A method, device and electronic device for lidar target recognition

Abstract

The present invention provides a method, apparatus and electronic device for lidar target recognition, which are applied to the field of automotive technology. In this method, after obtaining the point cloud to be processed by the lidar and performing structured processing on the point cloud to be processed to obtain a structured point cloud, the laser points corresponding to the ground points in the structured point cloud are invalidated to obtain the processed structured point cloud. Then, the laser points in the processed structured point cloud are clustered to obtain at least one point cloud cluster. Each point cloud cluster is respectively input into a feature extraction network to obtain the corresponding point cloud cluster feature vector. Finally, each point cloud cluster feature vector is input into a target recognition network to obtain the radar target. By performing structured processing and clustering processing on the point cloud to be processed, this method can significantly reduce the amount of data that the recognition model needs to process, and thus can simplify the network structure of the recognition model, reduce the hardware performance requirements for the electronic device, and effectively improve the recognition efficiency of the radar target.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobiles, and particularly relates to a method, device and electronic device for lidar target recognition. Background Art

[0002] In the fields of assisted driving and intelligent driving, in-vehicle lidar is widely used as the main means to detect obstacles around the vehicle. In each detection cycle, the in-vehicle lidar will feedback a point cloud composed of a number of laser points, and the vehicle-mounted controller completes the recognition of radar targets around the vehicle by analyzing the point cloud.

[0003] With the development of neural network technology, target recognition of point clouds through a target recognition model trained based on neural networks has become a widely used technical means. In practical applications, after the vehicle-mounted controller obtains the point cloud feedback by the lidar, it directly inputs the point cloud into the target recognition model, and then obtains the radar target recognition result.

[0004] [[ID=^{14}]]However, the inventor has found through research that the point cloud data volume feedback by the lidar is huge. Directly performing radar target recognition based on the point cloud results in a complex network model structure of the target recognition model, extremely large computational complexity, high requirements for the hardware performance of the vehicle-mounted controller, and low radar target recognition efficiency. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method, device and electronic device for lidar target recognition, which reduce the computational complexity in the radar target recognition process, reduce the requirements for hardware devices, and improve the recognition efficiency. The specific solutions are as follows:

[0006] In a first aspect, the present invention provides a method for lidar target recognition, including:

[0007] Obtaining the point cloud to be processed of the lidar;

[0008] Performing structural processing on the point cloud to be processed to obtain a structured point cloud;

[0009] Performing invalid processing on the laser points corresponding to the ground points in the structured point cloud to obtain a processed structured point cloud;

[0010] Clustering the laser points in the processed structured point cloud to obtain at least one point cloud cluster;

[0011] Respectively inputting each point cloud cluster into a feature extraction network to obtain corresponding point cloud cluster feature vectors;

[0012] Wherein, the feature extraction network takes the performance parameters of each laser point in the point cloud cluster as input and the feature vector of the point cloud cluster as output, and is obtained by training a neural network;

[0013] Input each of the point cloud cluster feature vectors into a target recognition network to obtain radar targets;

[0014] Among them, the target recognition network takes each of the point cloud cluster feature vectors as input and the radar targets included in each of the point cloud clusters as output, and is obtained by training a neural network.

[0015] Optionally, the structured processing of the to-be-processed point cloud to obtain a structured point cloud includes:

[0016] Use the beam number value and horizontal angle value corresponding to each laser point in the to-be-processed point cloud as the sorting index of the corresponding laser point;

[0017] Sort each laser point in the to-be-processed point cloud according to the sorting index to obtain a structured point cloud.

[0018] Optionally, the clustering of the laser points in the processed structured point cloud to obtain at least one point cloud cluster includes:

[0019] Take each valid laser point in the processed structured point cloud as a target laser point respectively;

[0020] Among them, the valid laser point is a laser point other than the laser point corresponding to the ground point in the processed structured point cloud;

[0021] Obtain the target clustering threshold corresponding to the target laser point;

[0022] Take the laser points in the processed structured point cloud that take the target laser point as the center and the sorting index is within a preset screening range as candidate laser points;

[0023] Calculate the Euclidean distance between each candidate laser point and the target laser point respectively to obtain the screening distance corresponding to each candidate laser point;

[0024] Cluster the candidate laser points with a screening distance less than the target clustering threshold and the target laser point to obtain the corresponding point cloud cluster.

[0025] Optionally, the obtaining of the target clustering threshold corresponding to the target laser point includes:

[0026] Calculate the Euclidean distance between the target laser point and the preset origin of the lidar to obtain the reference distance corresponding to the target laser point;

[0027] Determine the target clustering threshold according to the magnitude relationship between the reference distance and the reference distance threshold.

[0028] Optionally, the determining of the target clustering threshold according to the magnitude relationship between the reference distance and the reference distance threshold includes:

[0029] If the reference distance is less than the first reference distance threshold, the first clustering threshold is used as the target clustering threshold;

[0030] If the reference distance is greater than the second reference distance threshold, the second clustering threshold is used as the target clustering threshold;

[0031] If the reference distance is greater than or equal to the first reference distance threshold and less than or equal to the second reference distance threshold, the product of the reference distance and a preset proportionality coefficient is used as the target clustering threshold;

[0032] Wherein, the first reference distance threshold is less than the second reference distance threshold;

[0033] The first clustering threshold is less than the second clustering threshold.

[0034] Optionally, the step of using each valid laser point in the structured point cloud as a target laser point respectively includes:

[0035] According to the sequence of the sorting indexes of the valid laser points in the processed structured point cloud, each of the valid laser points is used as a target laser point respectively.

[0036] Optionally, the step of inputting each of the point cloud clusters into the feature extraction network to obtain corresponding point cloud cluster feature vectors includes:

[0037] Combining the three-dimensional coordinates and reflection intensities corresponding to the laser points in each of the point cloud clusters into a two-dimensional vector matrix;

[0038] Inputting the two-dimensional matrix into the feature extraction network, and after passing through the feature extraction network, the two-dimensional vector matrix is converted into a point cloud cluster feature vector of the same length, where the point cloud cluster feature vector is a one-dimensional vector.

[0039] Optionally, the step of inputting each of the point cloud cluster feature vectors into the target recognition network to obtain a radar target includes:

[0040] Combining each of the point cloud cluster feature vectors into a point cloud cluster feature matrix;

[0041] Inputting the point cloud cluster feature matrix into the target recognition network to obtain a radar target.

[0042] In a second aspect, the present invention provides a lidar target recognition device, including:

[0043] An acquisition unit, configured to acquire the point cloud to be processed of the lidar;

[0044] A first processing unit, configured to perform structured processing on the point cloud to be processed to obtain a structured point cloud;

[0045] A second processing unit for invalidating the laser points corresponding to the ground points in the structured point cloud to obtain a processed structured point cloud;

[0046] A clustering unit for clustering the laser points in the processed structured point cloud to obtain at least one point cloud cluster;

[0047] A feature extraction unit for respectively inputting each of the point cloud clusters into a feature extraction network to obtain corresponding point cloud cluster feature vectors;

[0048] Wherein, the feature extraction network takes the performance parameters of each laser point in the point cloud cluster as input and the feature vector of the point cloud cluster as output, and is obtained by training a neural network;

[0049] An identification unit for inputting each of the point cloud cluster feature vectors into a target identification network to obtain radar targets;

[0050] Wherein, the target identification network takes each of the point cloud cluster feature vectors as input and the radar targets included in each of the point cloud clusters as output, and is obtained by training a neural network.

[0051] In a third aspect, the present invention provides an electronic device, comprising: a memory and a processor;

[0052] The memory stores a program suitable for the processor to execute to implement the lidar target recognition method according to any one of the first aspects of the present invention.

[0053] Based on the above technical solutions, in the lidar target recognition method provided by the present invention, after obtaining the point cloud to be processed by the lidar and performing structured processing on the point cloud to be processed to obtain a structured point cloud, the laser points corresponding to the ground points in the structured point cloud are invalidated to obtain a processed structured point cloud, and then the laser points in the processed structured point cloud are clustered to obtain at least one point cloud cluster, each point cloud cluster is respectively input into a feature extraction network to obtain corresponding point cloud cluster feature vectors, and finally each point cloud cluster feature vector is input into a target recognition network to obtain radar targets. Compared with the prior art, this method can significantly reduce the amount of data that the recognition model needs to process through the structured processing and clustering processing of the point cloud to be processed, and thus can simplify the network structure of the recognition model. Through the adjustment of both the amount of data and the network structure, the hardware performance requirements for the electronic device are reduced. Correspondingly, in the case of using an electronic device with the same performance, the recognition efficiency of radar targets can be effectively improved by this method.

[0054] Further, when performing target recognition, first extract features from each point cloud cluster, and then perform radar target recognition. The point cloud cluster features obtained through the feature extraction model are more refined than the three-dimensional coordinates and reflection intensities in the point cloud to be processed. Compared with the prior art, the recognition effect is better than that of networks of the same level. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0056] Figure 1 is a flowchart of a lidar target recognition method provided by an embodiment of the present invention;

[0057] Figure 2 is a schematic diagram of the process of extracting a point cloud cluster feature vector provided by an embodiment of the present invention;

[0058] Figure 3 is a schematic diagram of a radar target recognition process provided by an embodiment of the present invention;

[0059] Figure 4 is a structural block diagram of a lidar target recognition device provided by an embodiment of the present invention;

[0060] Figure 5 is a structural block diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0062] The lidar target recognition method provided by the present invention is applied to an electronic device. The electronic device can be a controller corresponding to the lidar or other controllers capable of obtaining the point cloud feedback from the lidar. For example, when the lidar is applied to a vehicle, a vehicle controller or an autonomous driving controller can be selected. Of course, in some cases, it can also be applied to a server on the network side.

[0063] See Figure 1 , Figure 1It is a flowchart of the lidar target recognition method provided by an embodiment of the present invention. The process of the lidar target recognition method provided by this embodiment may include:

[0064] S100. Obtain the point cloud to be processed by the lidar.

[0065] The point cloud to be processed mentioned in the embodiment of the present invention refers to any frame of point cloud fed back during the operation of the lidar, which is the original data fed back by the lidar. Based on related technologies, the point cloud to be processed includes multiple laser points. Of course, the specific number of laser points is related to the number of objects within the detection range of the lidar and the performance parameters of the lidar itself.

[0066] S110. Perform structured processing on the point cloud to be processed to obtain a structured point cloud.

[0067] Assume that the point cloud to be processed is P0, and P0 can be expressed as [N0, M], where N0 is the number of all laser points in the point cloud to be processed, and M represents the three-dimensional coordinates and reflection intensity of any laser point. In addition, based on the relevant principles of the radar, each laser point can also correspond to a unique identification information based on the performance parameters of the lidar. Specifically, the performance parameters mentioned in this embodiment include the number of beams and the horizontal angle value of the lidar. Taking a lidar with 32 beams and a horizontal resolution of 0.2° as an example, the corresponding number of laser points should be 32 * 360 / 0.2 = 57600. Correspondingly, for any laser point, there is a unique line number value and horizontal angle value. Therefore, the line number value and horizontal angle value corresponding to each laser point can be used as the number of the corresponding laser point. In this embodiment, this number is defined as the sorting index.

[0068] Based on the foregoing content, during structured processing, sort the individual laser points in the point cloud to be processed according to the corresponding sorting index to obtain a structured point cloud. For example, first, for the first beam, sort the corresponding laser points in the order of the horizontal angle value from 0° to 360°; then, for the second beam, sort the corresponding laser points in the same order of the horizontal angle from 0° to 360°, and so on, until all laser points are sorted to obtain a structured point cloud.

[0069] S120. Invalidate the laser points corresponding to the ground points in the structured point cloud to obtain a processed structured point cloud.

[0070] Since the point cloud to be processed inevitably contains laser points corresponding to ground points, in order to reduce the computational load, it is necessary to mark the laser points corresponding to these ground points in the structured point cloud as invalid laser points, and finally obtain the processed structured point cloud. Optionally, in order to avoid affecting the structure of the structured point cloud, the laser points corresponding to the ground points can be marked as invalid laser points by setting them to null values. Specifically, the process of setting the laser points corresponding to the ground points to null values can be expressed as P0[i] = [nan, nan, nan, nan], where i is the sorting index corresponding to the ground points. Further, the processed structured point cloud can be marked as P.

[0071] It should be noted that the determination of the laser points corresponding to the ground points in the point cloud to be processed can be achieved in combination with related technologies, and the present invention does not limit this.

[0072] S130. Cluster the laser points in the processed structured point cloud to obtain at least one point cloud cluster.

[0073] When clustering the laser points of the processed structured point cloud, it should be ensured that each laser point corresponding to a non-ground point in the processed structured point cloud is traversed, and at the same time, it is ensured that each laser point corresponding to a non-ground point determines the point cloud cluster to which it belongs.

[0074] Optionally, based on the above principle, the laser points other than the laser points corresponding to the ground points in the processed structured point cloud are defined as valid laser points. According to the order of the sorting indexes of the valid laser points in the processed structured point cloud, each valid laser point is used as a target laser point, and then the target clustering threshold corresponding to the target laser point is determined. It should be noted here that in the method provided in the embodiments of the present invention, based on the Euclidean distance between the target laser point and the preset origin of the lidar, as the basis for selecting the target clustering threshold, it can be understood that the farther the target laser point is from the preset origin, the sparser the point cloud density, and the corresponding clustering threshold should be larger; the closer the target laser point is to the preset origin, the denser the point cloud density, and the corresponding clustering threshold should be smaller.

[0075] Specifically, calculate the Euclidean distance between the target laser point and the preset origin of the lidar. In this embodiment, this distance is defined as the reference distance corresponding to the target laser point, and then the target clustering threshold corresponding to the target laser point can be determined according to the following formula based on the magnitude relationship between the reference distance and the preset reference distance threshold.

[0076]

[0077] where T h represents the target clustering threshold corresponding to the target laser point;

[0078] Dp represents the reference distance of the target laser point;

[0079] D min represents the first reference distance threshold;

[0080] D max represents the second reference distance threshold, D max is greater than D min ;

[0081] R represents a preset proportionality coefficient;

[0082] T min represents the first clustering threshold;

[0083] T max represents the second clustering threshold, T max is greater than T min .

[0084] It can be seen from the above formula that if the reference distance of the target laser point is less than the first reference distance threshold, the first clustering threshold is used as the target clustering threshold; correspondingly, if the obtained reference distance is greater than the second reference distance threshold, the second clustering threshold is used as the target clustering threshold; if the reference distance of the target laser point is greater than or equal to the first reference distance threshold and less than or equal to the second reference distance threshold, the product of the reference distance and the preset proportionality coefficient is used as the target clustering threshold. Among them, the preset proportionality coefficient can be flexibly set based on the specific performance parameters of the lidar and the actual recognition requirements, and the present invention does not limit the specific value of the preset proportionality coefficient.

[0085] After determining the target clustering threshold corresponding to the target laser point according to the above content, it is possible to determine whether other laser points can be clustered into one category based on the Euclidean distance between the target laser point and other laser points. Specifically, if the Euclidean distance between any laser point and the target laser point is less than the target clustering threshold, the laser point can be clustered with the target laser point; on the contrary, if the Euclidean distance between the laser point and the target laser point is greater than or equal to the target clustering threshold, the two cannot be clustered into one category.

[0086] It can be imagined that according to the foregoing content, it can be known that the structured point cloud includes a large number of laser points. If the traversal method is used to determine whether other laser points other than the target laser point can be clustered with the target laser point respectively, the calculation amount is very large, and a large amount of hardware resources of the electronic device must be occupied.

[0087] To solve the above problems, in this embodiment, a preset screening range is set. Laser points in the structured point cloud that are centered on the target laser point and whose sorting indices are within this preset screening range are used as candidate laser points. Then, only the determination of whether the candidate laser points can be clustered with the target laser point is performed. That is, the Euclidean distances between each candidate laser point and the target laser point are calculated respectively to obtain the screening distances corresponding to the candidate laser points, and the candidate laser points with screening distances less than the target clustering threshold are clustered with the target laser point to obtain the corresponding point cloud clusters.

[0088] By setting the preset screening range, the computational complexity of the clustering process can be effectively reduced, thereby improving the efficiency of target recognition. In practical applications, the specific determination of the preset screening range needs to be flexibly selected in combination with the specific number of laser points in the structured point cloud, as well as the requirements for recognition accuracy and efficiency. The present invention does not limit the specific setting of the preset screening range.

[0089] After traversing all the laser points in the structured point cloud, at least one point cloud cluster can be obtained through clustering. Through the clustering process, the rough-graining process of the point cloud to be processed is realized.

[0090] In practical applications, the above clustering process can be implemented using the following program segment:

[0091]

[0092]

[0093] In the above program segment, point represents the laser point, point’s around are the laser points within the preset screening range around the point point, Dist(around,point) is the Euler distance between the target laser point and other laser points, and Threshold(point) is the target clustering threshold corresponding to the target laser point.

[0094] Using the above clustering method, the processed structured point cloud P is clustered to generate a point cloud cluster set (p0, p1, p2…, pN), where N is the total number of point cloud clusters after clustering. Among them, each point cloud cluster can be represented as a two-dimensional point cloud matrix [Npi, M], where Npi is the number of laser points in the pi-th point cloud cluster, which is the original information of the lidar sensor.

[0095] S140. Input each point cloud cluster into the feature extraction network respectively to obtain the corresponding point cloud cluster feature vectors.

[0096] The lidar target recognition method provided in this embodiment provides a pre-trained feature extraction network. This feature extraction network takes the performance parameters of each laser point in the point cloud cluster as input and outputs the feature vector of the point cloud cluster. It is obtained by training a neural network. Among them, the performance parameters of the laser point mainly include the three-dimensional coordinates and reflection intensity of the laser point.

[0097] After obtaining each point cloud cluster, the three-dimensional coordinates and reflection intensity corresponding to each laser point in each point cloud cluster are combined into a two-dimensional vector matrix. Further, the two-dimensional matrix is input into the aforementioned feature extraction network. After passing through the feature extraction network, the two-dimensional vector matrix is converted into a point cloud cluster feature vector of the same length. Among them, the point cloud cluster feature vector is a one-dimensional vector.

[0098] Specifically, refer to Figure 2 , Figure 2 It is a schematic diagram of the process of extracting the feature vector of the point cloud cluster by applying the feature extraction network provided in the embodiment of the present invention.

[0099] The input of the feature extraction network is the three-dimensional coordinates and reflection intensity of each laser point in the clustered point cloud clusters (p0, p1, p2... pN), and the output is the feature vector (Fp0, Fp1, Fp2,... FpN) of each point cloud cluster. N is the number of clustered point cloud clusters. The feature extraction process can be expressed by the following formula:

[0100] Fpi = f(pi), i = 0, 1, 2... N

[0101] Optionally, refer to Figure 2 , the feature extraction network provided in the embodiment of the present invention realizes the above process of f(pi). The feature extraction network can be extracted by a method similar to pointnet, and its network structure can be referred to Figure 2 as shown.

[0102] Since after passing through the feature extraction network, the feature of each point cloud cluster changes from a two-dimensional matrix [Npi, M] to a one-dimensional vector Fpi of the same length, all the feature vectors of the point cloud clusters can be combined into a two-dimensional matrix [N, F], where N is the number of clustered point cloud clusters, and F is the length of the feature vector output by the feature extraction network.

[0103] S150. Input the feature vectors of each point cloud cluster into the target recognition network to obtain the radar target.

[0104] Optionally, the embodiment of the present invention provides a pre-trained target recognition network. This target recognition network takes the feature vectors of each point cloud cluster as input and outputs the radar targets included in each point cloud cluster. It is obtained by training a neural network.

[0105] Based on the above, after the point cloud cluster feature vectors of each point cloud cluster are output in S140, the feature vectors of each point cloud cluster are combined into a point cloud cluster feature matrix [N, F], where N is the number of point cloud clusters after clustering, and F is the length of the point cloud cluster feature vector output by the aforementioned feature extraction network. Using this point cloud cluster feature matrix as the input, combined with specific requirements, the trained target recognition network is used to classify, bounding box regression, segment, etc. for each cluster of point clouds, and the radar target extraction in the entire point cloud to be processed is completed. This process can be referred to Figure 3 as shown.

[0106] Optionally, if other information needs to be included in the recognition result, such as the heading angle information, a network layer for recognizing the heading angle can also be added when training the aforementioned model.

[0107] In summary, the lidar target recognition method provided by the present invention can significantly reduce the amount of data that the recognition model needs to process through the structured processing and clustering processing of the point cloud to be processed, and thus can simplify the network structure of the recognition model. Through the adjustment of both the amount of data and the network structure, the hardware performance requirements for electronic devices are reduced. Correspondingly, in the case of using electronic devices with the same performance, the recognition efficiency of radar targets can be effectively improved by this method.

[0108] Taking a mechanical 32-line lidar as an example, the number of original point clouds is 1800 * 32. After clustering, about 300 point cloud clusters can be obtained, and the number of point clouds is reduced by nearly 200 times. At the same time, the number of point cloud clusters is related to the scene and has no obvious relationship with the number of beams and models of the lidar. That is, if a lidar with a higher number of beams is used, the amount of data reduced will be more obvious. For example, compared with an 80-line lidar, the number of its point clouds is nearly 500 times less.

[0109] Furthermore, feature extraction is first performed on each point cloud cluster, and then radar target recognition is performed. Compared with the prior art, the recognition effect is better than that of networks of the same level. The point cloud cluster features obtained through the feature extraction model are more refined than the three-dimensional coordinates and reflection intensities in the point cloud to be processed, and are more conducive to the subsequent first target recognition model to achieve better extraction and recognition effects with a simpler network.

[0110] Next, the lidar target recognition device provided by the embodiments of the present invention will be introduced. The lidar target recognition device described below can be considered as a functional module architecture that needs to be set in the central device to implement the lidar target recognition method provided by the embodiments of the present invention. The content described below can be referred to each other with the above content.

[0111] Figure 4 is a structural block diagram of a lidar target recognition device provided by an embodiment of the present invention. Referring to Figure 4 , the device may include:

[0112] An acquisition unit for acquiring the point cloud to be processed of the lidar;

[0113] A first processing unit for performing structured processing on the point cloud to be processed to obtain a structured point cloud;

[0114] A second processing unit for invalidating the laser points corresponding to the ground points in the structured point cloud to obtain a processed structured point cloud;

[0115] A clustering unit for clustering the laser points in the processed structured point cloud to obtain at least one point cloud cluster;

[0116] A feature extraction unit for respectively inputting each point cloud cluster into a feature extraction network to obtain a corresponding point cloud cluster feature vector;

[0117] Wherein, the feature extraction network takes the performance parameters of each laser point in the point cloud cluster as input and the feature vector of the point cloud cluster as output, and is obtained by training a neural network;

[0118] An identification unit for inputting each point cloud cluster feature vector into a target identification network to obtain a radar target;

[0119] Wherein, the target identification network takes each point cloud cluster feature vector as input and the radar targets included in each point cloud cluster as output, and is obtained by training a neural network.

[0120] Optionally, the first processing unit 20 for performing structured processing on the point cloud to be processed to obtain a structured point cloud includes:

[0121] Taking the line bundle value and the horizontal angle value corresponding to each laser point in the point cloud to be processed as the sorting index of the corresponding laser point;

[0122] Sorting each laser point in the point cloud to be processed according to the sorting index to obtain a structured point cloud.

[0123] Optionally, the clustering unit 40 for clustering the laser points in the processed structured point cloud to obtain at least one point cloud cluster includes:

[0124] Taking each valid laser point in the processed structured point cloud as a target laser point respectively;

[0125] Wherein, the valid laser point is a laser point other than the laser point corresponding to the ground point in the processed structured point cloud;

[0126] Obtaining the target clustering threshold corresponding to the target laser point;

[0127] Taking the laser points in the processed structured point cloud that are centered on the target laser point and whose sorting index is within a preset screening range as candidate laser points;

[0128] Calculate the Euclidean distance between each candidate laser point and the target laser point respectively to obtain the screening distance corresponding to each candidate laser point;

[0129] Cluster the candidate laser points with screening distances less than the target clustering threshold and the target laser point to obtain the corresponding point cloud clusters.

[0130] Optionally, the clustering unit 40 is used to obtain the target clustering threshold corresponding to the target laser point, including:

[0131] Calculate the Euclidean distance between the target laser point and the preset origin of the lidar to obtain the reference distance corresponding to the target laser point;

[0132] Determine the target clustering threshold according to the magnitude relationship between the reference distance and the reference distance threshold.

[0133] Optionally, the clustering unit 40 is used to determine the target clustering threshold according to the magnitude relationship between the reference distance and the reference distance threshold, including:

[0134] If the reference distance is less than the first reference distance threshold, take the first clustering threshold as the target clustering threshold;

[0135] If the reference distance is greater than the second reference distance threshold, take the second clustering threshold as the target clustering threshold;

[0136] If the reference distance is greater than or equal to the first reference distance threshold and less than or equal to the second reference distance threshold, take the product of the reference distance and the preset proportionality coefficient as the target clustering threshold;

[0137] Among them, the first reference distance threshold is less than the second reference distance threshold;

[0138] The first clustering threshold is less than the second clustering threshold.

[0139] Optionally, the clustering unit 40 is used to take each valid laser point in the structured point cloud as the target laser point respectively, including:

[0140] According to the sequence of the sorting indexes of each valid laser point in the processed structured point cloud, take each valid laser point as the target laser point respectively.

[0141] Optionally, the feature extraction unit 50 is used to input each point cloud cluster into the feature extraction network respectively to obtain the corresponding point cloud cluster feature vector, including:

[0142] Combine the three-dimensional coordinates and reflection intensities corresponding to each laser point in each point cloud cluster into a two-dimensional vector matrix;

[0143] Input the two-dimensional matrix into the feature extraction network. After passing through the feature extraction network, the two-dimensional vector matrix is converted into a point cloud cluster feature vector of the same length, where the point cloud cluster feature vector is a one-dimensional vector.

[0144] Optionally, the recognition unit 60 is configured to input each point cloud cluster feature vector into the target recognition network to obtain a radar target, including:

[0145] Combine each point cloud cluster feature vector into a point cloud cluster feature matrix;

[0146] Input the point cloud cluster feature matrix into the target recognition network to obtain a radar target.

[0147] Figure 5 The structural block diagram of the electronic device provided by the embodiment of the present invention is shown in Figure 5 As shown, it may include: at least one processor 100, at least one communication interface 200, at least one memory 300, and at least one communication bus 400;

[0148] In the embodiment of the present invention, the number of the processor 100, the communication interface 200, the memory 300, and the communication bus 400 is at least one, and the processor 100, the communication interface 200, and the memory 300 complete mutual communication through the communication bus 400; obviously, Figure 5 The communication connection schematic diagram of the processor 100, the communication interface 200, the memory 300, and the communication bus 400 shown is only optional;

[0149] Optionally, the communication interface 200 may be an interface of a communication module, such as an interface of a GSM module;

[0150] The processor 100 may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention. ]>

[0151] The memory 300 stores an application program, which may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.

[0152] Among them, the processor 100 is specifically configured to execute the application program in the memory to implement any one of the above-described lidar target recognition methods.

[0153] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0154] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0155] The steps of the methods or algorithms described in combination with the embodiments disclosed herein can be directly implemented by hardware, software modules executed by a processor, or a combination of the two. The software modules can be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0156] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the core idea or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for lidar target recognition, characterized in that, Including: Obtaining the point cloud to be processed of the lidar; Performing structured processing on the point cloud to be processed to obtain a structured point cloud; Performing invalid processing on the laser points corresponding to the ground points in the structured point cloud to obtain a processed structured point cloud; Clustering the laser points in the processed structured point cloud to obtain at least one point cloud cluster; Inputting each of the point cloud clusters into a feature extraction network respectively to obtain corresponding point cloud cluster feature vectors; Wherein, the feature extraction network takes the performance parameters of each laser point in the point cloud cluster as input and the feature vector of the point cloud cluster as output, and is obtained by training a neural network; Inputting each of the point cloud cluster feature vectors into a target recognition network to obtain radar targets; Wherein, the target recognition network takes each of the point cloud cluster feature vectors as input and the radar targets included in each of the point cloud clusters as output, and is obtained by training a neural network; Wherein, each laser point in the structured point cloud is sorted according to the sorting index of each laser point, and the clustering of the laser points in the processed structured point cloud to obtain at least one point cloud cluster includes: Taking each valid laser point in the processed structured point cloud as a target laser point respectively; the valid laser point is a laser point other than the laser point corresponding to the ground point in the processed structured point cloud; Obtaining a target clustering threshold corresponding to the target laser point; the obtaining of the target clustering threshold corresponding to the target laser point includes: calculating the Euclidean distance between the target laser point and a preset origin of the lidar to obtain a reference distance corresponding to the target laser point; determining the target clustering threshold according to the magnitude relationship between the reference distance and a reference distance threshold; the distance between the target laser point and the preset origin is directly proportional to the magnitude of the target clustering threshold; Taking the laser points in the processed structured point cloud that take the target laser point as the center and whose sorting index is within a preset screening range as candidate laser points; Calculating the Euclidean distance between each of the candidate laser points and the target laser point respectively to obtain a screening distance corresponding to each of the candidate laser points; Clustering the candidate laser points with a screening distance less than the target clustering threshold and the target laser point to obtain a corresponding point cloud cluster.

2. The lidar target recognition method according to claim 1, wherein The performing structured processing on the point cloud to be processed to obtain a structured point cloud includes: Taking the line bundle value and horizontal angle value corresponding to each laser point in the point cloud to be processed as the sorting index of the corresponding laser point; Sorting each laser point in the point cloud to be processed according to the sorting index to obtain a structured point cloud.

3. The lidar target recognition method according to claim 1, wherein The determining the target clustering threshold according to the magnitude relationship between the reference distance and the reference distance threshold includes: If the reference distance is less than a first reference distance threshold, taking a first clustering threshold as the target clustering threshold; If the reference distance is greater than a second reference distance threshold, taking a second clustering threshold as the target clustering threshold; If the reference distance is greater than or equal to the first reference distance threshold and less than or equal to the second reference distance threshold, taking the product of the reference distance and a preset proportionality coefficient as the target clustering threshold; Wherein, the first reference distance threshold is less than the second reference distance threshold; The first clustering threshold is less than the second clustering threshold.

4. The lidar target recognition method according to claim 2, characterized in that, Regarding each valid laser point in the structured point cloud as a target laser point respectively includes: Regarding each of the valid laser points as a target laser point respectively in the order of the sorting indexes of the valid laser points in the processed structured point cloud.

5. The lidar target recognition method according to claim 4, wherein The step of inputting each of the point cloud clusters into the feature extraction network respectively to obtain corresponding point cloud cluster feature vectors includes: Combining the three-dimensional coordinates and reflection intensities corresponding to the laser points in each of the point cloud clusters into a two-dimensional vector matrix; Inputting the two-dimensional vector matrix into the feature extraction network, and after passing through the feature extraction network, the two-dimensional vector matrix is converted into a point cloud cluster feature vector of the same length, where the point cloud cluster feature vector is a one-dimensional vector.

6. The lidar target recognition method according to claim 5, wherein The step of inputting each of the point cloud cluster feature vectors into the target recognition network to obtain radar targets includes: Combining each of the point cloud cluster feature vectors into a point cloud cluster feature matrix; Inputting the point cloud cluster feature matrix into the target recognition network to obtain radar targets.

7. A lidar target recognition device, characterized in that, It includes: An acquisition unit, configured to acquire the point cloud to be processed by the lidar; A first processing unit, configured to perform structured processing on the point cloud to be processed to obtain a structured point cloud; A second processing unit, configured to perform invalid processing on the laser points corresponding to the ground points in the structured point cloud to obtain a processed structured point cloud; A clustering unit, configured to cluster the laser points in the processed structured point cloud to obtain at least one point cloud cluster; A feature extraction unit, configured to input each of the point cloud clusters into the feature extraction network respectively to obtain corresponding point cloud cluster feature vectors; Wherein, the feature extraction network takes the performance parameters of each laser point in the point cloud cluster as input and the feature vector of the point cloud cluster as output, and is obtained by training a neural network; An identification unit, configured to input each of the point cloud cluster feature vectors into the target recognition network to obtain radar targets; Wherein, the target recognition network takes each of the point cloud cluster feature vectors as input and the radar targets included in each of the point cloud clusters as output, and is obtained by training a neural network; Wherein, each laser point in the structured point cloud is sorted according to the sorting index of each laser point, and the clustering unit clusters the laser points in the processed structured point cloud to obtain at least one point cloud cluster, including: Regarding each valid laser point in the processed structured point cloud as a target laser point respectively; the valid laser point is a laser point other than the laser point corresponding to the ground point in the processed structured point cloud; Obtaining the target clustering threshold corresponding to the target laser point; the step of obtaining the target clustering threshold corresponding to the target laser point includes: calculating the Euclidean distance between the target laser point and the preset origin of the lidar to obtain the reference distance corresponding to the target laser point; determining the target clustering threshold according to the magnitude relationship between the reference distance and the reference distance threshold; the distance between the target laser point and the preset origin is directly proportional to the magnitude of the target clustering threshold; Regarding the laser points in the processed structured point cloud that take the target laser point as the center and whose sorting indexes are within the preset screening range as candidate laser points; Calculate the Euclidean distance between each of the candidate laser points and the target laser point respectively to obtain the screening distance corresponding to each of the candidate laser points; Cluster the candidate laser points with a screening distance less than the target clustering threshold with the target laser point to obtain corresponding point cloud clusters.

8. An electronic device, characterized in that, Comprising: A memory and a processor; The memory stores a program suitable for execution by the processor to implement the lidar target recognition method according to any one of claims 1-6.

Images