Target recognition method based on feature matching

Abstract

The present invention discloses a target recognition method based on feature matching, comprising the following steps: Step 1, obtaining point cloud data collected by a laser radar, obtaining the two-dimensional position coordinates and reflection intensity of each laser measurement point; Step 2, processing the laser radar point cloud using a clustering algorithm based on an adaptive distance threshold, obtaining point cloud clusters representing different targets; Step 3, performing line segment fitting on each point cloud cluster, and extracting corresponding target features based on the fitted line segments; Step 4, collecting a large amount of experimental data, obtaining target features through the above clustering and feature extraction methods, and artificially marking the target type, thereby obtaining reference features corresponding to different types of targets; Step 5, matching multiple features of the target to be identified with the reference features of each type of target, and completing target type recognition. The target recognition method based on feature matching reduces the use of computing resources and improves recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent driving, and in particular to a target recognition method based on feature matching. Background Art

[0002] The current development of intelligent driving technology and the improvement of the system have enhanced the vehicle's autonomous perception of the environment, provided drivers with rich driving assistance and warning information, and improved driving safety. Among them, the recognition of multiple targets in the vehicle's surrounding environment based on on-board sensors is an important prerequisite for the intelligent driving system to complete many driving assistance tasks (such as adaptive cruise control, collision warning, etc.). LiDAR is a scanning sensor based on non-contact laser ranging. Its point cloud data can provide distance information with high accuracy and resolution. It is widely used as the main means of obstacle detection and identification around smart cars.

[0003] Target recognition methods based on LiDAR point clouds can be divided into two main categories: deep learning algorithms and non-deep learning algorithms. Deep learning algorithms use neural networks to process input point cloud data to complete target recognition, but the target recognition network model structure is complex, occupies more computing resources, and has high hardware performance requirements; non-deep learning algorithms obtain target recognition results based on set spatial rules, which are greatly affected by the accuracy of point cloud segmentation and feature extraction. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art.

[0005] To this end, the present invention proposes a target recognition method based on feature matching to reduce the use of computing resources and improve recognition accuracy.

[0006] The target recognition method based on feature matching according to an embodiment of the present invention comprises the following steps:

[0007] Step 1: Obtain point cloud data collected by the laser radar to obtain the two-dimensional position coordinates and reflection intensity of each laser measurement point;

[0008] Step 2: Process the lidar point cloud using a clustering algorithm based on an adaptive distance threshold to obtain point cloud clusters representing different targets;

[0009] Step 3: perform line segment fitting on each point cloud cluster, and extract corresponding target features based on the fitted line segments;

[0010] Step 4: Collect a large amount of experimental data, obtain target features through the above clustering and feature extraction methods, and manually mark the target types, so as to obtain reference features corresponding to different types of targets;

[0011] Step 5: Match multiple features of the target to be identified with reference features of each type of target to complete target type identification.

[0012] The beneficial effects of the present invention are as follows: (1) the present invention considers the relationship between the distribution of laser point cloud and the direction of laser pulse, and adopts a method based on adaptive distance threshold to cluster the point cloud, thereby improving the accuracy of point cloud clustering; (2) the present invention uses the standard deviation of the distance from the target point cloud cluster to the corresponding line segment and the standard deviation of the reflection intensity as reference features for target recognition, and uses a large amount of experimental data to construct a normal distribution of the reference features, thereby improving the target recognition performance; (3) the target recognition method of the present invention is not based on deep learning, and reduces the consumption of computing resources while ensuring the accuracy of target recognition.

[0013] According to one embodiment of the present invention, in the first step, a frame of laser radar point cloud data is composed of a group of measurement points sorted by angle, and each measurement point is represented by a two-dimensional coordinate and a reflection intensity in a laser radar coordinate system.

[0014] According to one embodiment of the present invention, in the second step, the following steps are specifically included:

[0015] Step 201, judging whether two adjacent measurement points are located on a target surface relatively parallel to the laser pulse according to the angular relationship between the two adjacent measurement points and the coordinate origin;

[0016] Step 202: when it is considered that the two points are located on a target surface relatively parallel to the laser pulse, a larger adaptive distance threshold is calculated; when it is considered that the two points are located on a target surface relatively perpendicular to the laser pulse, a smaller adaptive distance threshold is calculated;

[0017] Step 203, using the coordinates to calculate the Euclidean distance between the two points, when the Euclidean distance is greater than the adaptive distance threshold, it is considered that the two points do not belong to the same target; otherwise, the two points are classified as the same target.

[0018] According to one embodiment of the present invention, in the third step, the following steps are specifically included:

[0019] Step 301, performing line segment fitting within each point cloud cluster after clustering using an iterative endpoint fitting algorithm;

[0020] Step 302: based on the fitted line segments, target features are obtained, where the target features include the number of line segments, the length of each line segment, the standard deviation of the distance from the point cloud cluster to the corresponding line segment, and the standard deviation of the reflection intensity of the point cloud cluster.

[0021] According to an embodiment of the present invention, in the fourth step, the target types include vehicles, cyclists and pedestrians.

[0022] According to an embodiment of the present invention, the reference features corresponding to the different types of targets include:

[0023] Vehicle target reference features, the number of line segments is 1 or 2, the length interval of each line segment, the normal distribution parameters of the standard deviation of the distance between the point cloud cluster and the corresponding line segment, and the normal distribution parameters of the standard deviation of the reflection intensity of the point cloud cluster;

[0024] Reference features of the cyclist target, the number of line segments is 1 or 2, the length interval of each line segment, the normal distribution parameters of the standard deviation of the distance between the point cloud cluster and the corresponding line segment, and the normal distribution parameters of the standard deviation of the reflection intensity of the point cloud cluster;

[0025] Pedestrian target reference features, number of line segments 1, length interval of line segments, normal distribution parameters of the standard deviation of the distance between point cloud clusters and corresponding line segments, and normal distribution parameters of the standard deviation of the reflection intensity of point cloud clusters.

[0026] According to one embodiment of the present invention, in the fifth step, the specific steps are:

[0027] Extract multiple features of the target to be identified, determine whether the number of line segments of the target meets the requirement of a certain type of line segment number, whether the line segment length is within the reference interval, and use the normal distribution of the standard deviation of the distance between each type of point cloud cluster to the corresponding line segment to calculate the probability density of this feature of the target. Use the normal distribution of the standard deviation of the reflection intensity of the point cloud cluster to calculate the probability density of this feature of the target.

[0028] According to one embodiment of the present invention, target recognition is divided into the following two cases:

[0029] In the first case, when the target to be identified matches the reference feature of a certain type of target, that is, the number of line segments is consistent, the length of the line segments is within the reference interval, and the two calculated probability densities are greater than the set threshold, and it does not match the reference features of other types, then the target to be identified is classified as this type;

[0030] In the second case, when the target to be identified matches the reference features of both types of targets, the target type with the larger probability density of the standard deviation of the distance between the point cloud cluster and the corresponding line segment is selected as the target recognition result.

[0031] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 is a flow chart of a target recognition method based on feature matching of the present invention;

[0035] Figure 2 It is a schematic diagram of point cloud clustering in the target recognition method based on feature matching of the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] The target recognition method based on feature matching according to an embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0038] See Figure 1 The target recognition method based on feature matching of the present invention comprises the following steps:

[0039] Step 1: Obtain the point cloud data collected by the laser radar, and obtain the two-dimensional position coordinates and reflection intensity of each laser measurement point; in step 1, a frame of laser radar point cloud data consists of a group of measurement points sorted by angle, and each measurement point is represented by the two-dimensional coordinates and reflection intensity in the laser radar coordinate system.

[0040] Among them, the point cloud data collected by the laser radar in step 1 can be expressed as {(x n ,y n ,I n )|n=0,...,N}, where x n ,y n ,I n They represent the horizontal distance, vertical distance, and reflection intensity of the nth laser measurement point, respectively, and N is the number of laser measurement points. The measurement points in each frame of point cloud data are arranged in order according to the laser pulse scanning angle.

[0041] In the second step, a clustering algorithm based on an adaptive distance threshold is used to process the laser radar point cloud to obtain point cloud clusters representing different targets. In the second step, the following steps are specifically included:

[0042] Step 201, judging whether two adjacent measurement points are located on a target surface relatively parallel to the laser pulse according to the angular relationship between the two adjacent measurement points and the coordinate origin;

[0043] Step 202: when it is considered that the two points are located on a target surface relatively parallel to the laser pulse, a larger adaptive distance threshold is calculated; when it is considered that the two points are located on a target surface relatively perpendicular to the laser pulse, a smaller adaptive distance threshold is calculated;

[0044] Step 203, using the coordinates to calculate the Euclidean distance between the two points, when the Euclidean distance is greater than the adaptive distance threshold, it is considered that the two points do not belong to the same target; otherwise, the two points are classified as the same target, that is, when the Euclidean distance is less than the adaptive distance threshold, it is considered that the two points belong to the same target.

[0045] See Figure 2 , where p n-2 、p n-1 、p n and p n+1 are all measurement points, r n-1 is the measuring point p n-1 The radial distance, r n is the measuring point p n The radial distance of the first connecting line (measurement point p n The line connecting the coordinate origin) and the second line (the measuring point p n With measuring point p n-1 The angle formed by the second connecting line (the measuring point p n With measuring point p n-1 The line connecting the measuring point p n-1 The angle formed by the line connecting the coordinate system and the origin.

[0046] Use |α-β| to represent the angle relationship between two adjacent measurement points and the coordinate origin in step 201. The measurement point p n-1 With measuring point p n The distance threshold is calculated as follows:

[0047]

[0048] The specific meanings of the symbols in formula (1) are:

[0049] r n-1 Indicates the measurement point p n-1 The radial distance;

[0050] Indicates the angular resolution of the lidar;

[0051] λ represents the angle threshold used to calculate the maximum spacing;

[0052] ε represents the sensor error.

[0053] When |α-β| is less than the set angle threshold, the two points are considered to be on the target surface relatively perpendicular to the laser pulse, and a larger λ is set to obtain a smaller distance threshold;

[0054] When |α-β| is greater than the set angle threshold, the two points are considered to be on the target surface relatively parallel to the laser pulse, and a smaller λ is set to obtain a larger distance threshold; finally, by comparing the Euclidean distance between the two points with the distance threshold, it is determined whether the two points belong to the same target.

[0055] Step 3: perform line segment fitting on each point cloud cluster, and extract corresponding target features based on the fitted line segments; Step 3 specifically includes the following steps:

[0056] Step 301: perform line segment fitting using the Iterative End Point Fitting (IEPF) algorithm within each point cloud cluster after clustering, connect the first and last data points in each point cloud cluster to form a line segment, and calculate the maximum distance of other data points between these two points relative to this line segment; when the maximum distance is greater than a set threshold, the line segment is divided into two line segments at the data point corresponding to the maximum distance, and the above check is repeated until all line segments of the point cloud cluster do not need to be split again;

[0057] Step 302: based on the fitted line segments, target features are obtained, where the target features include the number of line segments, the length of each line segment, the standard deviation of the distance from the point cloud cluster to the corresponding line segment, and the standard deviation of the reflection intensity of the point cloud cluster.

[0058] Step 4: Collect a large amount of experimental data, obtain target features through the above clustering and feature extraction methods, and manually mark the target types to obtain reference features corresponding to different types of targets; in step 4, the target types include vehicles, cyclists and pedestrians. The reference features corresponding to different types of targets include:

[0059] The vehicle target reference features include the number of line segments, 1 or 2, and the length interval of each line segment [l min,v , l max,v ], the normal distribution parameter of the standard deviation of the distance between the point cloud cluster and the corresponding line segment Normal distribution parameters of the standard deviation of the reflection intensity of a point cloud cluster

[0060] The reference features of the cyclist target include the number of line segments, 1 or 2, and the length interval of each line segment [l min,c , l max,c ], the normal distribution parameter of the standard deviation of the distance between the point cloud cluster and the corresponding line segment Normal distribution parameters of the standard deviation of the reflection intensity of a point cloud cluster

[0061] The reference features of pedestrian targets include the number of line segments 1, the length interval of the line segments [l min,p , l max,p ], the normal distribution parameter of the standard deviation of the distance between the point cloud cluster and the corresponding line segment Normal distribution parameters of the standard deviation of the reflection intensity of a point cloud cluster

[0062] By manually labeling a large amount of experimental data, a large number of features corresponding to the three types of targets can be obtained. The segment length interval of each type of target is determined by the minimum and maximum values ​​of all segment length features of this type of target. According to the standard deviation of the distance between a large number of point cloud clusters and the corresponding line segments and the standard deviation of the reflection intensity of the point cloud clusters for each type of target, the corresponding mean and standard deviation can be calculated.

[0063] Step 5: Match multiple features of the target to be identified with reference features of each type of target to complete target type identification. In step 5, the specific steps are:

[0064] Extract multiple features of the target to be identified, determine whether the number of line segments of the target meets the requirement of a certain type of line segment number, whether the line segment length is within the reference interval, and use the normal distribution of the standard deviation of the distance between each type of point cloud cluster to the corresponding line segment to calculate the probability density of this feature of the target. Use the normal distribution of the standard deviation of the reflection intensity of the point cloud cluster to calculate the probability density of this feature of the target.

[0065] Target recognition is divided into the following two cases:

[0066] In the first case, when the target to be identified matches the reference feature of a certain type of target, that is, the number of line segments is consistent, the length of the line segments is within the reference interval, and the two calculated probability densities are greater than the set threshold, and it does not match the reference features of other types, then the target to be identified is classified as this type;

[0067] In the second case, when the target to be identified matches the reference features of both types of targets, the target type with the larger probability density of the standard deviation of the distance between the point cloud cluster and the corresponding line segment is selected as the target recognition result.

[0068] Further explanation of the above two situations:

[0069] In the first case, the target to be identified is represented by two line segments after clustering and line segment fitting, and the lengths of the two line segments are both within [l min,v , l max,v ], according to the normal distribution Calculate the probability density f of the standard deviation of the distance between the target point cloud cluster and the corresponding line segment d,mv Greater than the set threshold f d,v , according to the normal distribution

[0070] Calculate the probability density f of the standard deviation of the emission intensity of the target point cloud cluster i,mv Greater than the set threshold f i,v , and the target feature to be identified does not match one or more reference features of the cyclist or pedestrian target, then the target is a vehicle target;

[0071] In the second case, the target cluster to be identified is represented by a line segment after fitting with the line segment, and the length of the line segment is [l min,c , l max,c ], also in [l min,p , l max,p ], according to the normal distribution The probability density f calculated respectively d,mc 、f i,mc 、f d,mp 、f i,mp are greater than the corresponding threshold, then further by comparing f d,mc With f d,mp Complete target recognition; when f d,mc >f d,mp , the target is a cyclist, otherwise the target is a pedestrian.

[0072] In summary, compared with the prior art, the target recognition method based on feature matching provided by the present invention has the advantage of high target recognition accuracy, while solving the problem of high computing resource usage.

[0073] The target recognition method based on feature matching of the present invention has the following advantages: (1) The present invention considers the relationship between the distribution of laser point cloud and the direction of laser pulse, and adopts a method based on adaptive distance threshold to cluster the point cloud, thereby improving the accuracy of point cloud clustering; (2) The present invention uses the standard deviation of the distance from the target point cloud cluster to the corresponding line segment and the standard deviation of the reflection intensity as reference features for target recognition, and uses a large amount of experimental data to construct the normal distribution of the reference features, thereby improving the target recognition performance; (3) The target recognition method of the present invention is not based on deep learning, and reduces the consumption of computing resources while ensuring the accuracy of target recognition.

[0074] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A target recognition method based on feature matching, characterized in that: The following steps are involved: Step 1: Obtain point cloud data collected by the laser radar to obtain the two-dimensional position coordinates and reflection intensity of each laser measurement point; Step 2: Process the lidar point cloud using a clustering algorithm based on an adaptive distance threshold to obtain point cloud clusters representing different targets; The specific steps include: Step 201, judging whether two adjacent measurement points are located on a target surface relatively parallel to the laser pulse according to the angular relationship between the two adjacent measurement points and the coordinate origin; Step 202: when it is considered that the two points are located on a target surface relatively parallel to the laser pulse, a larger adaptive distance threshold is calculated; when it is considered that the two points are located on a target surface relatively perpendicular to the laser pulse, a smaller adaptive distance threshold is calculated; Step 203, using the coordinates to calculate the Euclidean distance between the two points, when the Euclidean distance is greater than the adaptive distance threshold, it is considered that the two points do not belong to the same target; otherwise, the two points are classified as the same target; Step 3: perform line segment fitting on each point cloud cluster, and extract corresponding target features based on the fitted line segments; Step 4: Collect a large amount of experimental data, obtain target features through the above clustering and feature extraction method, and manually mark the target type, so as to obtain reference features corresponding to different types of targets; the reference features corresponding to different types of targets include: Vehicle target reference features, the number of line segments is 1 or 2, the length interval of each line segment, the normal distribution parameters of the standard deviation of the distance between the point cloud cluster and the corresponding line segment, and the normal distribution parameters of the standard deviation of the reflection intensity of the point cloud cluster; Reference features of the cyclist target, the number of line segments is 1 or 2, the length interval of each line segment, the normal distribution parameters of the standard deviation of the distance between the point cloud cluster and the corresponding line segment, and the normal distribution parameters of the standard deviation of the reflection intensity of the point cloud cluster; Pedestrian target reference features, number of line segments 1, line segment length interval, normal distribution parameters of the standard deviation of the distance between the point cloud cluster and the corresponding line segment, and normal distribution parameters of the standard deviation of the reflection intensity of the point cloud cluster; Step 5: Match multiple features of the target to be identified with reference features of each type of target to complete target type identification; the specific steps are: Extract multiple features of the target to be identified, determine whether the number of line segments of the target meets the requirements of a certain type of line segment number, whether the line segment length is within the reference interval, and use the normal distribution of the standard deviation of the distance between each type of point cloud cluster to the corresponding line segment to calculate the probability density of this feature of the target. Use the normal distribution of the standard deviation of the reflection intensity of the point cloud cluster to calculate the probability density of this feature of the target. Target recognition is divided into the following two cases: In the first case, when the target to be identified matches the reference feature of a certain type of target, that is, the number of line segments is consistent, the length of the line segments is within the reference interval, and the two calculated probability densities are greater than the set threshold, and it does not match the reference features of other types, then the target to be identified is classified as this type; In the second case, when the target to be identified matches the reference features of both types of targets, the target type with the larger probability density of the standard deviation of the distance between the point cloud cluster and the corresponding line segment is selected as the target recognition result.

2. The target recognition method based on feature matching according to claim 1, characterized in that: In the first step, a frame of laser radar point cloud data is composed of a group of measurement points sorted by angle, and each measurement point is represented by a two-dimensional coordinate and reflection intensity in the laser radar coordinate system.

3. The target recognition method based on feature matching according to claim 1, characterized in that: In the third step, the following steps are specifically included: Step 301, performing line segment fitting within each point cloud cluster after clustering using an iterative endpoint fitting algorithm; Step 302: based on the fitted line segments, target features are obtained, where the target features include the number of line segments, the length of each line segment, the standard deviation of the distance from the point cloud cluster to the corresponding line segment, and the standard deviation of the reflection intensity of the point cloud cluster.

4. The target recognition method based on feature matching according to claim 1, characterized in that: In the fourth step, the target types include vehicles, cyclists and pedestrians.

Images