Indoor corridor bifurcation detection method based on multi-line laser radar

Abstract

The present invention discloses a method for detecting indoor corridor bifurcations based on multi-line laser radar. This method is used to detect road information at the location of a mobile robot when the mobile robot is automatically navigating. This method first defines a data structure that expresses intersection information, then uses the random sampling consensus algorithm RANSAC plane matching to perform ground detection on the original point cloud acquired by the laser radar to remove ground interference points; then, a beam model based on a polar coordinate system is established in the ground point cloud to detect the number and angle of intersections ahead; finally, based on the characteristics of the detection results of the beam model, the best matching intersection selection algorithm is used to select the final intersection bifurcation angle and position as the final detection result, and the result is then used for the navigation of the mobile robot. The present invention improves the detection speed and accuracy by removing the interference of the radar ground line by removing ground points, and removing low-density point clouds by statistical filtering; the beam model is used to extract intersection features, which effectively improves the detection accuracy of intersection features.

Description

Technical Field

[0001] The present invention relates to the field of point cloud recognition technology, and in particular to a method for detecting a bifurcation in an indoor corridor based on a multi-line laser radar. Background Art

[0002] Currently, many road intersection detection algorithms have been proposed in the remote sensing field. These methods are all based on aerial imagery, which has a large range and cannot be used in the field of autonomous driving. In the field of autonomous driving, video-based detection methods have been proposed in the past few years. However, these methods use complex image processing techniques, which require high computing power from the on-board computing unit and sacrifice processing speed. Poor lighting conditions (such as cloudy days, overexposure, and other interference from moving vehicles and pedestrians) still make road intersection detection in the field of autonomous driving extremely difficult and almost impossible.

[0003] Therefore, active sensors are now widely used. Kodagoda et al. used laser measurement to quickly extract two corresponding road edges or curbs, and used laser depth measurement and extended Kalman filters to detect Y-shaped, T-shaped, and X-shaped intersections. Wijesoma et al. fused a 2D laser scanner with a CCD camera to track road curbs. Aycard et al. used a variety of sensors including lidar and stereo vision, but mainly focused on the safety of road intersections, including the tasks of perception and risk assessment of moving vehicles and pedestrians, rather than finding intersections. They cannot provide comprehensive road perception information for vehicle path planning in autonomous driving. Moreover, the extraction of intersection features is often obtained at the intersection, and the corresponding intersection information cannot be obtained before reaching the intersection. These intersection distances and bifurcation angles are of great significance for advance planning of autonomous driving. Summary of the Invention

[0004] The present invention provides an indoor corridor bifurcation detection method based on multi-line laser radar, which can detect the intersection information ahead in advance during the mobile robot navigation task, including the intersection bifurcation angle and intersection position information.

[0005] The technical solution to achieve the purpose of the present invention is: a method for detecting indoor corridor bifurcations based on a multi-line laser radar, comprising the following steps:

[0006] Step 1: Define a data structure representing intersection information as the output of the multi-line LiDAR-based indoor corridor fork detection method. Use the random sampling consensus algorithm RANSAC to detect the ground surface, remove the points on the road surface from the original point cloud, and obtain a ground point cloud that includes the curb.

[0007] Step 2: Use the obtained ground point cloud to build a polar coordinate beam model mathematical model. This model can detect the angle of the intersection ahead relative to the polar coordinate. The intersection detection result is obtained by building the beam model.

[0008] Step 3: Use the best intersection matching algorithm to extract and filter the features of the obtained detection results, and select the best matching intersection information as the final detection result.

[0009] Compared with the existing technology, the present invention has the following significant advantages: (1) the use of lidar sensors can effectively solve the problem that image methods are limited by lighting conditions and is more robust to environmental changes; (2) the intersection can be detected before reaching the intersection, effectively detecting the intersection information in the passable area ahead and obtaining the detection results in advance. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is the algorithm flow chart of this method.

[0011] Figure 2 is the original image input by the lidar.

[0012] Figure 3 It is a point cloud image after the lidar selects the area of ​​interest.

[0013] Figure 4 It is the point cloud image of the area of ​​interest after removing the ground points.

[0014] Figure 5 This is the beam model effect diagram after the beam model is established.

[0015] Figure 6 is the final intersection detection result. DETAILED DESCRIPTION

[0016] This method is used to detect the road information of the mobile robot's location during its automatic navigation. The present invention is based on a multi-line laser radar indoor corridor fork detection method, comprising the following steps:

[0017] Step 1: Design the data structure class RoadCrossInfo for intersection information, which includes the following elements:

[0018] RoadCrossInfo -road_direction:vector <int> < / int> +getRoadDirection():int

[0019] The intersection information type is defined as RoadCrossInfo, which contains the member variable road_direction, the intersection angle information, and the member function getRoadDirection() as a method for extracting intersection information.

[0020] The ultimate goal of this method is to obtain the intersection information (if there is an intersection) in front of the mobile robot in each frame of input point cloud for subsequent navigation effect optimization.

[0021] Use RANSAC to detect ground point clouds. First, select the mobile robot's region of interest (ROI). For mobile robot navigation tasks, the ROI refers to the traversable area ahead of the road. On semi-structured roads, the ROI should focus on all scenes above and below a certain xy plane ahead of the mobile robot. In road detection tasks, intersection information must be detected in advance, so the ROI is selected as 10 meters in front, behind, and on the left and right sides:

[0022] {(x,y)|-10m≤x≤10m,-10m≤y≤10m}

[0023] When using RANSAC to fit a plane, two parameters need to be determined: the distance threshold d threshoid The distance threshold is used to determine whether a point cloud belongs to a sample point within or outside the RANSAC-fitted plane model. The iteration number n is used to iterate the loss function multiple times to obtain the optimal plane fitting parameters. In actual point cloud detection, these two parameters also need to be determined based on the density, resolution, and number of lines of the input point cloud.

[0024] Step 2: Use the beam model to obtain intersection information. The beam model is primarily a mathematical model based on a polar coordinate system. Therefore, it is necessary to determine the location of the polar coordinate system established within a frame of input point cloud data, i.e., the beam model's origin (x0, y0).

[0025] First, determine the origin of the beam model (x0, y0) based on the range of interest. In theory, (x0, y0) and the range of interest should satisfy the following relationship:

[0026] x0≤x max , y0≤y max

[0027] Step 3: Based on the candidate angle set of all detection results obtained in the third step, use the designed screening algorithm to calculate the angle with the highest confidence as the final intersection bifurcation angle.

[0028] By sorting all candidate angles in ascending order, two thresholds need to be determined: the separation angle θ similar and the merging angle θ merge When the difference between two adjacent angles Δθ≤θ merge , it is considered that the two angles point to the same intersection; when Δθ>θ similarWhen , the two angles are considered to point to two different intersections. This is how the candidate angle sequence is grouped. Each group finally represents a set T of angles pointing to the same intersection.

[0029] Next, determine the threshold θ remove Compare the angles within each intersection angle set T, when:

[0030]

[0031] The number of samples in the angle set T is considered too small to represent an intersection. For the angle set T that satisfies the above inequality, the intersection shape is determined based on its number: 2 represents no intersection, 3 represents T-shaped or Y-shaped intersection, 4 represents + intersection, and greater than 4 represents other intersections.

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Example

[0034] Combine Figure 1 The present invention is based on a multi-line laser radar indoor corridor fork detection method, the implementation steps are as follows:

[0035] Step 1: Implement the intersection information data structure class RoadCrossInfo as the final output of the algorithm.

[0036] Step 2: First, crop the region of interest from the input point cloud image. The original input image is as shown in the attached figure. Figure 2 As shown in the figure, the point cloud image of the area of ​​interest after cropping is shown in the attached figure. Figure 3 Then, perform RANSAC plane matching on the point cloud image in the area of ​​interest, record its index after finding the point cloud in the plane, remove the ground points in the point cloud image of the area of ​​interest, and only leave the ground points. Figure 4 shown.

[0037] Step 3: Build a beam model in the ground point cloud image to detect intersection information. The beam model's beam diagram is shown in the attached figure. Figure 5 In the beam graph obtained above, a set of candidate angles for the intersection bifurcation angles is obtained.

[0038] Step 4: In the candidate angle set obtained in step 3, use the screening algorithm to obtain the best matching angle segmentation set and intersection location information, such as Figure 6 shown.

[0039] In summary, the present invention improves detection speed and accuracy by removing ground points to remove interference from radar ground lines and by using statistical filtering to remove low-density point clouds; and uses a beam model to extract intersection features, effectively improving the detection accuracy of intersection features.

Claims

1. A method for detecting forks in indoor corridors based on multi-line laser radar, characterized in that: The steps include: Step 1: First, define a data structure representing intersection information as the output of the multi-line lidar-based indoor corridor fork detection method. Then, use the random sampling consensus algorithm (RANSAC) to construct a plane model, detect the ground surface in the original point cloud input by the radar, and remove points on the ground surface from the point cloud to obtain a ground point cloud that includes the curb. Step 2: Build a polar coordinate beam model mathematical model in the obtained ground point cloud, and use the beam model to obtain the results of the intersection detection ahead; the method is as follows: 1) Establish a polar coordinate system on the xy plane in the point cloud coordinate system, and determine the angle value θ of the point in the polar coordinate system based on the position coordinates (x, y, z) of the point, where (x o ,y o ) is the origin of the coordinate system: The point cloud is divided according to the angle θ of each point, and the points with the same θ value are grouped into a set Within, that is: 2) For each set Calculate the minimum distance d of the included points from the origin min With the maximum distance d max : x i ,y i is the coordinate of the point, i is the number of points; For each set Definition d k is the length of the beam within this angle: When d k =1 means that there may be no ground point within the kth angle, which is the candidate angle for the intersection fork; 3) Count all d k = 1 angle k, the continuous d k = 1 is divided into a group of angles k i , each group T i Finally, the candidate angle set ψ is formed k∈T i ,T i ∈{ψ}; Step 3: Use the best intersection matching algorithm to extract and filter the features of the intersection detection results, and select the best matching intersection information as the final detection result. The method is as follows: 1) In the candidate angle set ψ obtained in the second step, if the element T i The size is greater than the threshold num threshoid , then T i As the next candidate angle set, otherwise the angles in the set are considered to be false detections and discarded; 2) In the candidate angle set T i , calculate the corresponding segmentation angle s i : θ max T i The maximum angle value, θ min T i The smallest angle value in the final s i The set S is the best matching intersection set; 3) The number of angles contained in the set S is used as the basis for judging the intersection type, that is, the number of intersection forks is the number of angles in S, 2 represents no intersection, 3 represents T-shaped or Y-shaped intersection, 4 represents + intersection, and greater than 4 represents other intersections.

2. The method for detecting indoor corridor bifurcations based on multi-line laser radar according to claim 1, characterized in that: The random sampling consensus algorithm RANSAC described in the first step detects the ground and removes points on the ground of the point cloud as follows: 1) Define the data structure that expresses intersection information, including the number of forks and the angle of the forks; 2) Select the area of ​​interest in the ground point cloud, that is, whether there is an intersection within the area that the algorithm needs to detect; The region of interest is defined as: {(x,y)|-10m≤x≤10m,-10m≤y≤10m} (x,y) represents the coordinate value in the point cloud coordinate system; 3) Randomly select three points in the point cloud within the range of interest and calculate the corresponding plane equation: ax+by+cz+d=0 Among them, a, b, c, d are plane fitting parameters, {x, y, z} are the three-dimensional coordinates of the point, and points that meet the same plane equation are on the same plane; calculate the Euclidean distance d from all points in the input point cloud to the plane in sequence i : d i =|ax i +by i +cz i +d| Among them, {x i ,y i , z i } represents the three-dimensional coordinates of each point within the range of interest; then the threshold d is selected threshoid , if d i ≤d threshoid , then the point is considered to be a sample point inside the plane model, i.e., an interior point; otherwise, it is a sample point outside the plane model, i.e., an exterior point. The number of current interior points is recorded. 4) Repeat steps 2)-3) above to select the best fitting parameters, that is, the model parameters corresponding to the plane with the largest number of inliers. At the end of each iteration, an iteration end judgment factor is calculated based on the expected error rate, the number of best inliers, the total number of samples, and the current number of iterations. The decision to stop the iteration is based on the set number of iterations. 5) After the iteration is completed, the best model parameters are the final parameter estimates, that is, the sample points in the plane model are used as ground points, their indexes in the point cloud are recorded, and they are discarded in the element point cloud.

Images