding to the time. Indicates the first The point at the th Velocity in the direction of the time coordinate.

[0059] After the point cloud velocity coordinate matrix of each point is constructed, the motion velocity of the points in the point cloud set is calculated.

[0060] Specifically, obtain the first point of the point cloud set. Location information at any given time; predicting the first point in the point cloud set using Kalman filtering. Location information at any given time; based on the first Location information at time and the first The positional information at any given time is used to calculate the velocity and velocity vector of the points in the point cloud set.

[0061] Step S222: Determine whether the point cloud set belongs to a fast-moving obstacle or a suspected obstacle.

[0062] Specifically, if the velocity of a point in a point cloud set is above the speed threshold of 30 km / h, the point cloud set is determined to be a fast-moving obstacle; otherwise, it is a suspected obstacle.

[0063] Preliminary obstacle identification using speed thresholds allows for quick and efficient differentiation between different obstacle types. Setting a speed threshold acts like an intelligent filter, rapidly distinguishing fast-moving objects from slow-moving or stationary ones. This initial screening significantly narrows the scope of subsequent analysis, reduces unnecessary computation and processing time, and improves the overall efficiency of the detection system.

[0064] Regarding the selection of speed thresholds, in autonomous driving scenarios, generally speaking, when the speed of an obstacle reaches 30 km / h or higher relative to the autonomous vehicle, it can be considered a fast-moving obstacle. For example, on a highway, if a vehicle approaches from behind at a significantly higher speed than the vehicle itself, or if a vehicle suddenly cuts into front of the vehicle from the adjacent lane at high speed, these vehicles can be considered fast-moving obstacles.

[0065] Step S223: Calculate the direction of motion of the midpoint of the suspected obstacle.

[0066] Specifically, if the point cloud is determined to be a suspected obstacle, further identification is required. This identification is based on the direction of motion. Specifically: A velocity vector is obtained based on the position information of the midpoint of the suspected obstacle; the motion direction vector of the point is calculated based on the velocity vector; the azimuth and pitch angles of the midpoint of the suspected obstacle are calculated based on the motion direction vector as the motion direction of the midpoint. Calculating the motion direction from the position information is existing technology and will not be discussed further here.

[0067] Step S224: Determine whether the suspected obstacle point cloud set belongs to a false obstacle or a slow-moving obstacle.

[0068] The specific judgment logic is as follows: when the direction of movement of a point is inconsistent with the direction of movement of other points in the suspected obstacle, the point cloud set is determined to be a false obstacle; otherwise, it is determined to be a slow-moving obstacle. Inconsistent direction of movement means that the angle difference between that point and any other point is greater than a preset difference threshold.

[0069] After initially identifying suspected obstacles, the key highlight of this method is further identifying false obstacles based on their movement direction. In real-world environments, false obstacles are often generated by various interference factors, such as rain, snow, and blowing sand. The movement direction of these false obstacles exhibits significant irregularity and randomness, differing drastically from the movement direction characteristics of real obstacles. By analyzing the movement direction, the system can accurately capture this difference. For example, the movement direction of debris occasionally flying through a passageway may be influenced by various factors such as airflow, resulting in a clear discrepancy with the movement direction of surrounding real obstacles. In this case, the motion feature classification layer can accurately identify it as a false obstacle, avoiding misjudgment as a real obstacle, thereby significantly reducing the false detection rate.

[0070] In addition, considering that the direction of a pedestrian's arms and legs is usually not the same when they are moving, the movement direction of all corresponding points cannot be completely consistent. Therefore, a threshold can be set to prevent misjudgment in determining whether the movement direction is consistent.

[0071] For example, if a threshold of 10 is set, and assuming there are 15 points in the point cloud set whose motion direction is inconsistent with that point, then the point cloud set to which that point belongs is considered a false obstacle because its motion direction is inconsistent with the other points. If assuming there are 5 points in the point cloud set whose motion direction is inconsistent with that point, then the point cloud set to which that point belongs is considered a slow-moving obstacle because its motion direction is consistent with the other points.

[0072] Finally, in order for the feature extraction layer to process the data, the input layer needs to output the velocity and direction of motion of the point cloud set.

[0073] Step S230: Construct the feature extraction layer.

[0074] To elaborate, the feature extraction layer needs to receive motion features, namely speed and direction of motion, from the motion feature classification layer. Then, it performs three convolutions on the motion features. The specific operation of the three convolutions is the same as that of PointNet. For example, the first convolution extracts local features from the output of the feature extraction layer; the second convolution enhances the output of the first layer by extracting mid-scale features; and the third convolution extracts global features from the output of the second layer, ultimately resulting in a feature map of the three convolutions.

[0075] Furthermore, since the PointNet model itself has the ability to identify static obstacles, static obstacles in the 3D point cloud image can be obtained in this step.

[0076] Step S240: Construct a dense classification layer.

[0077] In detail, this includes: receiving the feature map output from the feature extraction layer; calculating the probability of the feature map belonging to the obstacle category through an activation function, where obstacle categories include: fast-moving obstacles, slow-moving obstacles, and pseudo-obstacles; using a multilayer perceptron (MLP) to extract the non-linear relationship between motion features and obstacle categories, and calculating the motion state probability of the point; and weightedly fusing the motion state probability and the obstacle category probability to obtain the obstacle category information of the point.

[0078] Specifically, dense classification layers typically apply an activation function, such as the softmax function. The softmax function first transforms the feature map output from the feature extraction layer into a probability distribution, which in this invention represents the probability of obstacle categories. Specifically, it assigns probabilities to categories such as fast-moving obstacles, slow-moving obstacles, static obstacles, and pseudo-obstacles.

[0079] Multilayer perceptrons (MLPs) possess powerful nonlinear fitting capabilities. Through the combination of multiple layers of neurons and the action of activation functions, they can automatically learn these complex nonlinear relationships, thereby more accurately classifying and understanding obstacles with different motion characteristics. Next, the MLP is used to extract the nonlinear relationship between motion features and obstacle categories, calculating the motion state probability of points.

[0080] Finally, the motion state probability and obstacle category probability are weighted and fused to obtain the obstacle category information of the point.

[0081] Obstacle category information refers to a comprehensive score that includes obstacle classification results and motion status results.

[0082] For example, the obstacle classification result is one of the following: static obstacle, fast-moving obstacle, slow-moving obstacle, and pseudo-obstacle. The corresponding motion state results can be speed 0, speed above 30km / h, speed 5m / s and not in the direction of motion, and speed between 5-30km / h and in the direction of motion, respectively.

[0083] In summary, this describes the construction process of the four layers of the obstacle detection model. Based on this, loss assessment and optimization processing are still required for the detection results of this model.

[0084] The loss function measures the difference between the model's predictions and the actual results, thus guiding the model to optimize parameters for obstacle classification and 3D point cloud reconstruction. The loss function can be calculated using mean squared error loss, mean absolute error loss, or cross-entropy loss. The model's loss consists of three parts: classification loss, localization loss, and motion loss.

[0085] Classification loss measures the difference between the model's predicted probability of an obstacle's category and the true category. Localization loss measures the difference between the model's predicted location of an obstacle and the true location of the obstacle. Motion loss measures the difference between the model's predicted probability of the obstacle's direction of motion and the true direction of motion.

[0086] Finally, calculate the total loss of the obstacle. :

[0087] ;

[0088] in, and To preset the weighting coefficients, Represents the classification loss of obstacles, Indicates the location loss of obstacles, Loss of motion vectors of obstacles;

[0089] about and Accurate obstacle classification is crucial for autonomous driving decision-making. For example, differentiating between different types of obstacles such as pedestrians, vehicles, and traffic signs allows the vehicle to respond differently. Therefore, appropriately increasing the... Strengthen the weight of classification tasks. If accurate obstacle location is more critical during vehicle movement, such as accurately determining the obstacle's position to avoid a collision, then... The value may need to be relatively large. For example, in urban roads, vehicles frequently need to avoid pedestrians and other vehicles, making the localization task more important, and the value can be set to... =0.4, =0.6. However, in certain scenarios, such as parking lots, accurately identifying parking space markings and obstacle types may be more important, and a setting could be made... =0.6, =0.4.

[0090] The PointNet model also involves optimization algorithms, typically using stochastic gradient descent (SGD) and its variants, such as Adagrad, Adadelta, RMSProp, Adam, and AdamW, to update the model's parameters. These algorithms adaptively adjust the learning rate based on the gradient information of the loss function to accelerate model convergence and avoid getting trapped in local optima. This part represents existing technology and will not be discussed in detail here.

[0091] Step S300: When the point cloud set belongs to a false obstacle, remove the point cloud of the false obstacle from the 3D point cloud image.

[0092] like Figure 4 As shown, after the above steps, obstacle category information such as false obstacles, fast-moving obstacles, slow-moving obstacles, and static obstacles in the 3D point cloud image can be identified. Figure 4 As shown, static obstacles such as buildings (31), false obstacles such as dust (32), and the ground (33) are mainly presented. The point cloud set belonging to false obstacles needs to be removed from the 3D point cloud image, including the dust (32). This prevents the autonomous vehicle from braking or automatically changing course to avoid hazards due to incorrect obstacle recognition. The 3D point cloud image after removing the dust is shown below. Figure 5 As shown in the image, ground 33 is not present. Ground 33 is usually removed after the 3D image is acquired to prevent it from affecting the subsequent detection of obstacles.

[0093] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for detecting roadway obstacles based on three-dimensional point clouds, characterized in that, include: Acquire a 3D point cloud image and obtain multiple point cloud sets from the 3D point cloud image; The constructed obstacle detection model is used to process any of the point cloud sets to identify the category information of the obstacles to which the point cloud set belongs; When the point cloud set belongs to a false obstacle, the point cloud of the false obstacle is removed from the 3D point cloud image; The obstacle detection model includes a motion feature classification layer, used to obtain the category of obstacles at points in the point cloud set, and output the velocity and direction of motion of the point cloud set, including: The point cloud set is acquired, and the velocity of the points in the point cloud set is calculated. If the velocity of the points in the point cloud set is above a velocity threshold, the point cloud set is determined to be a fast-moving obstacle; otherwise, it is a suspected obstacle. Calculating the motion direction of the midpoint of the suspected obstacle includes: obtaining a velocity vector based on the position information of the midpoint of the suspected obstacle; calculating the motion direction vector of the point based on the velocity vector; and calculating the azimuth and pitch angles of the midpoint of the suspected obstacle based on the motion direction vector. A threshold is set. If the movement direction of a point in a suspected obstacle is inconsistent with the movement direction of other points in the suspected obstacle, and the number of other points with inconsistent movement directions is greater than the threshold, then the point cloud set to which the point belongs is determined to be a false obstacle; otherwise, the point cloud set to which the point belongs is determined to be a slow-moving obstacle.

2. The method for detecting roadway obstacles based on three-dimensional point clouds according to claim 1, characterized in that, The calculation of the speed of points in the point cloud set includes: Obtain the current position information of the points in the point cloud set; Predict the position information of points in the point cloud set at the next moment using Kalman filtering; The velocity of the points in the point cloud set is calculated based on the current position information and the position information at the next moment.

3. The method for detecting roadway obstacles based on three-dimensional point clouds according to claim 1, characterized in that, The obstacle detection model also includes loss calculation, which includes: Calculate the classification loss for each obstacle. Location loss due to obstacles Loss of motion vectors of obstacles ; Calculate the total loss of the obstacle. ,in, and These are preset weighting coefficients.

4. The method for detecting roadway obstacles based on three-dimensional point clouds according to claim 1, characterized in that, The obstacle detection model further includes an input layer, which includes: Obtain the coordinate data of any point in the point cloud set; Construct its coordinate data matrix based on the coordinate data; Construct a point cloud data set containing all points in the point cloud set based on the coordinate data matrix.

5. The method for detecting roadway obstacles based on three-dimensional point clouds according to claim 1, characterized in that, The obstacle detection model further includes a feature extraction layer, which comprises: The motion features of the point cloud set output by the motion feature classification layer are received, and the motion features include velocity and direction of motion. The motion features are subjected to three convolutions to obtain a three-layer convolutional feature map.

6. The method for detecting roadway obstacles based on three-dimensional point clouds according to claim 5, characterized in that, The obstacle detection model further includes a dense classification layer, which comprises: Receive the feature map output by the feature extraction layer; The probability of the feature map belonging to an obstacle category is calculated by an activation function. The obstacle categories include: fast-moving obstacles, slow-moving obstacles, and pseudo-obstacles. The nonlinear relationship between motion features and obstacle category is extracted using a multilayer perceptron (MLP), and the motion state probability of the midpoint of the suspected obstacle is calculated. The probability of the motion state and the probability of the obstacle category are weighted and fused to obtain the obstacle category information at the midpoint of the suspected obstacle.

7. The method for detecting roadway obstacles based on three-dimensional point clouds according to claim 1, characterized in that, The acquisition of the 3D point cloud image includes: Acquire 3D point cloud images; The three-dimensional point cloud image is then denoised.

8. The method for detecting roadway obstacles based on three-dimensional point clouds according to claim 1, characterized in that, The step of obtaining multiple point cloud sets from the 3D point cloud image includes: Mean offset clustering is applied to the point clouds in the three-dimensional point cloud image to obtain multiple point cloud sets in the three-dimensional point cloud image.

9. The method for detecting roadway obstacles based on three-dimensional point clouds according to claim 1, characterized in that, The direction of motion of a point in the suspected obstacle is inconsistent with the direction of motion of other points in the suspected obstacle when the angle difference between the point and any other point is greater than a preset difference threshold.

Images