Target identification method and device, computer equipment and readable storage medium
By projecting and eliminating unreasonable data on multi-frame point cloud data, weak voxel grids are enhanced, and the problem of low target recognition accuracy in traditional methods is solved, and the recognition accuracy and adaptability of the target recognition model are improved.
Patent Information
- Application Number
- CN202411037698.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-15
AI Technical Summary
Traditional target recognition model training based on non-time sequence point clouds has problems such as target dispersion and inability to fully present target details, resulting in low accuracy of target recognition.
By obtaining the continuously collected multi-frame point cloud data, using the positioning information and object labeling information of adjacent frame point cloud data for projection, unreasonable point cloud data are eliminated, and point cloud data of weak voxel grids are enhanced to form a more complete Point cloud data is used to train target recognition models.
It improves the accuracy of target recognition of the target recognition model, solves the problem of high data time and speed requirements in traditional methods, and data fusion causes target characteristics to change or become background, and enhances the model's ability to identify targets.
Smart Images

Figure CN118960718A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of target detection, and in particular to a target recognition method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Art
[0002] In the fields of autonomous driving, medicine, smart home, security inspection and protection, there is a wide demand for target detection technology. To ensure the safety of users, target detection technology must be real-time and accurate. Point cloud-based target detection has been widely used in industry, academia and life due to its high-precision positioning capability and fast and real-time recognition capability.
[0003] In traditional methods, non-temporal point clouds are used to train target recognition models, and the trained target recognition models are used for target recognition. However, this non-temporal point cloud has challenges such as target dispersion and inability to fully present target details. Therefore, the target recognition model trained by traditional methods has the problem of low target recognition accuracy. Summary of the invention
[0004] Based on this, it is necessary to provide a target recognition method, device, computer equipment, computer-readable storage medium and computer program product that can improve the accuracy of target recognition in response to the above technical problems.
[0005] In a first aspect, the present application provides a target recognition method, comprising:
[0006] Obtain multiple frames of point cloud data obtained by continuous acquisition, and obtain the positioning information and object annotation information corresponding to each frame of point cloud data;
[0007] For any frame of point cloud data, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data, the multiple adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the projected point cloud data and the projected annotation information;
[0008] The target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition.
[0009] In one embodiment, based on the positioning information and object annotation information corresponding to each of the plurality of adjacent frame point cloud data, the plurality of adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the projected point cloud data and the projected annotation information, including:
[0010] Based on the positioning information of the point cloud data of the multiple adjacent frames, the point cloud data of the multiple adjacent frames are projected onto the point cloud data of the current frame to obtain the initial projected point cloud data;
[0011] Determine the intersection-and-union ratio between the object annotation information corresponding to each adjacent frame point cloud data and the object annotation information corresponding to the current frame point cloud data;
[0012] According to the intersection-and-union ratios corresponding to the point cloud data of multiple adjacent frames, unreasonable point cloud data in the initial projection point cloud data are eliminated to obtain the projection point cloud data;
[0013] The projection annotation information is determined according to the intersection-and-union ratios and object annotation information corresponding to each of the plurality of adjacent frame point cloud data.
[0014] In one embodiment, each frame of point cloud data has multiple object annotation information corresponding to it; according to the intersection-and-union ratios corresponding to the multiple adjacent frames of point cloud data, unreasonable point cloud data in the initial projected point cloud data is eliminated to obtain projected point cloud data, including:
[0015] In the case where any intersection-and-union ratio is greater than a first preset value and less than a second preset value, determining target adjacent frame point cloud data corresponding to the intersection-and-union ratio, and determining target object labeling information from a plurality of object labeling information corresponding to the target adjacent frame point cloud data;
[0016] Point cloud data corresponding to the target object annotation information in the initial projection point cloud data is eliminated to obtain projection point cloud data, wherein the first preset value is less than the second preset value.
[0017] In one embodiment, each frame of point cloud data has multiple object annotation information corresponding to it; and determining the projection annotation information according to the intersection-and-union ratio and the object annotation information corresponding to each of the multiple adjacent frames of point cloud data includes:
[0018] In the case where any intersection-and-union ratio is equal to a first preset value, determining the target adjacent frame point cloud data corresponding to the intersection-and-union ratio, and determining the target object annotation information from the multiple object annotation information corresponding to the target adjacent frame point cloud data; and using the target object annotation information and the multiple object annotation information corresponding to the current frame point cloud data together as the projection annotation information;
[0019] In the case where any intersection-over-union ratio is greater than the first set value, multiple object annotation information corresponding to the current frame point cloud data is used as projection annotation information.
[0020] In one embodiment, training a target recognition model based on projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data includes:
[0021] Determine the object category corresponding to at least one voxel grid in each frame of point cloud data;
[0022] Determine a distribution matrix corresponding to each object category according to the distribution of each object category in the projected point cloud data corresponding to each frame of point cloud data;
[0023] For the projected point cloud data corresponding to any frame of point cloud data, all object categories in the projected point cloud data are traversed, and weak voxel grids are determined in at least one voxel grid of the projected point cloud data according to the distribution matrix corresponding to each object category, and the point cloud data corresponding to the weak voxel grid in the projected point cloud data is enhanced;
[0024] The target recognition model is trained based on the enhanced projection point cloud data and projection annotation information corresponding to each frame of point cloud data.
[0025] In one embodiment, determining a weak voxel grid in at least one voxel grid of the projected point cloud data according to the distribution matrix corresponding to each object category includes:
[0026] When any element value in any distribution matrix is less than a preset threshold, a voxel grid corresponding to the element value in at least one voxel grid of the projected point cloud data is used as a weak voxel grid.
[0027] In one embodiment, the point cloud data corresponding to the weak voxel grid in the projected point cloud data is enhanced, including:
[0028] The point cloud data corresponding to the weak voxel grid in the multi-frame point cloud data is used as supplementary point cloud data;
[0029] The supplementary point cloud data is used to replace the point cloud data corresponding to the weak voxel grid in the projected point cloud data to obtain enhanced projected point cloud data.
[0030] In a second aspect, the present application also provides a target recognition device, comprising:
[0031] An acquisition module is used to acquire multiple frames of point cloud data acquired continuously, and to acquire the positioning information and object annotation information corresponding to each frame of point cloud data;
[0032] A projection module is used to project multiple adjacent frame point cloud data onto the current frame point cloud data based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data for any frame point cloud data, so as to obtain the projected point cloud data and the projection annotation information;
[0033] The recognition module is used to train a target recognition model based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition.
[0034] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0035] Obtain multiple frames of point cloud data obtained by continuous acquisition, and obtain the positioning information and object annotation information corresponding to each frame of point cloud data;
[0036] For any frame of point cloud data, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data, the multiple adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the projected point cloud data and the projected annotation information;
[0037] The target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition.
[0038] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the following steps are implemented:
[0039] Obtain multiple frames of point cloud data obtained by continuous acquisition, and obtain the positioning information and object annotation information corresponding to each frame of point cloud data;
[0040] For any frame of point cloud data, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data, the multiple adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the projected point cloud data and the projected annotation information;
[0041] The target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition.
[0042] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:
[0043] Obtain multiple frames of point cloud data obtained by continuous acquisition, and obtain the positioning information and object annotation information corresponding to each frame of point cloud data;
[0044] For any frame of point cloud data, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data, the multiple adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the projected point cloud data and the projected annotation information;
[0045] The target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition.
[0046] In the above-mentioned target recognition method, device, computer equipment, computer-readable storage medium and computer program product, multiple frames of point cloud data are continuously collected and have a temporal sequence. For any frame of point cloud data, the adjacent frame point cloud data adjacent to the current frame point cloud data is projected into the current frame point cloud data, which is conducive to fusing the point cloud data of the same object at different observation angles at multiple different times, making the current frame point cloud data more complete. Training the target recognition model with more complete point cloud data can improve the target recognition accuracy of the target recognition model. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0048] Figure 1 A diagram of an application environment of a target recognition method in an embodiment;
[0049] Figure 2 is a schematic flow chart of a target recognition method in one embodiment;
[0050] Figure 3 Schematic diagram of the overall process of a target recognition method in one embodiment;
[0051] Figure 4 is a structural block diagram of a target recognition device in one embodiment;
[0052] Figure 5 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0053] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0054] In the fields of autonomous driving, medicine, smart home, security inspection and protection, there is a wide demand for target detection technology. To ensure the safety of users, target detection technology must be real-time and accurate. Point cloud-based target detection has been widely used in industry, academia and life due to its high-precision positioning capability and fast and real-time recognition capability.
[0055] A traditional target recognition method uses non-sequential point clouds to train target recognition models, and the trained target recognition models are used for target recognition. However, this non-sequential point cloud has challenges such as target dispersion and inability to fully present target details. Therefore, the target recognition model trained by the traditional method has the problem of low target recognition accuracy.
[0056] Another traditional target recognition method completely projects the point cloud data of the previous and next two frames into the point cloud data of the current frame, so as to train the target recognition model based on the projected point cloud data. However, this method has high requirements on the acquisition time interval of the two frames of point cloud data and the relative speed between the point cloud acquisition device and the object. In a short time interval, when the point cloud acquisition device moves at a slow speed, projecting the point cloud data of the previous and next two frames into the point cloud of the current frame can enrich the point cloud data of each object in the point cloud of the current frame. However, when there is a large acquisition time interval (only the previous and next two frames can be used, and the effect will decrease if there are more) or there is a large relative speed between the point cloud acquisition device and the object, the same object in the two frames of point cloud data will have a large position interval after projection, resulting in the formation of abnormal object point cloud or the object becomes the background, thereby affecting the effect of the model. In addition, in practical applications, it is difficult to obtain point cloud data that meets the requirements.
[0057] Therefore, an embodiment of the present application proposes a target recognition method, in which multiple frames of point cloud data are collected continuously and have a time sequence relationship. For any frame of point cloud data, the adjacent frame point cloud data adjacent to the current frame point cloud data is projected into the current frame point cloud data, which is beneficial to the fusion of point cloud data of the same object at different times and different observation angles, making the current frame point cloud data more complete. Training the target recognition model with more complete point cloud data can improve the target recognition accuracy of the target recognition model.
[0058] The target recognition method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. This embodiment is illustrated by applying the method to the server 104. It can be understood that the method can also be applied to the terminal 102, and can also be applied to a system including the terminal 102 and the server 104, and is implemented through the interaction between the terminal 102 and the server 104. The server 104 obtains the multi-frame point cloud data obtained by continuous acquisition, and obtains the positioning information and object annotation information corresponding to each frame of point cloud data; for any frame of point cloud data, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frames of point cloud data, the multiple adjacent frames of point cloud data are projected to the current frame of point cloud data to obtain the projection point cloud data and projection annotation information; the target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition. The terminal 102 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car devices, projection devices, etc. Portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. Head-mounted devices may be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. The server 104 may be an independent physical server, or a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.
[0059] In an exemplary embodiment, Figure 2 As shown, a target recognition method is provided, which is applied to Figure 1 The server 104 in the example is used as an example to illustrate the method, which includes the following steps 202 to 206. Among them:
[0060] Step 202, obtain multiple frames of point cloud data acquired continuously, and obtain positioning information and object annotation information corresponding to each frame of point cloud data.
[0061] The point cloud data is data collected by a point cloud collection device, for example, the point cloud collection device may be a radar sensor.
[0062] The point cloud acquisition device obtains multiple frames of point cloud data through continuous acquisition, and the multiple frames of point cloud data have a time sequence.
[0063] For different application fields, the point cloud acquisition device can be installed in different installation devices, such as vehicles, medical equipment, smart home devices, and security inspection and security equipment. The positioning information may include the rotation angle information of the point cloud acquisition device when the point cloud acquisition device collects each frame of point cloud data, as well as the location information of the installation device.
[0064] Object annotation information refers to the annotation information corresponding to different objects in the point cloud data, and the object annotation information may include the size information, orientation angle information, etc. of the object. There may be multiple object annotation information corresponding to each frame of point cloud data.
[0065] During the data collection process of the point cloud acquisition device, there is often at least one object in the environment, such as a vehicle, pedestrian, roadblock, etc. Multiple frames of point cloud data are submitted to the annotation provider for object annotation, and the object annotation information corresponding to each frame of point cloud data will be obtained.
[0066] For example, for the scene where the radar in the vehicle collects point cloud data, the server obtains a set of n frames of point cloud data collected continuously, which can be expressed as [ ,..., ], the positioning information corresponding to this set of point cloud data can be expressed as , where the positioning information of the i-th frame can be expressed as:
[0067] ego i =[ R i T i 0 1 ]
[0068] here Represents the rotation matrix of the radar when collecting the i-th frame of point cloud data, Represents the translation matrix of the vehicle in the positioning map when collecting the i-th frame of point cloud data.
[0069] The object annotation information corresponding to this set of point cloud data can be If the object annotation information corresponding to the i-th frame point cloud data contains m objects, then we can have info i =[ sample 1 , sample 2 ,..., sample m ] Each object can be represented as , where (x, y, z) represents the coordinates of the center point of the object, (l, w, h) represents the length, width, and height of the object, and yaw represents the orientation angle of the object.
[0070] Step 204 , for any frame of point cloud data, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data, project the multiple adjacent frame point cloud data onto the current frame point cloud data to obtain the projected point cloud data and the projected annotation information.
[0071] Among them, since multiple frames of point cloud data have a time sequence relationship, the server can sort the multiple frames of point cloud data according to the time sequence relationship to obtain a point cloud data sequence. Adjacent frame point cloud data refers to the point cloud data frame adjacent to the current frame point cloud data in the point cloud data sequence. Multiple adjacent frame point cloud data may include multiple frames of point cloud data that are located before the current frame point cloud data in the point cloud data sequence, and before the current frame point cloud data.
[0072] Projection refers to the process of adding multiple adjacent frame point cloud data to the current frame point cloud data. With the projection of point cloud data, object annotation information can also be projected to obtain projection annotation information.
[0073] The projected point cloud data integrates the current frame point cloud data and multiple adjacent frame point cloud data. The projected annotation information integrates the object annotation information corresponding to the current frame point cloud data and the object annotation information corresponding to multiple adjacent frame point cloud data. After projection, the point cloud of each object in the current frame is enriched, which is conducive to improving the recognition accuracy of the model.
[0074] Step 206 , training a target recognition model based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and using the trained target recognition model for target recognition.
[0075] The server may perform the above projection process on each frame of point cloud data to obtain the projection point cloud data and projection annotation information corresponding to each frame of point cloud data, or may perform the above projection process once every preset frame. The projection point cloud data and projection annotation information corresponding to each of the obtained multiple frames of point cloud data may be used as training samples to train a target recognition model, which may be a machine learning model. Using the trained target recognition model for target recognition is beneficial to improving the accuracy of target recognition.
[0076] In the above target recognition method, multiple frames of point cloud data are collected continuously and have a temporal sequence. For any frame of point cloud data, the adjacent frame of point cloud data adjacent to the current frame of point cloud data is projected into the current frame of point cloud data, which is beneficial to the fusion of point cloud data of the same object at different times and different observation angles, making the current frame of point cloud data more complete. Training the target recognition model with more complete point cloud data can improve the target recognition accuracy of the target recognition model.
[0077] In an exemplary embodiment, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data, the multiple adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the projected point cloud data and the projection annotation information, including: based on the positioning information of the multiple adjacent frame point cloud data, the multiple adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the initial projected point cloud data; the intersection-and-union ratio between the object annotation information corresponding to each adjacent frame point cloud data and the object annotation information corresponding to the current frame point cloud data is determined; according to the intersection-and-union ratio corresponding to each of the multiple adjacent frame point cloud data, unreasonable point cloud data in the initial projected point cloud data is eliminated to obtain the projected point cloud data; according to the intersection-and-union ratio and object annotation information corresponding to each of the multiple adjacent frame point cloud data, the projection annotation information is determined.
[0078] The server projects multiple adjacent frame point cloud data to the current frame point cloud data. The projection process of any adjacent frame point cloud data is described as an example. For example, the adjacent frame point cloud data is the kth frame, which is recorded as , the current frame point cloud data is the tth frame, recorded as . Any point in , the point is projected to The point obtained is recorded as ,but:
[0079]
[0080]
[0081] According to this, we can All points in the projected In the example above, we get the initial projection point cloud data. .
[0082] In addition, the object annotation information of the kth frame can also be Object annotation information projected to the tth frame in sequence Above, recorded as .in, The jth target Project to The result after projection is recorded as sample j ' =[ x j ' , y j ' , z j ' , l j ' , w j ' , h j ' , yaw j ' ] ,in .
[0083] Can be converted to a rotation matrix . Can be converted to a translation matrix , similarly we can get , then:
[0084]
[0085] From this, we can find sample j ' =[ x j ' , y j ' , z j ' , l j ' , w j ' , h j ' , yaw j ' ] .
[0086] When the acquisition time interval between two consecutive frames of point cloud data is large, or the relative speed between the point cloud acquisition device and the object is high, the same object in different frames will have a large position interval after projection, resulting in the formation of abnormal object point clouds or the object becoming the background, thereby affecting the target recognition effect of the model. In order to solve this problem, the embodiment of the present application proposes to directly project multiple adjacent frames of point cloud data to obtain the initial projection point cloud data to remove unreasonable point cloud data, thereby avoiding the existence of overly singular feature structures in the projected point cloud data.
[0087] Specifically, the server can calculate the intersection and union ratio of the object annotation information corresponding to each adjacent frame point cloud data and the object annotation information corresponding to the current frame point cloud data, so as to remove unreasonable point cloud data according to the intersection and union ratio. The intersection and union ratio is used to characterize the degree of overlap between the object annotation information corresponding to the adjacent frame point cloud data and the object annotation information corresponding to the current frame point cloud data. For example, the higher the intersection and union ratio, the higher the degree of overlap between the adjacent frame point cloud data and the current frame point cloud data; the lower the intersection and union ratio, the lower the degree of overlap between the adjacent frame point cloud data and the current frame point cloud data.
[0088] The server can further determine the projection annotation information based on the intersection-and-union ratios corresponding to the point cloud data of multiple adjacent frames, combined with the object annotation information. For example, if the intersection-and-union ratio is greater than a preset value, the object annotation information corresponding to the point cloud data of the current frame can be used as the projection annotation information; if the intersection-and-union ratio is less than the preset value, the object annotation information corresponding to the point cloud data of the current frame and the object annotation information corresponding to the point cloud data of the adjacent frames can be used together as the projection annotation information.
[0089] In this embodiment, initial projected point cloud data is obtained by projecting multiple adjacent frame point cloud data onto the current frame point cloud data, and then unreasonable point cloud data is eliminated based on the intersection-and-union ratio between the object annotation information corresponding to the adjacent frame point cloud data and the object annotation information corresponding to the current frame point cloud data, and the projection annotation information is determined in combination with the object annotation information. This is beneficial for avoiding overly strange feature structures caused by the projection process, and can also prevent the projected point cloud from being too far from the original point cloud, causing the object features to become background features, affecting the model's recognition accuracy of the target.
[0090] In an exemplary embodiment, each frame of point cloud data corresponds to multiple object annotation information; according to the intersection-and-union ratios corresponding to each of the multiple adjacent frame point cloud data, unreasonable point cloud data in the initial projection point cloud data is eliminated to obtain projected point cloud data, including: when any intersection-and-union ratio is greater than a first preset value and less than a second preset value, the target adjacent frame point cloud data corresponding to the intersection-and-union ratio is determined, and the target object annotation information is determined in the multiple object annotation information corresponding to the target adjacent frame point cloud data; the point cloud data corresponding to the target object annotation information in the initial projection point cloud data is eliminated to obtain the projection point cloud data, wherein the first preset value is less than the second preset value.
[0091] Each frame of point cloud data may include point cloud data of multiple objects. In this case, there are multiple corresponding object annotation information.
[0092] The intersection-and-union ratio corresponding to each frame of point cloud data includes the intersection-and-union ratios corresponding to multiple objects. If any intersection-and-union ratio is greater than the first preset value and less than the second preset value, it indicates that the overlap between the annotation information of an object in a certain frame of point cloud data and the corresponding object annotation information in the current frame of point cloud data is between the first preset value and the second preset value. For example, the first preset value is 0, and the second preset value is 0.9. At this time, the server uses the adjacent frame point cloud data corresponding to the intersection-and-union ratio as the target adjacent frame point cloud data, and uses the object annotation information corresponding to the intersection-and-union ratio in the target adjacent frame point cloud data as the target object annotation information.
[0093] Since the initial projected point cloud data includes all adjacent frame point cloud data projected, the server regards the point cloud data with an intersection-and-union ratio between (0.1, 0.9) as unreasonable point cloud data and deletes it from the initial projected point cloud data to obtain the projected point cloud data. When the intersection-and-union ratio is within other ranges, the projected point cloud data can remain unchanged as the initial projected point cloud data.
[0094] For example, due to May be related to The object annotation information in the adjacent frame or the previously projected object annotation information overlaps, so the overlapping part can be removed. The object annotation information corresponding to the point cloud data of adjacent frames is expressed as , the object annotation information corresponding to the projected point cloud data is expressed as sample ' =[ x ' , y ' , z ' , l ' , w ' , h ' , yaw ' ] . Calculate the intersection and union ratio of the two ,if , then in Delete exist The projected point cloud data is framed in the middle.
[0095] In the present embodiment, when the intersection-and-union ratio is between the first preset value and the second preset value, the point cloud data of the object corresponding to the intersection-and-union ratio will produce a ghosting effect on the corresponding point cloud data in the current frame point cloud data. Therefore, the embodiment of the present application proposes to regard the point cloud data corresponding to the intersection-and-union ratio as unreasonable point cloud data, and remove it from the initial projected point cloud data. This can avoid the ghosting problem caused by the projection of point cloud data of multiple adjacent frames when the acquisition time interval between adjacent frames and the current frame is too long, or the relative speed between the point cloud acquisition device and the object is too high, thereby affecting the model's accurate recognition of the object.
[0096] In an exemplary embodiment, each frame of point cloud data corresponds to multiple object annotation information; projection annotation information is determined based on the intersection-and-union ratios and object annotation information corresponding to each of the multiple adjacent frame point cloud data, including: when any intersection-and-union ratio is equal to a first preset value, determining the target adjacent frame point cloud data corresponding to the intersection-and-union ratio, and determining the target object annotation information among the multiple object annotation information corresponding to the target adjacent frame point cloud data; using the target object annotation information and the multiple object annotation information corresponding to the current frame point cloud data together as projection annotation information; when any intersection-and-union ratio is greater than the first set value, using the multiple object annotation information corresponding to the current frame point cloud data as projection annotation information.
[0097] There are multiple object annotation information corresponding to each frame of point cloud data, and any intersection-and-union ratio is equal to the first preset value, indicating that there is an object annotation information in a certain frame of point cloud data and the corresponding object annotation information in the current frame of point cloud data overlaps to a degree equal to the first preset value. For example, the first preset value is 0. At this time, the server uses the adjacent frame point cloud data corresponding to the intersection-and-union ratio as the target adjacent frame point cloud data, and uses the object annotation information corresponding to the intersection-and-union ratio in the target adjacent frame point cloud data as the target object annotation information.
[0098] When the intersection-and-union ratio is equal to the first set value, the server uses the target object annotation information and the multiple object annotation information corresponding to the current frame point cloud data as the projection annotation information. In this way, the projection annotation information retains the object annotation information of the adjacent frame and the current frame at the same time, and since the intersection-and-union ratio is 0, this will not cause singular features.
[0099] When the intersection-over-union ratio is greater than a first set value, only the multiple object annotation information corresponding to the point cloud data of the current frame is used as the projection annotation information, and the target object annotation information in the adjacent frame is not retained.
[0100] For example, if , then Join In; if , then As .
[0101] In this embodiment, when any intersection-and-union ratio is equal to the first preset value, the projection annotation information simultaneously retains the object annotation information corresponding to the current frame and the adjacent frame; when any intersection-and-union ratio is greater than the first set value, the projection annotation information only retains the object annotation information corresponding to the current frame. This helps to avoid singular feature structures caused by the projection process, and avoid the object features becoming background features due to the projected point cloud data being too far away from the original point cloud data, which helps to improve the model's recognition accuracy of the target.
[0102] In an exemplary embodiment, a target recognition model is trained based on the projected point cloud data and projection annotation information corresponding to each frame of point cloud data, including: determining the object category corresponding to at least one voxel grid in each frame of point cloud data; determining the distribution matrix corresponding to each object category according to the distribution of each object category in the projected point cloud data corresponding to each frame of point cloud data; for the projected point cloud data corresponding to any frame of point cloud data, traversing all object categories in the projected point cloud data, determining weak voxel grids in at least one voxel grid of the projected point cloud data according to the distribution matrices corresponding to each object category, and enhancing the point cloud data corresponding to the weak voxel grid in the projected point cloud data; training the target recognition model based on the enhanced projected point cloud data and projection annotation information corresponding to each frame of point cloud data.
[0103] The voxel grid refers to a grid structure obtained by dividing the point cloud data into cubic units. Each cubic unit is called a voxel grid. Each frame of point cloud data can be divided into at least one voxel grid.
[0104] The object category refers to the category of the object indicated by the object identification information, such as vehicle category, pedestrian category, roadblock category, etc. Since each object can often occupy at least one voxel grid in each frame of point cloud data, the server can obtain the object category corresponding to each voxel grid.
[0105] The distribution refers to the number of times each object category appears in different voxel grids in different frames of point cloud data. The size of the distribution matrix can be the same as the size of the voxel grid contained in each frame of point cloud data. For example, if the voxel grid of each frame of point cloud data is 10×10, then the size of the distribution matrix is 10×10. The element value of each position in the distribution matrix corresponding to each object category is equal to the number of times the object category appears in the corresponding voxel grid. Since the distribution matrix is determined based on the projected point cloud data corresponding to each frame of point cloud data, each object category may appear multiple times in the same voxel grid.
[0106] A weak voxel grid refers to a voxel grid of a certain object category that appears infrequently in the projected point cloud data. When the model is trained, the projected point cloud data corresponding to each frame of point cloud data will be trained in sequence. If it is found during training that a certain object category exists in the projected point cloud data corresponding to the frame of point cloud data, the weak voxel grid can be enhanced to ensure that the weak voxel grid can be continuously learned during the model training process, which is beneficial to improving the model's target recognition accuracy.
[0107] In this embodiment, by determining the distribution matrix corresponding to each object category, the weak voxel grid is identified and enhanced to ensure that the weak voxel grid can be continuously learned during the model training process, which is beneficial to improving the target recognition accuracy of the model.
[0108] In an exemplary embodiment, a weak voxel grid is determined in at least one voxel grid of the projected point cloud data according to the distribution matrix corresponding to each object category, including: when any element value in any distribution matrix is less than a preset threshold, the voxel grid with the corresponding element value in at least one voxel grid of the projected point cloud data is used as a weak voxel grid.
[0109] Among them, the determination of weak voxel grids can be determined by whether the element value in the distribution matrix is less than a preset threshold. If the element value in the distribution matrix is less than the preset threshold, it means that the number of occurrences of the object category at the voxel grid corresponding to the element value is small. For such weak voxel grids, they may not be effectively learned during the model training process, thereby affecting the target recognition effect of the model.
[0110] In this embodiment, when the element value in the distribution matrix is less than a preset threshold, the voxel grid corresponding to the element value is used as a weak voxel grid, which is helpful for identifying the locations where the object category appears less frequently, thereby facilitating the subsequent enhancement processing of the point cloud at the corresponding location to improve the target recognition accuracy of the model.
[0111] In an exemplary embodiment, point cloud data corresponding to weak voxel grids in projected point cloud data are enhanced, including: using point cloud data corresponding to weak voxel grids in multiple frames of point cloud data as supplementary point cloud data; and replacing point cloud data corresponding to weak voxel grids in projected point cloud data with the supplementary point cloud data to obtain enhanced projected point cloud data.
[0112] Among them, since the number of times the object category appears at the position corresponding to the weak voxel grid is small, the embodiment of the present application eliminates the supplementary point cloud data at the position of the weak voxel grid to achieve the enhancement of the point cloud data.
[0113] The server can use the point cloud data corresponding to the weak voxel grid in the multi-frame point cloud data as supplementary point cloud data. Specifically, the point cloud data frame in which the weak voxel grid of the object category exists in the multi-frame point cloud data can be used as the target frame point cloud data, and the point cloud data corresponding to the weak voxel grid in the target frame point cloud data can be used as supplementary point cloud data.
[0114] The server uses the supplementary point cloud data to replace the point cloud data corresponding to the weak voxel grid in the projected point cloud data, so as to ensure that the point cloud data of the corresponding object category is added to the weak voxel grid in the projected point cloud data, thereby enhancing the projected point cloud data.
[0115] In some embodiments, if the supplementary point cloud data are point cloud data in multiple frames of point cloud data, the supplementary point cloud data corresponding to each frame can be added to the weak voxel grid of the projected point cloud data in a preset order, and the original point cloud data in the weak voxel grid of the projected point cloud data is deleted. As the supplementary point cloud data is added, the distribution matrix is updated in real time until the element value corresponding to the weak voxel grid exceeds the preset threshold, and the addition of the supplementary point cloud data is stopped.
[0116] In this embodiment, by determining supplementary point cloud data from multiple frames of point cloud data and replacing the point cloud data of the weak voxel grid, the weak position in the projected point cloud data is enhanced, which is beneficial to improving the recognition effect of the model.
[0117] To explain the target recognition method and effect in this solution in detail, a most detailed embodiment is described below:
[0118] Take the application of target recognition method in autonomous driving scenario as an example, the radar installed in the vehicle is used as the point cloud acquisition device. Figure 3 The figure shows a schematic diagram of the overall process of a target recognition method in an embodiment. Among them, the server obtains multiple frames of point cloud data obtained by continuous acquisition, and obtains the positioning information and object annotation information corresponding to each frame of point cloud data; for any frame of point cloud data, based on the positioning information of multiple adjacent frames of point cloud data, multiple adjacent frames of point cloud data are projected to the current frame of point cloud data to obtain initial projection point cloud data without skipping adjacent frames; the object annotation information corresponding to each of the multiple adjacent frames of point cloud data is projected to the current frame, and reasonable results are screened and unreasonable results are eliminated; the distribution matrix corresponding to each object category is counted to enhance the point cloud data with weak distribution; the target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data; and the trained target recognition model is used to perform target recognition.
[0119] In the above target recognition method, multiple frames of point cloud data are collected continuously and have a sequential relationship in time. For any frame of point cloud data, the adjacent frame of point cloud data adjacent to the current frame of point cloud data is projected into the current frame of point cloud data, which is conducive to the fusion of point cloud data of the same object at different observation angles at multiple different times, making the current frame of point cloud data more complete. Training the target recognition model with more complete point cloud data can improve the target recognition accuracy of the target recognition model. In addition, it solves the problem that the traditional method has high requirements for data time and speed, can only use data within two frames, and has high data requirements and is difficult to collect; solves the problem that data fusion in the traditional method will cause the target feature to change or the target to become the background; solves the problem that the model trained by the traditional method has poor recognition accuracy for certain voxel areas. The embodiment of the present application utilizes multiple frames of point cloud data with a time series relationship to train a target recognition model, which improves the problem of abnormal foreground and background features caused by mixed time series, greatly improves the target recognition accuracy of the trained target recognition model, and can be adapted to various special data sets, greatly reducing the requirements for the data set; by enhancing the weak points of the data distribution of each object category, the weak parts of the data set are targetedly enhanced, greatly improving the recognition ability of the object category in this area.
[0120] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.
[0121] Based on the same inventive concept, the embodiment of the present application also provides a target recognition device for implementing the target recognition method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more target recognition device embodiments provided below can refer to the limitations on the target recognition method above, and will not be repeated here.
[0122] In an exemplary embodiment, Figure 4 As shown, a target recognition device 400 is provided, comprising: an acquisition module 420, a projection module 440 and a recognition module 460, wherein:
[0123] The acquisition module 420 is used to acquire multiple frames of point cloud data acquired continuously, and acquire the positioning information and object annotation information corresponding to each frame of point cloud data;
[0124] The projection module 440 is used to project the point cloud data of multiple adjacent frames onto the current frame point cloud data for any frame point cloud data based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data to obtain the projected point cloud data and the projected annotation information;
[0125] The recognition module 460 is used to train a target recognition model based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition.
[0126] In the above-mentioned target recognition device, multiple frames of point cloud data are collected continuously and have a time sequence relationship. For any frame of point cloud data, the adjacent frame of point cloud data adjacent to the current frame of point cloud data is projected into the current frame of point cloud data, which is beneficial to the fusion of point cloud data of the same object at different observation angles at multiple different times, making the current frame of point cloud data more complete. Training the target recognition model with more complete point cloud data can improve the target recognition accuracy of the target recognition model.
[0127] In one embodiment, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data, the multiple adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the projected point cloud data and the projection annotation information. The projection module 440 is also used to: based on the positioning information of the multiple adjacent frame point cloud data, project the multiple adjacent frame point cloud data onto the current frame point cloud data to obtain the initial projected point cloud data; determine the intersection-and-union ratio between the object annotation information corresponding to each adjacent frame point cloud data and the object annotation information corresponding to the current frame point cloud data; according to the intersection-and-union ratio corresponding to each of the multiple adjacent frame point cloud data, eliminate the unreasonable point cloud data in the initial projected point cloud data to obtain the projected point cloud data; determine the projection annotation information according to the intersection-and-union ratio and object annotation information corresponding to each of the multiple adjacent frame point cloud data.
[0128] In one embodiment, each frame of point cloud data corresponds to multiple object annotation information; according to the intersection-and-union ratios corresponding to multiple adjacent frame point cloud data, unreasonable point cloud data in the initial projection point cloud data is eliminated to obtain projected point cloud data, and the projection module 440 is also used to: when any intersection-and-union ratio is greater than a first preset value and less than a second preset value, determine the target adjacent frame point cloud data corresponding to the intersection-and-union ratio, and determine the target object annotation information in the multiple object annotation information corresponding to the target adjacent frame point cloud data; eliminate the point cloud data corresponding to the target object annotation information in the initial projection point cloud data to obtain the projection point cloud data, wherein the first preset value is less than the second preset value.
[0129] In one embodiment, there are multiple object annotation information corresponding to each frame of point cloud data; the projection annotation information is determined according to the intersection-and-union ratios and object annotation information corresponding to the multiple adjacent frame point cloud data. The projection module 440 is also used to: when any intersection-and-union ratio is equal to a first preset value, determine the target adjacent frame point cloud data corresponding to the intersection-and-union ratio, and determine the target object annotation information among the multiple object annotation information corresponding to the target adjacent frame point cloud data; use the target object annotation information and the multiple object annotation information corresponding to the current frame point cloud data together as the projection annotation information; when any intersection-and-union ratio is greater than the first set value, use the multiple object annotation information corresponding to the current frame point cloud data as the projection annotation information.
[0130] In one embodiment, a target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each frame of point cloud data, and the recognition module 460 is also used to: determine the object category corresponding to at least one voxel grid in each frame of point cloud data; determine the distribution matrix corresponding to each object category according to the distribution of each object category in the projection point cloud data corresponding to each frame of point cloud data; for the projection point cloud data corresponding to any frame of point cloud data, traverse all object categories in the projection point cloud data, determine weak voxel grids in at least one voxel grid of the projection point cloud data according to the distribution matrix corresponding to each object category, and enhance the point cloud data corresponding to the weak voxel grid in the projection point cloud data; train the target recognition model based on the enhanced projection point cloud data and projection annotation information corresponding to each frame of point cloud data.
[0131] In one embodiment, a weak voxel grid is determined in at least one voxel grid of the projected point cloud data according to the distribution matrix corresponding to each object category. The identification module 460 is also used to: when any element value in any distribution matrix is less than a preset threshold, the voxel grid with the corresponding element value in at least one voxel grid of the projected point cloud data is used as a weak voxel grid.
[0132] In one embodiment, the point cloud data corresponding to the weak voxel grid in the projected point cloud data is enhanced, and the identification module 460 is also used to: use the point cloud data corresponding to the weak voxel grid in the multi-frame point cloud data as supplementary point cloud data; and use the supplementary point cloud data to replace the point cloud data corresponding to the weak voxel grid in the projected point cloud data to obtain enhanced projected point cloud data.
[0133] Each module in the above target recognition device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.
[0134] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 5 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store multi-frame point cloud data, positioning information, object annotation information, projection point cloud data, projection annotation information and target recognition models. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a target recognition method is implemented.
[0135] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0136] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.
[0137] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.
[0138] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.
[0139] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.
[0140] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.
[0141] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.
[0142] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.
Claims
1. A target recognition method, It is characterized in that The method comprises: Obtain multiple frames of point cloud data obtained by continuous acquisition, and obtain the positioning information and object annotation information corresponding to each frame of point cloud data; For any frame of point cloud data, based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data, the multiple adjacent frame point cloud data are projected onto the current frame point cloud data to obtain the projected point cloud data and the projected annotation information; The target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition.
2. The method according to claim 1, It is characterized in that The method of projecting the point cloud data of the multiple adjacent frames onto the point cloud data of the current frame based on the positioning information and the object annotation information corresponding to each of the point cloud data of the multiple adjacent frames to obtain the projected point cloud data and the projected annotation information includes: Based on the positioning information of the point cloud data of the multiple adjacent frames, the point cloud data of the multiple adjacent frames are projected onto the point cloud data of the current frame to obtain the initial projected point cloud data; For any adjacent frame point cloud data, determining an intersection-over-combination ratio between the object annotation information corresponding to the adjacent frame point cloud data and the object annotation information corresponding to the current frame point cloud data; Eliminate unreasonable point cloud data in the initial projected point cloud data according to the intersection-and-union ratios corresponding to the point cloud data of multiple adjacent frames, and obtain projected point cloud data; The projection annotation information is determined according to the intersection-and-union ratios and object annotation information corresponding to each of the plurality of adjacent frame point cloud data.
3. The method according to claim 2, It is characterized in that There are multiple object annotation information corresponding to each frame of point cloud data; according to the intersection-and-union ratios corresponding to the multiple adjacent frames of point cloud data, unreasonable point cloud data in the initial projected point cloud data are eliminated to obtain the projected point cloud data, including: In the case where any intersection-and-union ratio is greater than a first preset value and less than a second preset value, determining target adjacent frame point cloud data corresponding to the intersection-and-union ratio, and determining target object annotation information from a plurality of object annotation information corresponding to the target adjacent frame point cloud data; The point cloud data corresponding to the target object annotation information in the initial projected point cloud data is eliminated to obtain projected point cloud data, wherein the first preset value is less than the second preset value.
4. The method according to claim 2, It is characterized in that There are multiple object annotation information corresponding to each frame of point cloud data; the projection annotation information is determined according to the intersection-and-union ratio and object annotation information corresponding to each of the multiple adjacent frames of point cloud data, including: In the case where any intersection-and-union ratio is equal to a first preset value, determining the target adjacent frame point cloud data corresponding to the intersection-and-union ratio, and determining the target object annotation information from the multiple object annotation information corresponding to the target adjacent frame point cloud data; and using the target object annotation information and the multiple object annotation information corresponding to the current frame point cloud data together as projection annotation information; In the case where any intersection-over-union ratio is greater than a first set value, multiple object annotation information corresponding to the current frame point cloud data is used as projection annotation information.
5. The method according to claim 1, It is characterized in that The target recognition model is trained based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, including: Determine the object category corresponding to at least one voxel grid in each frame of point cloud data; Determine a distribution matrix corresponding to each object category according to the distribution of each object category in the projected point cloud data corresponding to each frame of point cloud data; For the projected point cloud data corresponding to any frame of point cloud data, all object categories in the projected point cloud data are traversed, and a weak voxel grid is determined in at least one voxel grid of the projected point cloud data according to the distribution matrix corresponding to each object category, and the point cloud data corresponding to the weak voxel grid in the projected point cloud data is enhanced; The target recognition model is trained based on the enhanced projection point cloud data and projection annotation information corresponding to each frame of point cloud data.
6. The method according to claim 5, It is characterized in that The step of determining a weak voxel grid in at least one voxel grid of the projected point cloud data according to the distribution matrix corresponding to each object category includes: When any element value in any distribution matrix is less than a preset threshold, a voxel grid corresponding to the element value in at least one voxel grid of the projection point cloud data is used as a weak voxel grid.
7. The method according to claim 5, It is characterized in that The step of enhancing the point cloud data corresponding to the weak voxel grid in the projected point cloud data comprises: Using the point cloud data corresponding to the weak voxel grid in the multi-frame point cloud data as supplementary point cloud data; The supplementary point cloud data is used to replace the point cloud data corresponding to the weak voxel grid in the projected point cloud data to obtain enhanced projected point cloud data.
8. A target recognition device, It is characterized in that The device comprises: An acquisition module is used to acquire multiple frames of point cloud data acquired continuously, and to acquire the positioning information and object annotation information corresponding to each frame of point cloud data; A projection module is used to project multiple adjacent frame point cloud data onto the current frame point cloud data based on the positioning information and object annotation information corresponding to each of the multiple adjacent frame point cloud data for any frame point cloud data, so as to obtain the projected point cloud data and the projection annotation information; The recognition module is used to train a target recognition model based on the projection point cloud data and projection annotation information corresponding to each of the multiple frames of point cloud data, and the trained target recognition model is used for target recognition.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program. It is characterized in that When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.
