Target recognition model training method and device and point cloud processing method and device

Abstract

The invention provides a target recognition model training method and device and a point cloud processing method and device, and relates to the field of smart mines, automatic driving and unmanned vehicles. The training method comprises the following steps: constructing a training data set, wherein the training data set comprises point cloud data about a target object and label data corresponding to the point cloud data; an initial network model is constructed, the initial network model comprises a feature extraction module and a plurality of task execution modules, the input end of each task execution module is connected with the output end of the feature extraction module, and at least part of the task execution modules in the plurality of task execution modules are mutually independent; and based on the training data set, training the initial network model until a predetermined performance index is reached, and obtaining a target recognition model. According to the invention, the flexibility and efficiency of the target identification model are improved, so that the target identification model can process more complex point cloud data, and effective identification of the target object in the mine unmanned operation environment is realized.

Description

Technical Field

[0001] The present application relates to the fields of smart mines, autonomous driving, and unmanned vehicles, and specifically to a method and device for training a target recognition model, and a method and device for processing a point cloud. Background Art

[0002] LiDAR point cloud detection technology can not only identify the location, size and category of traffic participants, but also label each point in the real-time point cloud data. Therefore, this technology plays a vital role in the field of unmanned driving, especially in mines.

[0003] However, the application of the above technologies in the mine unmanned driving scenario faces some challenges. Specifically, there are some objects in the mining area that will affect vehicle safety and driving efficiency. However, due to the distribution characteristics of such objects in the mining area and the lack of feature information of such objects themselves, the detection accuracy is low, which has become a technical problem that needs to be solved urgently in the field of mine unmanned driving. Summary of the invention

[0004] In view of this, the embodiments of the present application provide a method and device for training a target recognition model, and a method and device for processing a point cloud.

[0005] In a first aspect, an embodiment of the present application provides a method for training a target recognition model, comprising: constructing a training data set, the training data set comprising point cloud data about a target object and label data corresponding to the point cloud data, the label data being used to annotate and identify the target object; constructing an initial network model, the initial network model comprising a feature extraction module and multiple task execution modules, wherein an input end of each task execution module is connected to an output end of the feature extraction module, and at least some of the multiple task execution modules are independent of each other; based on the training data set, training the initial network model until a predetermined performance indicator is reached to obtain a target recognition model.

[0006] In combination with the first aspect, in certain implementations of the first aspect, the initial network model is trained based on a training data set, including: in a first stage, based on the training data set, the full link of the initial network model is trained; in a second stage, the weight value of the feature extraction module and the weight values ​​of other task execution modules except the target task execution module in a plurality of task execution modules are locked; based on the training data set, the target task execution module is trained; and the second stage training process is repeated for other task execution modules in a specific order until all task execution modules have been trained separately as target task modules.

[0007] In combination with the first aspect, in certain implementations of the first aspect, the feature extraction module includes a shared module, the multiple task execution modules include a first task module and a second task module, the first task module and the second task module are two interrelated modules, and the shared module is at least used to extract feature data common to the first task module and the second task module.

[0008] In combination with the first aspect, in certain implementations of the first aspect, the first task module is at least used to perform a bounding box detection task, and the second task module is at least used to perform a semantic segmentation task. In the process of training the full link of the initial network model, the first task module and the second task module jointly participate in the training to achieve mutual learning of the label data corresponding to the first task module and the second task module.

[0009] In combination with the first aspect, in certain implementations of the first aspect, constructing a training data set includes: obtaining original point cloud data collected for a target object and first label data corresponding to the original point cloud data; downsampling the original point cloud data according to a preset ratio to obtain downsampled point cloud data; adjusting the first label data to obtain second label data corresponding to the downsampled point cloud data; and using the original point cloud data and the first label data corresponding to the original point cloud data, the downsampled point cloud data, and the second label data corresponding to the downsampled point cloud data as the training data set.

[0010] In combination with the first aspect, in certain implementations of the first aspect, the first label data includes a bounding box label and a semantic segmentation label corresponding to the target object; adjusting the first label data to obtain second label data corresponding to the downsampled point cloud data includes: adjusting the semantic segmentation label in the first label data to obtain a semantic segmentation label that matches the downsampled point cloud data; using the semantic segmentation label that matches the downsampled point cloud data and the bounding box label in the first label data as the second label data corresponding to the downsampled point cloud data.

[0011] In combination with the first aspect, in certain implementations of the first aspect, the semantic segmentation label in the first label data is adjusted to obtain a semantic segmentation label that matches the downsampled point cloud data, including: obtaining a position index of a point in the downsampled point cloud data in the original point cloud data; using the position index to retrieve a semantic segmentation label corresponding to the point in the downsampled point cloud data from the semantic segmentation label in the first label data, retaining the retrieved semantic segmentation label, and deleting the semantic segmentation label of the point removed during the downsampling process, to obtain a semantic segmentation label that matches the downsampled point cloud data.

[0012] In combination with the first aspect, in certain implementations of the first aspect, the target object includes pedestrians and / or target obstacles in the mining area, and the target obstacle includes an obstacle whose volume is smaller than a preset threshold.

[0013] In a second aspect, an embodiment of the present application provides a point cloud processing method, including: acquiring point cloud data to be detected; using a target recognition model to analyze the point cloud data to be detected to obtain a target object contained in the point cloud data to be detected, wherein the target recognition model is obtained based on the training method described in the first aspect.

[0014] In combination with the second aspect, in certain implementations of the second aspect, a target recognition model is used to analyze the point cloud data to be detected to obtain the target object contained in the point cloud data to be detected, including: using the target recognition model to analyze the point cloud data to be detected to obtain a candidate bounding box that preliminarily defines the spatial position of the target object; determining the semantic segmentation results corresponding to each point cloud within the candidate bounding box; if the number of point clouds of the target object as a result of the semantic segmentation is greater than a preset number threshold, the candidate bounding box is a valid bounding box, and the valid bounding box and the semantic segmentation result of the point cloud within the valid bounding box are output as the target object; and / or, if the number of point clouds of the target object as a result of the semantic segmentation is less than or equal to the preset number threshold, the point cloud within the candidate bounding box is deleted from the point cloud data to be detected.

[0015] In combination with the second aspect, in certain implementations of the second aspect, a target recognition model is used to analyze the point cloud data to be detected to obtain the target object contained in the point cloud data to be detected, and also includes: if there are a target number of point clouds whose semantic segmentation results are target objects within the target range outside the valid bounding box, then based on the position data of the point clouds whose semantic segmentation results are target objects, the valid bounding box is re-determined, and the re-determined valid bounding box is output.

[0016] In a third aspect, an embodiment of the present application provides a training device for a target recognition model, comprising: a first construction module, used to construct a training data set, the training data set comprising point cloud data about a target object and label data corresponding to the point cloud data, the label data being used to annotate and identify the target object; a second construction module, used to construct an initial network model, the initial network model comprising a feature extraction module and multiple task execution modules, wherein an input end of each task execution module is connected to an output end of the feature extraction module, and at least some of the multiple task execution modules are independent of each other; a training module, used to train the initial network model based on the training data set until a predetermined performance indicator is reached to obtain a target recognition model.

[0017] In a fourth aspect, an embodiment of the present application provides a point cloud processing device, comprising: an acquisition module, used to acquire point cloud data to be detected; an analysis module, used to analyze the point cloud data to be detected using a target recognition model to obtain a target object contained in the point cloud data to be detected, wherein the target recognition model is obtained based on the training method described in the first aspect.

[0018] In a fifth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program for executing the method described in the first aspect and / or the second aspect.

[0019] In a sixth aspect, an embodiment of the present application provides an electronic device, comprising: a processor; a memory for storing processor executable instructions; the processor is used to execute the method described in the first aspect and / or the second aspect.

[0020] In this application, by collecting and annotating the point cloud data of the target object, the training data set can provide rich data that is consistent with the actual mining environment, which helps the target recognition model to better understand and identify the target objects in the mining area. Secondly, the design of the initial network model, especially the combination of the feature extraction module and multiple task execution modules, enables the trained target recognition model to extract key features from the point cloud data and independently process different tasks, thereby improving the flexibility and efficiency of the target recognition model. In addition, the task execution modules are independent of each other. The network model of this structure can process complex point cloud data, and through feature extraction and multi-task parallel processing, it can achieve effective recognition of target objects in the unmanned driving operation environment of the mine. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 The figure is a flow chart of a method for training a target recognition model provided in one embodiment of the present application.

[0023] Figure 2 Shown is a schematic diagram of the structure of an initial network model provided in an embodiment of the present application.

[0024] Figure 3 The figure is a schematic diagram of the process of training the initial network model provided by an embodiment of the present application.

[0025] Figure 4 The figure is a schematic diagram of a process of constructing a training data set provided in an embodiment of the present application.

[0026] Figure 5 The figure is a schematic diagram of a process of obtaining second tag data provided by an embodiment of the present application.

[0027] Figure 6The figure is a flow chart of a point cloud processing method provided in one embodiment of the present application.

[0028] Figure 7 The figure is a schematic diagram of a process of obtaining a target object contained in the point cloud data to be detected provided by an embodiment of the present application.

[0029] Figure 8 Shown is a schematic diagram of the structure of a training device for a target recognition model provided in one embodiment of the present application.

[0030] Fig. 9 Shown is a schematic structural diagram of a point cloud processing device provided in one embodiment of the present application.

[0031] Fig.10 Shown is a schematic structural diagram of an electronic device provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0033] Figure 1 FIG. 1 is a flow chart of a method for training a target recognition model provided by an embodiment of the present application. Figure 1 As shown, the method includes the following steps.

[0034] Step S110, constructing a training data set.

[0035] Specifically, the training data set includes point cloud data about the target object and label data corresponding to the point cloud data, and the label data is used to mark and identify the target object. It can be understood that these label data are the basis for training the initial network model and represent the training target of the model.

[0036] In some embodiments, the target object includes operating equipment, roads, ores, etc. In other embodiments, the target object includes pedestrians and / or target obstacles in a mining area, and the target obstacles include obstacles with a volume less than a preset threshold. Accordingly, the label data indicates whether the point cloud data belongs to a pedestrian or a target obstacle. In this way, when the initial network model processes these point cloud data with label data during the training process, it can improve the accuracy of identifying the target object by learning the relationship between the label data and the point cloud data.

[0037] The body of an autonomous vehicle is usually equipped with one or more laser radars. During the driving process of the vehicle, the laser radar of the body will scan the surrounding environment in real time and generate raw point cloud data. This data contains the spatial position, shape and other possible attribute information of the target object, such as reflection intensity. In order to construct an effective training data set, these raw point cloud data need to be preprocessed, including filtering and denoising, coordinate system correction, etc. Then, the preprocessed point cloud data is labeled. For example, each point cloud data or a specific area in the data is marked as a specific category. In addition, during the labeling process, the target objects will be classified more finely according to different application requirements, such as distinguishing them according to size, shape or behavioral characteristics.

[0038] Step S120, constructing an initial network model.

[0039] Figure 2 FIG. 1 is a schematic diagram of the structure of an initial network model provided by an embodiment of the present application. Figure 2 As shown, the initial network model includes a feature extraction module and multiple task execution modules, wherein the input end of each task execution module is connected to the output end of the feature extraction module, and at least some of the multiple task execution modules are independent of each other.

[0040] Specifically, the feature extraction module is the front-end part of the initial network model. It reduces the dimension of the input point cloud data through a series of computing layers (such as convolutional layers, pooling layers, etc.), and extracts features that are useful for subsequent tasks, including the shape, size, texture, edge and other information of the target object.

[0041] After the feature extraction module, there are also multiple task execution modules. Each task execution module is responsible for completing a specific task, such as classification, detection or segmentation, and the task execution module receives the data processed by the feature extraction module as input. In addition, each task execution module can perform tasks independently, that is, they can work in parallel to improve the efficiency of the target recognition model.

[0042] Step S130, based on the training data set, the initial network model is trained until a predetermined performance index is reached to obtain a target recognition model.

[0043] During the training process, the initial network model gradually adjusts its internal parameters and weights by learning the point cloud data and corresponding label data in the training dataset to improve the recognition accuracy of the target object. Specifically, at the beginning of the training, the parameters of the initial network model are randomly initialized. As the training progresses, the initial network model calculates the prediction results through forward propagation, and updates the parameters according to the difference between the prediction results and the real label data through the back propagation algorithm. In addition, during the training process, the performance of the initial network model is regularly evaluated on the validation dataset to monitor the learning progress of the model and make adjustments. Once the initial network model reaches the predetermined performance indicators on the validation dataset, such as accuracy, recall rate or F1 score, the training can be stopped. Ultimately, the resulting target recognition model can accurately identify the target object from the point cloud data, providing reliable target recognition capabilities for unmanned mine operation vehicles.

[0044] In this embodiment, by collecting and annotating the point cloud data of the target object, the training data set can provide rich data that is consistent with the actual mining environment, which helps the target recognition model to better understand and identify the target objects in the mining area. Secondly, the design of the initial network model, especially the combination of the feature extraction module and multiple task execution modules, enables the trained target recognition model to extract key features from the point cloud data and independently process different tasks, thereby improving the flexibility and efficiency of the target recognition model. In addition, the task execution modules are independent of each other. The network model of this structure can process complex point cloud data, and through feature extraction and multi-task parallel processing, it can achieve effective recognition of target objects in the unmanned operation environment of the mine.

[0045] In this embodiment, the target objects include pedestrians and / or target obstacles, wherein pedestrians involve driving safety issues, and target obstacles involve wear and tear issues such as sudden braking of vehicles. Therefore, the target recognition model can realize efficient and reliable detection of pedestrians and / or target obstacles, providing a comprehensive basis and guarantee for the safety and wear and tear issues of operating vehicles in the field of unmanned driving in mines, that is, improving the driving safety of related vehicles, and reducing problems such as sudden braking caused by misdetection of target obstacles and tire wear caused by missed detection.

[0046] Figure 3 FIG. 1 is a flow chart of training an initial network model provided by an embodiment of the present application. Figure 1 Based on the embodiment shown, Figure 3 The embodiment shown is described below in detail. Figure 3 The embodiment shown and Figure 1 The differences and similarities of the illustrated embodiments are not described in detail.

[0047] like Figure 3As shown, in this embodiment, the initial network model is trained based on the training data set, including the following steps.

[0048] Step S310, in the first stage, the full link of the initial network model is trained based on the training data set.

[0049] Specifically, the full link of the initial network model refers to the entire network structure from the input layer to the output layer. In the first stage, each module of the initial network model participates in learning, that is, the weights of all modules are updated. On the one hand, it is convenient for the initial network model to learn and process all tasks at the same time. On the other hand, it can train the feature extraction module to extract the features required by all task execution modules, laying the foundation for subsequent optimization training or optimization of specific task modules.

[0050] Step S320, in the second stage, lock the weight value of the feature extraction module and the weight values ​​of other task execution modules among the multiple task execution modules except the target task execution module; train the target task execution module based on the training data set.

[0051] In the second stage, the weights of the feature extraction module and other task execution modules except the target task execution module are locked, that is, their weights are no longer updated, ensuring that these modules remain unchanged during the training process and focus on optimizing the performance of the target task execution module. This strategy allows the initial network model to be further refined on the existing common features, focusing on extracting features that are more useful for individual task execution modules.

[0052] Step S330 , repeating the second stage training process for other task execution modules in a specific order until all task execution modules have been trained separately as target task modules.

[0053] In this step, the second-stage training process is repeated in a specific order for other task execution modules except the target task execution module that has been trained separately, to ensure that each task execution module in the initial network model can be optimized and improved, thereby improving the performance of the entire initial network model on all tasks.

[0054] For example, assume that the initial network model needs to perform three tasks simultaneously: pedestrian detection, obstacle classification, and road segmentation. In the second stage, the task execution module for pedestrian detection has been trained separately. Next, it is required to perform similar separate training on the task execution modules corresponding to the two tasks in a specific order, such as obstacle classification first and road segmentation later.

[0055] Through this training process, each task execution module can be optimized specifically, so that the object recognition model can achieve a high performance level when processing all tasks. Moreover, this phased and orderly training method helps to balance and improve the performance of the object recognition model on multiple tasks, and finally forms a comprehensive and efficient multi-task learning model.

[0056] Combination Figure 3 In the illustrated embodiment, in some embodiments of the present application, the feature extraction module includes a shared module, the multiple task execution modules include a first task module and a second task module, the first task module and the second task module are two interrelated modules, and the shared module is at least used to extract feature data common to the first task module and the second task module.

[0057] In this embodiment, through the shared module, the initial network model can learn common features that contribute to both task execution modules, which helps to reduce the redundancy of the initial network model and improve the utilization of weights. For example, if the first task module is responsible for pedestrian detection and the second task module is responsible for obstacle classification, the shared module will extract shared features that are important to both, such as shape, size, and texture. In this way, each task execution module can focus on extracting information that is more specific to their respective tasks from the common features.

[0058] In addition, since the first task module and the second task module are interrelated, the initial network model can further improve performance through this correlation. For example, if obstacle detection can provide contextual information for pedestrian detection, or pedestrian location can assist obstacle classification, the initial network model can use this correlation to improve the accuracy of training the two task execution modules.

[0059] In some embodiments of the present application, the first task module is at least used to perform a bounding box detection task, and the second task module is at least used to perform a semantic segmentation task. In the process of training the full link of the initial network model, the first task module and the second task module jointly participate in the training to achieve mutual learning of the label data corresponding to the first task module and the second task module.

[0060] Specifically, the bounding box detection task and the semantic segmentation task generate corresponding prediction results, and these prediction results are optimized based on the label data corresponding to the specific task. In this way, the features and patterns learned by the initial network model when performing one task can have a positive impact on the performance of another task. For example, the bounding box detection task identifies the locations of pedestrians and target obstacles, and this information can help the semantic segmentation task to more accurately divide the drivable area and pedestrian area.

[0061] In this embodiment, the knowledge learned by the initial network model on one task can be used to improve the performance of other tasks, thereby achieving positive transfer between tasks. Ultimately, this mutual learning process helps to improve the overall performance of the object recognition model on all tasks, making the object recognition model more powerful and flexible in practical applications.

[0062] Figure 4 FIG. 1 is a flow chart of constructing a training data set according to an embodiment of the present application. Figure 1 Based on the embodiment shown, Figure 4 The embodiment shown is described below in detail. Figure 4 The embodiment shown and Figure 1 The differences and similarities of the illustrated embodiments are not described in detail.

[0063] like Figure 4 As shown, in this embodiment, constructing a training data set includes the following steps.

[0064] Step S410: acquiring original point cloud data collected for the target object and first label data corresponding to the original point cloud data.

[0065] Step S420, down-sample the original point cloud data according to a preset ratio to obtain down-sampled point cloud data.

[0066] Specifically, downsampling can simplify the computational burden of the target recognition model while retaining the key features and structural information of the original point cloud data. Exemplarily, a downsampling ratio (i.e., a preset ratio) is preset, which determines how many points in the original point cloud data will be retained. For example, if the preset ratio is 10%, only 10 points out of every 100 points in the original point cloud data will be retained. Next, a suitable downsampling method is selected to downsample the original point cloud data. Exemplarily, the downsampling methods include random sampling, uniform grid sampling, farthest point sampling, etc. Among them, random sampling randomly selects points in the original point cloud data for retention; uniform grid sampling divides the space corresponding to the original point cloud data into grids and ensures that only one point is retained in each grid; farthest point sampling iteratively selects the point in the original point cloud data that is farthest from the selected point set. Further, points are selected from the original point cloud data according to the selected downsampling method and preset ratio. After the downsampling operation is performed, the obtained point cloud data set is the downsampled point cloud data.

[0067] Step S430: adjust the first label data to obtain second label data corresponding to the downsampled point cloud data.

[0068] It is understandable that in the downsampling process, many points in the original point cloud data are removed, and the label data corresponding to these removed points cannot be correctly associated with the downsampled point cloud data if they are not adjusted accordingly.

[0069] In one example, adjusting the first label data includes merging the first label data. Specifically, for adjacent or close points in the downsampled point cloud data, if their label data are the same or similar, the labels of the points with the same or similar label data can be merged to reduce the amount of the first label data while retaining important classification information.

[0070] In another example, adjusting the first label data includes interpolation. Specifically, for some attribute labels, the second label data of the downsampled point cloud data can be estimated by an interpolation method. For example, if the downsampled point cloud data contains height information, the height label data of the downsampled point cloud data can be determined by interpolating the height information of surrounding points.

[0071] Step S440: Use the original point cloud data and the first label data corresponding to the original point cloud data, the downsampled point cloud data and the second label data corresponding to the downsampled point cloud data as training data sets.

[0072] Specifically, the original point cloud data and its corresponding first label data, the downsampled point cloud data and its corresponding second label data are merged. In this way, the training data set will contain two groups of data, one group is the unprocessed original data, and the other group is the downsampled data.

[0073] In this embodiment, the training data set includes the original point cloud data and its corresponding first label data, so that the original network model can learn the complete and detailed features of the target object. The first label data is adjusted to obtain the second label data corresponding to the downsampled point cloud data, ensuring the consistency of the second label data with the downsampled point cloud data. This adjustment helps the original network model understand that after the data is compressed or simplified, it maintains its recognition ability. In addition, the downsampling process simulates the sensor resolution limitation or data processing capacity limitation that may be encountered in actual applications, so that the original network model can maintain a high recognition accuracy rate even when facing sparse data.

[0074] Figure 5 The figure is a schematic diagram of a process of obtaining second tag data provided by an embodiment of the present application. Figure 4 Based on the embodiment shown, Figure 5 The embodiment shown is described below in detail. Figure 5 The embodiment shown and Figure 4 The differences and similarities of the illustrated embodiments are not described in detail.

[0075] like Figure 5As shown, in this embodiment, the first label data includes a bounding box label and a semantic segmentation label corresponding to the target object. Adjusting the first label data to obtain second label data corresponding to the downsampled point cloud data includes the following steps.

[0076] Step S510: adjust the semantic segmentation label in the first label data to obtain a semantic segmentation label that matches the downsampled point cloud data.

[0077] The bounding box label describes the position of the target object in space, usually represented by the minimum bounding rectangle of the target object. During the downsampling process, although the density of the point cloud will change, the overall position and shape of the target object will not change significantly, so the bounding box label does not need to be adjusted like the semantic segmentation label. The semantic segmentation label indicates the category corresponding to each point in the point cloud data. During the downsampling process, some points will be removed, which makes the semantic segmentation label in the first label data no longer applicable.

[0078] Specifically, if an area in the original point cloud data contains multiple points, and only one point is retained after downsampling, the semantic segmentation label of the retained point needs to comprehensively consider the label information of all points in the original area to ensure the accuracy and representativeness of the semantic segmentation label corresponding to the downsampled point cloud data. For example, in the original point cloud data, there is a cubic obstacle, which consists of 100 points, and each point is labeled as a "target obstacle" category. During the downsampling process, only one or several points in this cube are retained. In order to determine the semantic segmentation labels of the retained points, it is necessary to check the semantic segmentation labels of all points in the original cube and determine the semantic segmentation labels of the retained points based on this information. In this way, even if the number of points is reduced, it can be ensured that the semantic segmentation labels of the downsampled point cloud data can accurately reflect the semantic information of the scene.

[0079] In other embodiments, the semantic segmentation label in the first label data is adjusted to obtain a semantic segmentation label that matches the downsampled point cloud data, including: obtaining a position index of a point in the downsampled point cloud data in the original point cloud data; using the position index to retrieve a semantic segmentation label corresponding to the point in the downsampled point cloud data from the semantic segmentation label in the first label data, retaining the retrieved semantic segmentation label, and deleting the semantic segmentation label of the point removed during the downsampling process to obtain a semantic segmentation label that matches the downsampled point cloud data.

[0080] Exemplarily, the position index of the point recorded during the downsampling process is used to find which point in the original point cloud data corresponds to each point in the downsampled point cloud data. Since each point in the original point cloud data already has an associated semantic segmentation label, indicating the category of the object or scene part to which the point belongs, these position indexes can be used to retrieve the semantic segmentation label corresponding to the downsampled point cloud data. In this embodiment, for each point in the downsampled point cloud data, the semantic segmentation label retrieved from the original point cloud data is retained. In this way, each point in the downsampled point cloud data will have an accurate semantic segmentation label that reflects its category in the original scene. In addition, many points will be removed during the downsampling process. For these removed points, their semantic segmentation labels are no longer needed, so they are deleted from the first label data set to ensure the compactness and accuracy of the second label data.

[0081] Through this process, a second labeled data that completely matches the downsampled point cloud data can be created. This labeled data reduces the complexity of the data while retaining enough information for effective learning and analysis, and can be directly used to train the initial network model without worrying about the inconsistency between the second labeled data and the downsampled point cloud data due to downsampling.

[0082] Step S520: Use the semantic segmentation label that matches the down-sampled point cloud data and the bounding box label in the first label data as second label data corresponding to the down-sampled point cloud data.

[0083] As mentioned earlier, during the downsampling process, although the density of the point cloud will change, the overall position and shape of the target object usually will not change significantly, so the bounding box labels in the first labeled data can be retained without modification.

[0084] The adjusted semantic segmentation label and the retained bounding box label are combined to form the second label data. In this way, for each point in the downsampled point cloud data, there is not only its semantic information (semantic segmentation label) but also the location information of the target object (bounding box label).

[0085] In this embodiment, by downsampling the original point cloud data, the number of points is reduced, the storage requirements and computational complexity are reduced, and the efficiency of the original network model training and reasoning is improved. During the downsampling process, the semantic segmentation labels are adjusted to ensure that each retained point can accurately represent the semantic information of its original area, which helps the original network model learn more precise and representative features, thereby maintaining high accuracy when performing target object recognition. By retaining the bounding box labels, the original network model can capture the overall spatial structure and location information of the target object without losing the overall understanding of the target. By combining the original bounding box labels and the adjusted semantic segmentation labels, the original network model can be trained on data of different scales and densities, which helps to improve the adaptability and generalization ability of the original network model to various scenarios.

[0086] Figure 6 FIG. 1 is a flow chart of a point cloud processing method provided by an embodiment of the present application. Figure 6 As shown, the method includes the following steps.

[0087] Step S610, obtaining point cloud data to be detected.

[0088] Exemplarily, the point cloud data to be detected is point cloud data containing target objects collected by a laser radar on the vehicle body when the autonomous driving vehicle is operating in a mining area.

[0089] Step S620: Analyze the point cloud data to be detected using the target recognition model to obtain the target object contained in the point cloud data to be detected.

[0090] Specifically, the target recognition model is obtained based on the training method in the aforementioned embodiment. Exemplarily, the point cloud data to be detected is input into the target recognition model, and the feature extraction module of the target recognition model extracts key features from the input point cloud data, including the geometric position, shape, size, color, intensity, etc. of the points. Then, using the key features, multiple task execution modules classify each point in the point cloud data to be detected and identify different target objects. Then, the target recognition model generates a bounding box label for each target object and assigns a semantic segmentation label to each point in the point cloud data to be detected. Finally, information about the target object contained in the point cloud data to be detected is output, including the category, position, size, etc. of the target object.

[0091] In this embodiment, the structural design of the target recognition model enables it to extract key features from the point cloud data to be detected and independently process different tasks, thereby being able to quickly and accurately identify the target object in the point cloud data to be detected. In addition, the multiple task execution modules of the target recognition model are independent of each other, so it can process more complex point cloud data to be detected and realize the effective recognition of target objects in the unmanned driving operation environment of the mine.

[0092] Figure 7 The figure shows a flow chart of obtaining a target object contained in the point cloud data to be detected provided by an embodiment of the present application. Figure 6 Based on the embodiment shown, Figure 7 The embodiment shown is described below in detail. Figure 7 The embodiment shown and Figure 6 The differences and similarities of the illustrated embodiments are not described in detail.

[0093] like Figure 7 As shown, in this embodiment, the target recognition model is used to analyze the point cloud data to be detected to obtain the target object contained in the point cloud data to be detected, including the following steps.

[0094] Step S710: Analyze the point cloud data to be detected using the target recognition model to obtain a candidate bounding box that preliminarily defines the spatial position of the target object.

[0095] The object recognition model identifies regions that may represent the target object through the extracted key features and generates a candidate bounding box for each region. The candidate bounding box is usually a rectangular box that provides a preliminary positioning of the target object in three-dimensional space. It should be noted that these boxes may not be precise, but they provide a starting point for subsequent processing and can be further refined.

[0096] In some embodiments, the target recognition model also generates a target object candidate list, each entry of which includes a candidate bounding box and a confidence score associated with the target object. The confidence score indicates the target recognition model's prediction certainty of the true target object contained in the candidate bounding box.

[0097] Step S720, determining the semantic segmentation results corresponding to each point cloud within the candidate bounding box.

[0098] Furthermore, the target recognition model analyzes the point cloud data within the candidate bounding box point by point to determine a semantic segmentation result for each point, such as "pedestrian", "vehicle", "target obstacle", etc. Similarly, if the candidate bounding box contains non-target objects, the target recognition model also needs to assign a correct label to each point of the object.

[0099] Step S730: If the number of point clouds of the target object in the semantic segmentation result is greater than a preset number threshold, the candidate bounding box is a valid bounding box, and the semantic segmentation result of the valid bounding box and the point cloud in the valid bounding box is output as the target object.

[0100] Exemplarily, the preset quantity threshold is an empirical value used to determine whether the candidate bounding box contains enough point clouds of the target object, thereby determining the validity of the candidate bounding box.

[0101] Specifically, the number of point clouds of the target object in the candidate bounding box is compared with a preset number threshold. If the number of point clouds of the target object is greater than the preset number threshold, the candidate bounding box is considered valid. At this time, the semantic segmentation result of the candidate bounding box and the point cloud in the box is used as the final output. This information represents the target object recognized by the target recognition model.

[0102] For example, in an autonomous driving scenario, the target recognition model needs to identify pedestrians around it. Then, the target recognition model first analyzes the point cloud data to be detected and generates several candidate bounding boxes containing pedestrians. Then, the target recognition model performs semantic segmentation on the point clouds within these candidate bounding boxes to identify the point clouds belonging to pedestrians. If the number of point clouds of identified pedestrians in a candidate bounding box exceeds the preset number threshold (for example, the preset number threshold is 50), the candidate bounding box is considered valid. Finally, the target recognition model outputs this valid bounding box and all point clouds identified as pedestrians in the box as detected pedestrian targets.

[0103] Step S740: If the semantic segmentation result shows that the number of point clouds of the target object is less than or equal to a preset number threshold, the point clouds within the candidate bounding box are deleted from the point cloud data to be detected.

[0104] Accordingly, if the number of point clouds of the target object in the candidate bounding box is less than or equal to the preset number threshold, it means that the target object in the candidate bounding box may be incomplete, unclear or insufficient to constitute a valid target object. In this case, these point clouds are deleted from the point cloud data to be detected. After deleting the point clouds that are less than the preset number threshold, the point cloud data to be detected is further processed or used for subsequent analysis and model training.

[0105] In this embodiment, the deletion operation helps to clean up the point cloud data to be detected, remove information that may cause misjudgment or is unimportant, reduce the computational burden of subsequent processing steps, and help improve the accuracy and efficiency of the target recognition model when processing the remaining point cloud data. For example, when the target recognition model processes the point cloud data to be detected, it generates a candidate bounding box that contains a series of point clouds. After semantic segmentation, only 30 point clouds are identified as "pedestrians." Then it will be required to delete these point clouds. In this way, only those candidate bounding boxes that contain enough point clouds of target objects will be retained as valid detection results. In this way, it can be ensured that the output of the target recognition model is more accurate and the possibility of false positives and negatives is reduced.

[0106] In this embodiment, the target recognition model analyzes the point cloud data to be detected, preliminarily defines the spatial position of the target object, and generates a candidate bounding box, so that potential target objects can be quickly identified. The semantic segmentation results clarify the categories of each point cloud in the candidate bounding box, which not only enhances the target recognition model's ability to understand the scene, but also provides support for the precise positioning of the target object. By judging the preset quantity threshold, it is ensured that only the candidate bounding boxes containing enough point clouds of target objects are considered valid, thereby improving the accuracy and reliability of target object detection.

[0107] Combination Figure 7 As shown, in other embodiments of the present application, a target recognition model is used to analyze the point cloud data to be detected to obtain the target object contained in the point cloud data to be detected, and also includes: if there are a target number of point clouds whose semantic segmentation results are the target objects within the target range outside the valid bounding box, then based on the position data of the point clouds whose semantic segmentation results are the target objects, the valid bounding box is re-determined, and the re-determined valid bounding box is output.

[0108] It is understandable that there may be point clouds of other target objects outside the effective bounding box. In order to ensure that all target objects are accurately detected and included in the bounding box, this embodiment proposes an adjustment mechanism. Specifically, if the semantic segmentation results show that there are point clouds of target objects of the target number within the target range outside the effective bounding box, the target recognition model will re-determine the bounding box of the target object based on these additional semantic segmentation results. Exemplarily, the target range is 0.5 meters and the number of targets is 5.

[0109] In some embodiments, the process of redetermining the effective bounding box involves expanding or adjusting the existing effective bounding box to ensure that all relevant point cloud data are included to more accurately reflect the actual spatial distribution of the target object. Ultimately, the target recognition model outputs these redetermined effective bounding boxes and the semantic segmentation results of the point cloud within the effective bounding box as the final output of the target recognition model.

[0110] In this embodiment, not only can the target objects in the point cloud data to be detected be identified, but further exploration can be performed outside the initially determined effective bounding box to find target objects that may be missed, thereby ensuring that the target objects are accurately detected. When a sufficient number of target object point clouds are found outside the effective bounding box, the target recognition model will readjust the effective bounding box based on the semantic segmentation results of these point clouds to ensure that all target objects are included, which not only improves the coverage of target object detection, but also enhances the target recognition model's ability to understand the scene, making the output results more in line with the actual situation. In addition, by redefining the effective bounding box, the target recognition model can provide more accurate target position and shape information so that the autonomous driving system can make safer driving decisions.

[0111] Combination of the above Figures 1 to 7 , describes the method embodiment of the present application in detail, and the following is combined with Figure 8 and Fig. 9 , describes the device embodiment of the present application in detail. It should be understood that the description of the method embodiment corresponds to the description of the device embodiment, so the parts not described in detail can refer to the previous method embodiment.

[0112] Figure 8 FIG. 1 is a schematic diagram of the structure of a training device for a target recognition model provided by an embodiment of the present application. Figure 8 As shown, the target recognition model training device 80 provided in the embodiment of the present application includes:

[0113] A first construction module 810 is used to construct a training data set, where the training data set includes point cloud data about a target object and label data corresponding to the point cloud data, where the label data is used to mark and identify the target object;

[0114] A second construction module 820 is used to construct an initial network model, wherein the initial network model includes a feature extraction module and a plurality of task execution modules, wherein an input end of each task execution module is connected to an output end of the feature extraction module, and at least some of the plurality of task execution modules are independent of each other;

[0115] The training module 830 is used to train the initial network model based on the training data set until the predetermined performance index is reached to obtain the target recognition model.

[0116] In one embodiment of the present application, the training module 830 is also used to, in the first stage, train the full link of the initial network model based on the training data set; in the second stage, lock the weight value of the feature extraction module and the weight values ​​of other task execution modules in multiple task execution modules except the target task execution module; train the target task execution module based on the training data set; repeat the second stage training process for other task execution modules in a specific order until all task execution modules have been trained separately as target task modules.

[0117] In one embodiment of the present application, the feature extraction module includes a shared module, and the multiple task execution modules include a first task module and a second task module. The first task module and the second task module are two interrelated modules, and the shared module is at least used to extract feature data common to the first task module and the second task module.

[0118] In one embodiment of the present application, the first task module is at least used to perform a bounding box detection task, and the second task module is at least used to perform a semantic segmentation task. In the process of training the full link of the initial network model, the first task module and the second task module jointly participate in the training to achieve mutual learning of the label data corresponding to the first task module and the second task module.

[0119] In one embodiment of the present application, the first construction module 810 is also used to obtain the original point cloud data collected for the target object and the first label data corresponding to the original point cloud data; downsample the original point cloud data according to a preset ratio to obtain downsampled point cloud data; adjust the first label data to obtain the second label data corresponding to the downsampled point cloud data; and use the original point cloud data and the first label data corresponding to the original point cloud data, the downsampled point cloud data and the second label data corresponding to the downsampled point cloud data as training data sets.

[0120] In one embodiment of the present application, the first construction module 810 is also used to adjust the semantic segmentation label in the first label data to obtain a semantic segmentation label that matches the downsampled point cloud data; and use the semantic segmentation label that matches the downsampled point cloud data and the bounding box label in the first label data as the second label data corresponding to the downsampled point cloud data.

[0121] In one embodiment of the present application, the first construction module 810 is also used to obtain the position index of the points in the downsampled point cloud data in the original point cloud data; use the position index to retrieve the semantic segmentation label corresponding to the point in the downsampled point cloud data from the semantic segmentation label in the first label data, retain the retrieved semantic segmentation label, and delete the semantic segmentation label of the point removed in the downsampling process, so as to obtain the semantic segmentation label that matches the downsampled point cloud data.

[0122] In one embodiment of the present application, the target object includes pedestrians and / or target obstacles in the mining area, and the target obstacle includes an obstacle whose volume is smaller than a preset threshold.

[0123] Fig. 9 FIG. 1 is a schematic diagram of the structure of a point cloud processing device provided by an embodiment of the present application. Fig. 9 As shown, the point cloud processing device 90 provided in the embodiment of the present application includes:

[0124] An acquisition module 910 is used to acquire point cloud data to be detected;

[0125] The analysis module 920 is used to analyze the point cloud data to be detected by using the target recognition model to obtain the target object contained in the point cloud data to be detected, wherein the target recognition model is obtained based on the training method in the above-mentioned embodiment.

[0126] In one embodiment of the present application, the analysis module 920 is also used to use the target recognition model to analyze the point cloud data to be detected to obtain a candidate bounding box that preliminarily defines the spatial position of the target object; determine the semantic segmentation results corresponding to each point cloud within the candidate bounding box; if the semantic segmentation result is that the number of point clouds of the target object is greater than a preset number threshold, the candidate bounding box is a valid bounding box, and the valid bounding box and the semantic segmentation result of the point cloud within the valid bounding box are output as the target object; and / or, if the semantic segmentation result is that the number of point clouds of the target object is less than or equal to the preset number threshold, the point cloud within the candidate bounding box is deleted from the point cloud data to be detected.

[0127] In one embodiment of the present application, the analysis module 920 is also used to, if there are a target number of point clouds whose semantic segmentation results are target objects within the target range outside the effective bounding box, redetermine the effective bounding box based on the position data of the point clouds whose semantic segmentation results are target objects, and output the redetermined effective bounding box.

[0128] Below, reference Fig.10 To describe an electronic device according to an embodiment of the present application. Fig.10 Shown is a schematic structural diagram of an electronic device provided by an exemplary embodiment of the present application.

[0129] like Fig.10 As shown, the electronic device 100 includes one or more processors 1001 and a memory 1002 .

[0130] The processor 1001 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 100 to perform desired functions.

[0131] The memory 1002 may include one or more computer program products, and the computer program product may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 1001 may run the program instructions to implement the methods of the various embodiments of the present application described above and / or other desired functions. Various contents such as training data sets, target recognition models, and point cloud data to be detected may also be stored in the computer-readable storage medium.

[0132] In one example, the electronic device 100 may further include: an input device 1003 and an output device 1004 , and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0133] The input device 1003 may include, for example, a keyboard, a mouse, etc.

[0134] The output device 1004 can output various information to the outside, including training data sets, target recognition models, point cloud data to be detected, etc. The output device 1004 can include, for example, a display, a speaker, a printer, a communication network and its connected remote output device, etc.

[0135] Of course, to simplify, Fig.10 Only some of the components related to the present application in the electronic device 100 are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, according to specific application situations, the electronic device 100 may also include any other appropriate components.

[0136] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the method according to various embodiments of the present application described above in this specification.

[0137] The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0138] In addition, an embodiment of the present application may also be a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, enables the processor to execute the steps of the method according to various embodiments of the present application described above in this specification.

[0139] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0140] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.

[0141] The block diagrams of the devices, apparatuses, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.

[0142] It should also be noted that in the apparatus, device and method of the present application, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0143] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0144] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for training a target recognition model, characterized in that: include: Constructing a training data set, wherein the training data set includes point cloud data about a target object and label data corresponding to the point cloud data, wherein the label data is used to mark and identify the target object; Constructing an initial network model, the initial network model comprising a feature extraction module and a plurality of task execution modules, wherein an input end of each of the task execution modules is connected to an output end of the feature extraction module, and at least some of the plurality of task execution modules are independent of each other; Based on the training data set, the initial network model is trained until a predetermined performance indicator is reached to obtain a target recognition model.

2. The target recognition model training method according to claim 1, characterized in that: The step of training the initial network model based on the training data set includes: In the first stage, based on the training data set, the full link of the initial network model is trained; In the second stage, the weight value of the feature extraction module and the weight values ​​of other task execution modules among the multiple task execution modules except the target task execution module are locked; based on the training data set, the target task execution module is trained; The training process of the second stage is repeated for the other task execution modules in a specific order until all the task execution modules have been trained separately as target task modules.

3. The target recognition model training method according to claim 2, characterized in that: The feature extraction module includes a shared module, the multiple task execution modules include a first task module and a second task module, the first task module and the second task module are two interrelated modules, and the shared module is at least used to extract feature data common to the first task module and the second task module.

4. The target recognition model training method according to claim 3, characterized in that: The first task module is at least used to perform a bounding box detection task, and the second task module is at least used to perform a semantic segmentation task. In the process of training the full link of the initial network model, the first task module and the second task module jointly participate in the training to achieve mutual learning of the label data corresponding to the first task module and the second task module.

5. The method for training a target recognition model according to any one of claims 1 to 4, characterized in that: The constructing of the training data set comprises: Acquire original point cloud data collected for the target object and first label data corresponding to the original point cloud data; Down-sampling the original point cloud data according to a preset ratio to obtain down-sampled point cloud data; Adjusting the first label data to obtain second label data corresponding to the downsampled point cloud data; The original point cloud data and the first label data corresponding to the original point cloud data, the downsampled point cloud data and the second label data corresponding to the downsampled point cloud data are used as the training data set.

6. The method for training a target recognition model according to claim 5, characterized in that: The first label data includes a bounding box label and a semantic segmentation label corresponding to the target object; and adjusting the first label data to obtain second label data corresponding to the downsampled point cloud data includes: Adjusting the semantic segmentation label in the first label data to obtain a semantic segmentation label that matches the downsampled point cloud data; The semantic segmentation label matched with the down-sampled point cloud data and the bounding box label in the first label data are used as second label data corresponding to the down-sampled point cloud data.

7. The method for training a target recognition model according to claim 6, characterized in that: The step of adjusting the semantic segmentation label in the first label data to obtain a semantic segmentation label matching the downsampled point cloud data includes: Obtaining a position index of a point in the downsampled point cloud data in the original point cloud data; The semantic segmentation label corresponding to the point in the downsampled point cloud data is retrieved from the semantic segmentation label in the first label data using the position index, the retrieved semantic segmentation label is retained, and the semantic segmentation label of the point removed in the downsampling process is deleted to obtain a semantic segmentation label matching the downsampled point cloud data.

8. The method for training a target recognition model according to any one of claims 1 to 7, characterized in that: The target objects include pedestrians and / or target obstacles in the mining area, and the target obstacles include obstacles whose volume is smaller than a preset threshold.

9. A point cloud processing method, characterized in that: include: Obtain the point cloud data to be detected; The target recognition model is used to analyze the point cloud data to be detected to obtain the target object contained in the point cloud data to be detected, wherein the target recognition model is obtained based on the training method described in any one of claims 1 to 8.

10. The point cloud processing method according to claim 9, characterized in that: The step of analyzing the point cloud data to be detected by using a target recognition model to obtain a target object contained in the point cloud data to be detected includes: Analyzing the point cloud data to be detected using the target recognition model to obtain a candidate bounding box that preliminarily defines the spatial position of the target object; Determine the semantic segmentation result corresponding to each point cloud within the candidate bounding box; If the number of point clouds of the target object as a result of the semantic segmentation is greater than a preset number threshold, the candidate bounding box is a valid bounding box, and the semantic segmentation result of the valid bounding box and the point cloud within the valid bounding box is output as the target object; and / or, If the semantic segmentation result is that the number of point clouds of the target object is less than or equal to the preset number threshold, the point clouds within the candidate bounding box are deleted from the point cloud data to be detected.

11. The point cloud processing method according to claim 10, characterized in that: The step of analyzing the point cloud data to be detected by using a target recognition model to obtain a target object contained in the point cloud data to be detected further includes: If there are a target number of point clouds whose semantic segmentation results are the target object within the target range outside the valid bounding box, the valid bounding box is re-determined based on the position data of the point clouds whose semantic segmentation results are the target object, and the re-determined valid bounding box is output.

12. A training device for a target recognition model, characterized in that: include: A first construction module is used to construct a training data set, wherein the training data set includes point cloud data about a target object and label data corresponding to the point cloud data, wherein the label data is used to mark and identify the target object; A second construction module is used to construct an initial network model, wherein the initial network model includes a feature extraction module and a plurality of task execution modules, wherein an input end of each of the task execution modules is connected to an output end of the feature extraction module, and at least some of the plurality of task execution modules are independent of each other; The training module is used to train the initial network model based on the training data set until a predetermined performance indicator is reached to obtain the target recognition model.

13. A point cloud processing device, characterized in that: include: An acquisition module is used to acquire point cloud data to be detected; An analysis module is used to analyze the point cloud data to be detected using a target recognition model to obtain a target object contained in the point cloud data to be detected, wherein the target recognition model is obtained based on the training method described in any one of claims 1 to 8.

Images