Unmanned vehicle obstacle identification method and system

Abstract

The invention relates to an unmanned vehicle obstacle identification method and system, and the method comprises the steps: obtaining surrounding point cloud data in real time through employing a laser radar sensor equipped on an unmanned vehicle, carrying out the noise filtering and down-sampling of the point cloud data, and calculating the normal vector and local point cloud density of each point in the point cloud data; for each target point, calculating the included angle between the target point and the normal of the neighborhood point, calculating the normal consistency weight of the neighborhood point through the normal included angle, calculating the statistical filtering weight of the neighborhood point according to the local point cloud density, and taking the product of the normal consistency weight and the statistical filtering weight as the weight of the neighborhood point; calculating a fast point feature histogram of the target point by using the weight of the neighborhood point, the simple point feature histogram of the neighborhood point and the simple point feature histogram of the target point; and segmenting the point cloud through the positions of the points in the point cloud and the fast point feature histograms, classifying the segmented regions by using a classifier according to the fast point feature histogram of each segmented region, and determining the type and the distance of the obstacle.

Description

Technical Field

[0001] The present invention relates to the field of obstacle recognition, and in particular to an unmanned vehicle obstacle recognition method and system. Background Art

[0002] Unmanned vehicles include a variety of forms, from driverless cars to park logistics vehicles and express delivery robots. These unmanned vehicles can perceive the surrounding environment in real time, realize autonomous navigation, path planning and task execution by integrating advanced sensor technologies such as LiDAR, cameras, radar and ultrasonic sensors. In the field of logistics and distribution, unmanned vehicles can efficiently complete warehouse management, cargo handling and last-mile delivery tasks, significantly improve operational efficiency and reduce labor costs; in park management, unmanned vehicles can perform patrol monitoring, facility maintenance and environmental cleaning, and improve the intelligent management level of the park. In addition, the application of unmanned vehicles in industries such as medical care, manufacturing, and agriculture is also expanding, such as drug delivery robots in hospitals, automated transport vehicles in factories, and precision fertilization and harvesting robots in agriculture. These unmanned vehicles can not only improve operating efficiency and accuracy, but also replace manual operations in dangerous or harsh environments to ensure the safety of operators.

[0003] However, in practical applications, the point cloud data obtained by unmanned vehicles is often affected by many factors, such as sensor noise, environmental complexity, and non-uniform distribution of point cloud density, all of which may lead to errors and delays in obstacle recognition. Specifically, sensor noise introduces a large number of isolated or abnormal points, which are often mistaken as valid features in traditional feature extraction methods, thereby affecting the accuracy of the entire obstacle recognition. In addition, the density of point cloud data may vary significantly in different areas. For example, in open roads and densely built areas, the density changes of point clouds will cause feature descriptors to generate redundant information in high-density areas, while in low-density areas, there may be a lack of sufficient feature details. This density non-uniformity makes it difficult for traditional fast point feature histograms (FPFH) to balance computational efficiency and comprehensive feature expression when processing point cloud data in different environments, thereby limiting its application in real-time unmanned vehicle obstacle recognition. Summary of the invention

[0004] In order to improve the accuracy and calculation speed of obstacle recognition of unmanned vehicles, in a first aspect, the present invention provides an unmanned vehicle obstacle recognition method, the method comprising: The laser radar sensor equipped on the unmanned vehicle is used to obtain the surrounding point cloud data in real time, filter the noise and downsample the point cloud data, and calculate the normal vector and local point cloud density of each point in the point cloud data; For each target point, the angle between the target point and the normal of the neighboring point is calculated, the normal consistency weight of the neighboring point is calculated by the normal angle, and the statistical filtering weight of the neighboring point is calculated according to the local point cloud density, and the product of the normal consistency weight and the statistical filtering weight is used as the weight of the neighboring point; the fast point feature histogram of the target point is calculated using the weight of the neighboring point and the simple point feature histogram of the neighboring point and the simple point feature histogram of the target point; The point cloud is segmented by the position of the point in the point cloud and the fast point feature histogram. The classifier is used to classify the segmented area according to the fast point feature histogram of each segmented area, and the type and distance of the obstacle are determined.

[0005] Optionally, the point cloud is segmented by the position of the point in the point cloud and the fast point feature histogram, specifically: The position coordinates of each point and the fast point feature histogram are normalized respectively and then feature fusion is performed. The point cloud is clustered according to the fusion features using a clustering method or a machine learning method. After clustering, each cluster is a segmented area.

[0006] Optionally, the classifier is used to classify the segmented regions according to the fast point feature histogram of each segmented region, and determine the type and distance of the obstacle, specifically: Cluster the fast point feature histogram of each segmented area to obtain the average fast point feature histogram of each cluster; The average fast point feature histogram of all clusters in the segmented area is input into the classifier to obtain the probability of obstacles and the types of obstacles; The segmented area with a probability greater than the threshold is obtained, and the average distance from multiple points in the segmented area closest to the current distance of the unmanned vehicle to the unmanned vehicle is calculated using point cloud computing, and the average distance is used as the distance from the obstacle to the unmanned vehicle.

[0007] Optionally, the normal consistency weight of the neighborhood point is calculated by the normal angle, specifically: Obtain the point cloud data set collected during the operation of the unmanned vehicle, and calculate the standard deviation of the normal angle between any two points in the point cloud data set; The product of the standard deviation and the adjustment coefficient is used as the standard deviation of the Gaussian function, and the normal angle is input into the Gaussian function to obtain the normal consistency weight.

[0008] Optionally, the statistical filtering weights of the neighborhood points are calculated according to the local point cloud density, specifically: Obtain a density threshold, calculate the ratio of the local point cloud density of the neighborhood point to the density threshold, if the ratio is greater than 1, the statistical filtering weight is 1, if the ratio is less than 1, the statistical filtering weight is the ratio.

[0009] In another aspect, the present invention provides an unmanned vehicle obstacle recognition system, the system comprising: The point cloud preprocessing module uses the laser radar sensor equipped by the unmanned vehicle to obtain the surrounding point cloud data in real time, performs noise filtering and downsampling on the point cloud data, and calculates the normal vector and local point cloud density of each point in the point cloud data; The feature calculation module is used to calculate the angle between the target point and the normal of the neighboring point for each target point, calculate the normal consistency weight of the neighboring point through the normal angle, and calculate the statistical filtering weight of the neighboring point according to the local point cloud density, and use the product of the normal consistency weight and the statistical filtering weight as the weight of the neighboring point; calculate the fast point feature histogram of the target point using the weight of the neighboring point and the simple point feature histogram of the neighboring point and the simple point feature histogram of the target point; The obstacle recognition module is used to segment the point cloud according to the position of the point in the point cloud and the fast point feature histogram, classify the segmented area according to the fast point feature histogram of each segmented area using a classifier, and determine the type and distance of the obstacle.

[0010] Optionally, the point cloud is segmented by the position of the point in the point cloud and the fast point feature histogram, specifically: The position coordinates of each point and the fast point feature histogram are normalized respectively and then feature fusion is performed. The point cloud is clustered according to the fusion features using a clustering method or a machine learning method. After clustering, each cluster is a segmented area.

[0011] Optionally, the classifier is used to classify the segmented regions according to the fast point feature histogram of each segmented region, and determine the type and distance of the obstacle, specifically: Cluster the fast point feature histogram of each segmented area to obtain the average fast point feature histogram of each cluster; The average fast point feature histogram of all clusters in the segmented area is input into the classifier to obtain the probability of obstacles and the types of obstacles; The segmented area with a probability greater than the threshold is obtained, and the average distance from multiple points in the segmented area closest to the current distance of the unmanned vehicle to the unmanned vehicle is calculated using point cloud computing, and the average distance is used as the distance from the obstacle to the unmanned vehicle.

[0012] Optionally, the normal consistency weight of the neighborhood point is calculated by the normal angle, specifically: Obtain the point cloud data set collected during the operation of the unmanned vehicle, and calculate the standard deviation of the normal angle between any two points in the point cloud data set; The product of the standard deviation and the adjustment coefficient is used as the standard deviation of the Gaussian function, and the normal angle is input into the Gaussian function to obtain the normal consistency weight.

[0013] Optionally, the statistical filtering weights of the neighborhood points are calculated according to the local point cloud density, specifically: Obtain a density threshold, calculate the ratio of the local point cloud density of the neighborhood point to the density threshold, if the ratio is greater than 1, the statistical filtering weight is 1, if the ratio is less than 1, the statistical filtering weight is the ratio.

[0014] The present invention introduces normal consistency weights and statistical filtering weights in the FPFH feature calculation process to effectively suppress the interference of noise points and adapt to changes in point cloud density. The normal consistency weights calculate the angle between the target point and the normal of the neighborhood point, and assign higher weights to the neighborhood points with consistent normal directions, thereby reducing the negative impact of noise points on the FPFH feature due to inconsistent normal directions. At the same time, the statistical filtering weights are adjusted according to the local density of the neighborhood points. High-density areas are assigned lower weights to avoid feature redundancy, while low-density areas are assigned higher weights to supplement feature deficiencies. The present invention not only ensures the calculation efficiency of high-density areas, but also improves the feature expression capabilities of low-density areas, thereby improving recognition accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of Embodiment 1; Figure 2 This is a structural diagram of the second embodiment. DETAILED DESCRIPTION

[0016] In this article, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0018] Embodiment 1: This embodiment provides a method for identifying obstacles of an unmanned vehicle, the method comprising: S1 uses the lidar sensor equipped on the unmanned vehicle to obtain the surrounding point cloud data in real time, performs noise filtering and downsampling on the point cloud data, and calculates the normal vector and local point cloud density of each point in the point cloud data; The high-precision laser radar sensor installed on the unmanned vehicle can collect three-dimensional point cloud data of the environment around the unmanned vehicle in real time. The acquired raw point cloud data is firstly subjected to noise filtering using, for example, statistical outlier removal algorithms or the like, to remove noise points that are abnormally far away from their neighboring points, thereby improving the quality and reliability of the point cloud data.

[0019] For the denoised point cloud data, downsampling is performed using methods such as density-aware downsampling. Specifically, different downsampling ratios are set according to the density level of each point. For example, the downsampling ratio is set higher in high-density areas and lower in low-density areas. After noise filtering and downsampling, the normal vector is calculated for each point to obtain the surface geometric direction information of the point. In addition, the k-nearest neighbor algorithm is used to calculate the point cloud density of each point in its local area to obtain the local point cloud density.

[0020] S2, for each target point, calculate the angle between the target point and the normal of the neighboring point, calculate the normal consistency weight of the neighboring point through the normal angle, and calculate the statistical filtering weight of the neighboring point according to the local point cloud density, and take the product of the normal consistency weight and the statistical filtering weight as the weight of the neighboring point; calculate the fast point feature histogram of the target point using the weight of the neighboring point and the simple point feature histogram of the neighboring point and the simple point feature histogram of the target point; When the unmanned vehicle collects point clouds, there will inevitably be noise, which has a great impact on the subsequent recognition of obstacles, especially small obstacles. In addition, the noise points in the point cloud often appear as isolated points or outliers in spatial distribution. Through statistical filtering methods, each neighborhood point can be further given a weight based on its local density to suppress isolated noise points.

[0021] Specifically, the angle between the target point and the normal of the neighboring point is calculated, and the normal consistency weight of the neighboring point is calculated by the normal angle. In one embodiment, the point cloud data set collected during the operation of the unmanned vehicle is obtained, and the standard deviation of the normal angle between any two points in the point cloud data set is counted; the product of the standard deviation and the adjustment coefficient is used as the standard deviation of the Gaussian function, and the normal angle is input into the Gaussian function to obtain the normal consistency weight. A large number of point cloud data sets collected by the unmanned vehicle through the laser radar sensor during operation are obtained, and the standard deviation of the normal angle between any two points in the point cloud data set is counted. The standard deviation is an important indicator for measuring the consistency of the normal direction, which reflects the overall distribution of the normal direction in the point cloud data. The obtained normal angle standard deviation is multiplied by the preset adjustment coefficient to obtain the standard deviation parameter of the Gaussian function; wherein the adjustment coefficient is used to control the attenuation rate of the Gaussian function, so as to adapt to the needs under different environments and point cloud density conditions. The normal angle between each target point and its neighboring point is input into the Gaussian function, and the corresponding normal consistency weight is calculated using the previously determined standard deviation parameter. The Gaussian function makes the neighborhood points with smaller normal angles obtain higher weights, while the neighborhood points with larger angles obtain lower weights, thereby effectively suppressing the negative impact of noise points and points with inconsistent directions on feature description, and improving the accuracy and robustness of the fast point feature histogram (FPFH).

[0022] When calculating the statistical filtering weight, a density threshold needs to be determined first. The density threshold is used to distinguish high-density areas from low-density areas in the point cloud data. The density threshold can be determined as a representative value by statistically analyzing a large amount of collected point cloud data. Calculate the local point cloud density of each neighborhood point, that is, the number or density of points around the target point in the neighborhood range. Then, compare this local density value with the previously determined density threshold to obtain a ratio. This ratio reflects the relative relationship between the density level of the current neighborhood point and the preset standard. Assume that the density threshold is 20 points per cubic meter. If the local point cloud density of a neighborhood point is 25 points, then the ratio of its density to the threshold is 1.25; if the local point cloud density of another neighborhood point is 10 points, then the ratio is 0.5. According to the calculated density ratio, the statistical filtering weight is assigned to the neighborhood point to dynamically adjust its influence in the FPFH feature calculation.

[0023] After calculating the normal consistency weight and statistical filtering weight of the neighborhood point of the target point, the product of the two is used as the weight of the neighborhood point. In another alternative embodiment, the reciprocal of the distance from the target point to the neighborhood point is calculated, and the product of the reciprocal and the product of the two is used as the weight of the neighborhood point.

[0024] After obtaining the weights of the neighborhood points, the fast point feature histogram of the target point is calculated according to the FPFH calculation formula. The specific calculation process is to calculate the simple point feature histogram (SPFH) of all neighborhood points of the target point, weight the SPFH of the neighborhood points using the weights of the neighborhood points, and then average the weighted SPFH of all neighborhood points to obtain a first result; calculate the simple point feature histogram (SPFH) of the target point, and use the sum of the simple point feature histogram of the target point and the first result as the FPFH of the target point.

[0025] S3, segmenting the point cloud by the position of the point in the point cloud and the fast point feature histogram, using a classifier to classify the segmented area according to the fast point feature histogram of each segmented area, and determining the type and distance of the obstacle.

[0026] In order to effectively identify and classify obstacles, the point cloud data is segmented, that is, the point cloud is divided into multiple areas with similar features or spatial relationships. The segmentation process depends on the position coordinates of each point (such as x, y, z coordinates) and its corresponding fast point feature histogram (FPFH). Specifically, the position coordinates of each point and the fast point feature histogram are normalized respectively, and then feature fusion is performed. The point cloud is clustered according to the fusion features using a clustering method or a machine learning method. After clustering, each cluster is a segmented area. The position coordinates of the points in the point cloud data are normalized, and then the FPFH is normalized. The normalized position coordinates and FPFH are spliced, and then the splicing results are clustered, for example, using DBSCAN for clustering. DBSCAN can automatically identify clusters of any shape. In another embodiment, the position of the points in the point cloud and the fast point feature histogram are input into the neural network, and the points are classified according to the neural network, and then the points of the same type are divided into a segmented area.

[0027] After completing the point cloud segmentation, each segmented area corresponds to a potential obstacle. In order to further determine the specific types of these obstacles (such as pedestrians, vehicles, static obstacles, etc.), a classifier is used to classify the FPFH feature histogram of each area. The classifier learns the characteristic patterns of different obstacle types to achieve accurate classification of the newly segmented area; wherein, the classifier preferably uses a support vector machine, random forest or neural network model.

[0028] In an optional embodiment, the classifier is used to classify the segmented regions according to the fast point feature histogram of each segmented region, and determine the type and distance of the obstacle, specifically: Cluster the fast point feature histogram of each segmented area to obtain the average fast point feature histogram of each cluster; The average fast point feature histogram of all clusters in the segmented area is input into the classifier to obtain the probability of obstacles and the types of obstacles; The segmented area with a probability greater than the threshold is obtained, and the average distance from multiple points in the segmented area closest to the current distance of the unmanned vehicle to the unmanned vehicle is calculated using point cloud computing, and the average distance is used as the distance from the obstacle to the unmanned vehicle.

[0029] After the point cloud data segmentation is completed, each segmented area contains multiple points and their corresponding FPFH feature histograms. In order to further simplify the feature representation and extract more representative geometric information, this embodiment performs cluster analysis on the FPFH features in each segmented area to form several clusters, and calculates the average FPFH feature histogram of each cluster. Specifically, the K-Means clustering algorithm is used to cluster the FPFH feature histograms in each segmented area. For each cluster, the dimension-by-dimensional average of all FPFH feature histograms within it is calculated to form the average FPFH feature histogram of the cluster. After obtaining the average FPFH feature histogram of each cluster, these feature vectors are input into a pre-trained classifier to predict the obstacle type and probability. After the classifier completes the obstacle type prediction, the segmented areas with probabilities higher than the preset threshold are further screened out to ensure that only high-confidence recognition results are subsequently processed. At the same time, by calculating the average distance of multiple points closest to the unmanned vehicle in these high-confidence areas, such as 10 or 20 points, the actual distance between the obstacle and the unmanned vehicle is determined, providing accurate distance information for path planning and obstacle avoidance decisions.

[0030] Embodiment 2: This embodiment provides an obstacle recognition system for an unmanned vehicle, such as Figure 2 As shown, the system comprises: The point cloud preprocessing module uses the laser radar sensor equipped by the unmanned vehicle to obtain the surrounding point cloud data in real time, performs noise filtering and downsampling on the point cloud data, and calculates the normal vector and local point cloud density of each point in the point cloud data; The feature calculation module is used to calculate the angle between the target point and the normal of the neighboring point for each target point, calculate the normal consistency weight of the neighboring point through the normal angle, and calculate the statistical filtering weight of the neighboring point according to the local point cloud density, and use the product of the normal consistency weight and the statistical filtering weight as the weight of the neighboring point; calculate the fast point feature histogram of the target point using the weight of the neighboring point and the simple point feature histogram of the neighboring point and the simple point feature histogram of the target point; The obstacle recognition module is used to segment the point cloud according to the position of the point in the point cloud and the fast point feature histogram, classify the segmented area according to the fast point feature histogram of each segmented area using a classifier, and determine the type and distance of the obstacle.

[0031] Optionally, the point cloud is segmented by the position of the point in the point cloud and the fast point feature histogram, specifically: The position coordinates of each point and the fast point feature histogram are normalized respectively and then feature fusion is performed. The point cloud is clustered according to the fusion features using a clustering method or a machine learning method. After clustering, each cluster is a segmented area.

[0032] Optionally, the classifier is used to classify the segmented regions according to the fast point feature histogram of each segmented region, and determine the type and distance of the obstacle, specifically: Cluster the fast point feature histogram of each segmented area to obtain the average fast point feature histogram of each cluster; The average fast point feature histogram of all clusters in the segmented area is input into the classifier to obtain the probability of obstacles and the types of obstacles; The segmented area with a probability greater than the threshold is obtained, and the average distance from multiple points in the segmented area closest to the current distance of the unmanned vehicle to the unmanned vehicle is calculated using point cloud computing, and the average distance is used as the distance from the obstacle to the unmanned vehicle.

[0033] Optionally, the normal consistency weight of the neighborhood point is calculated by the normal angle, specifically: Obtain the point cloud data set collected during the operation of the unmanned vehicle, and calculate the standard deviation of the normal angle between any two points in the point cloud data set; The product of the standard deviation and the adjustment coefficient is used as the standard deviation of the Gaussian function, and the normal angle is input into the Gaussian function to obtain the normal consistency weight.

[0034] Optionally, the statistical filtering weights of the neighborhood points are calculated according to the local point cloud density, specifically: Obtain a density threshold, calculate the ratio of the local point cloud density of the neighborhood point to the density threshold, if the ratio is greater than 1, the statistical filtering weight is 1, if the ratio is less than 1, the statistical filtering weight is the ratio.

[0035] Through the description of the above implementation methods, technicians in this field can clearly understand that each implementation method can be implemented by adding a necessary general hardware platform, and of course can also be implemented by combining hardware and software. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a computer product, and the present invention can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it, and other embodiments may also be used. Although the present invention has been described in detail with reference to the aforementioned embodiments, ordinary technicians in this field should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying obstacles of an unmanned vehicle, characterized in that: The method comprises: The laser radar sensor equipped on the unmanned vehicle is used to obtain the surrounding point cloud data in real time, filter the noise and downsample the point cloud data, and calculate the normal vector and local point cloud density of each point in the point cloud data; For each target point, the angle between the target point and the normal of the neighboring point is calculated, the normal consistency weight of the neighboring point is calculated by the normal angle, and the statistical filtering weight of the neighboring point is calculated according to the local point cloud density, and the product of the normal consistency weight and the statistical filtering weight is used as the weight of the neighboring point; the fast point feature histogram of the target point is calculated using the weight of the neighboring point and the simple point feature histogram of the neighboring point and the simple point feature histogram of the target point; The point cloud is segmented by the position of the point in the point cloud and the fast point feature histogram. The classifier is used to classify the segmented area according to the fast point feature histogram of each segmented area, and the type and distance of the obstacle are determined.

2. The method according to claim 1, characterized in that The point cloud segmentation by the position of the point in the point cloud and the fast point feature histogram is specifically as follows: The position coordinates of each point and the fast point feature histogram are normalized respectively and then feature fusion is performed. The point cloud is clustered according to the fusion features using a clustering method or a machine learning method. After clustering, each cluster is a segmented area.

3. The method according to claim 1, characterized in that The classifier is used to classify the segmented areas according to the fast point feature histogram of each segmented area, and determine the type and distance of the obstacle, specifically: Cluster the fast point feature histogram of each segmented area to obtain the average fast point feature histogram of each cluster; The average fast point feature histogram of all clusters in the segmented area is input into the classifier to obtain the probability of obstacles and the types of obstacles; The segmented area with a probability greater than the threshold is obtained, and the average distance from multiple points in the segmented area closest to the current distance of the unmanned vehicle to the unmanned vehicle is calculated using point cloud computing, and the average distance is used as the distance from the obstacle to the unmanned vehicle.

4. The method according to claim 1, characterized in that The calculation of the normal consistency weight of the neighborhood point by the normal angle is specifically: Obtain the point cloud data set collected during the operation of the unmanned vehicle, and calculate the standard deviation of the normal angle between any two points in the point cloud data set; The product of the standard deviation and the adjustment coefficient is used as the standard deviation of the Gaussian function, and the normal angle is input into the Gaussian function to obtain the normal consistency weight.

5. The method according to claim 1, characterized in that The statistical filtering weights of the neighborhood points are calculated according to the local point cloud density, specifically: Obtain a density threshold, calculate the ratio of the local point cloud density of the neighborhood point to the density threshold, if the ratio is greater than 1, the statistical filtering weight is 1, if the ratio is less than 1, the statistical filtering weight is the ratio.

6. An obstacle recognition system for an unmanned vehicle, characterized in that: The system comprises: The point cloud preprocessing module uses the laser radar sensor equipped by the unmanned vehicle to obtain the surrounding point cloud data in real time, performs noise filtering and downsampling on the point cloud data, and calculates the normal vector and local point cloud density of each point in the point cloud data; The feature calculation module is used to calculate the angle between the target point and the normal of the neighboring point for each target point, calculate the normal consistency weight of the neighboring point through the normal angle, and calculate the statistical filtering weight of the neighboring point according to the local point cloud density, and use the product of the normal consistency weight and the statistical filtering weight as the weight of the neighboring point; calculate the fast point feature histogram of the target point using the weight of the neighboring point and the simple point feature histogram of the neighboring point and the simple point feature histogram of the target point; The obstacle recognition module is used to segment the point cloud according to the position of the point in the point cloud and the fast point feature histogram, classify the segmented area according to the fast point feature histogram of each segmented area using a classifier, and determine the type and distance of the obstacle.

7. The system according to claim 6, characterized in that The point cloud segmentation by the position of the point in the point cloud and the fast point feature histogram is specifically as follows: The position coordinates of each point and the fast point feature histogram are normalized respectively and then feature fusion is performed. The point cloud is clustered according to the fusion features using a clustering method or a machine learning method. After clustering, each cluster is a segmented area.

8. The system according to claim 6, characterized in that The classifier is used to classify the segmented areas according to the fast point feature histogram of each segmented area, and determine the type and distance of the obstacle, specifically: Cluster the fast point feature histogram of each segmented area to obtain the average fast point feature histogram of each cluster; The average fast point feature histogram of all clusters in the segmented area is input into the classifier to obtain the probability of obstacles and the types of obstacles; The segmented area with a probability greater than the threshold is obtained, and the average distance from multiple points in the segmented area closest to the current distance of the unmanned vehicle to the unmanned vehicle is calculated using point cloud computing, and the average distance is used as the distance from the obstacle to the unmanned vehicle.

9. The system according to claim 6, characterized in that The calculation of the normal consistency weight of the neighborhood point by the normal angle is specifically: Obtain the point cloud data set collected during the operation of the unmanned vehicle, and calculate the standard deviation of the normal angle between any two points in the point cloud data set; The product of the standard deviation and the adjustment coefficient is used as the standard deviation of the Gaussian function, and the normal angle is input into the Gaussian function to obtain the normal consistency weight.

10. The system according to claim 6, characterized in that The statistical filtering weights of the neighborhood points are calculated according to the local point cloud density, specifically: Obtain a density threshold, calculate the ratio of the local point cloud density of the neighborhood point to the density threshold, if the ratio is greater than 1, the statistical filtering weight is 1, if the ratio is less than 1, the statistical filtering weight is the ratio.

Images