Point cloud data deduplication method and storage medium

Abstract

The application relates to a point cloud data deduplication method and a storage medium. The method comprises the following steps: acquiring point cloud data and picture data in the same time period; obtaining a first time point cloud image and a second time point cloud image according to a first rule; extracting feature vectors of the first time point cloud image and the second time point cloud image according to a second rule; comparing the feature vectors of the first time point cloud image and the second time point cloud image by using a preset means; and saving the second time point cloud image and a second time point cloud image corresponding to the second time point cloud image if the similarity of the feature vectors of the first time point cloud image and the second time point cloud image does not satisfy a preset condition. The point cloud data deduplication method and the storage medium can efficiently and quickly obtain deduplicated data, and improve deduplication efficiency and accuracy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method for deduplicating point cloud data and a storage medium. Background Technology

[0002] In the fields of artificial intelligence and autonomous driving, vehicle-to-infrastructure (V2I) technology is becoming increasingly mature. When vehicle-side or roadside monitoring is involved, the data collection time intervals for identifying road conditions and vehicles on the road are very short to ensure accuracy and precision. While this avoids omissions and improves recognition accuracy, the large amount of data increases the difficulty of analysis and computation. To reduce this difficulty, we usually add a deduplication step to remove highly similar data from similar time periods. However, the deduplication method and accuracy have a significant impact on the data; therefore, a highly efficient and accurate deduplication method for point cloud data is urgently needed. Summary of the Invention

[0003] To address or partially address the problems existing in related technologies, this application provides a point cloud data deduplication method and storage medium. This point cloud data deduplication method and storage medium can efficiently and quickly obtain deduplicated data, improving deduplication efficiency and accuracy.

[0004] The first aspect of this application provides a method for deduplicating point cloud data, including:

[0005] Acquire point cloud data and image data that are within the same time period;

[0006] The point cloud image of the first time point and the point cloud image of the next adjacent time point are obtained according to the first rule;

[0007] According to the second rule, the image data within the time period is framed to obtain at least one first framed image corresponding to the point cloud image at the first time point and at least one second framed image corresponding to the point cloud image at the second time point.

[0008] Extract the feature vectors of the first frame-picking image and the second frame-picking image respectively;

[0009] The feature vector of the second frame-drop image is compared with the feature vector of the first frame-drop image using a preset method. If the similarity between the feature vector of the second frame-drop image and the feature vector of the first frame-drop image does not meet the preset condition, the second frame-drop image and the second time-point cloud image corresponding to the second time-point cloud image are saved, and the similarity between the second time-point cloud image and its adjacent next time-point cloud image is judged based on the second time-point cloud image.

[0010] In one implementation, comparing the feature vector of the second frame-drop image with the feature vector of the first frame-drop image using a preset method further includes:

[0011] If the similarity between the feature vector of at least one second-time frame image and the feature vector of the first-time frame image meets a preset condition, then the second-time frame image corresponding to the second-time point cloud image and the second-time point cloud image are deleted, and the first-time point cloud image is used as a benchmark to determine the similarity between the next-time point cloud image adjacent to the second-time point cloud image, or...

[0012] If the similarity between the feature vector of at least one second frame image and the feature vector of the first frame image meets a preset condition, then the first frame image corresponding to the first time point cloud image and the first time point cloud image are deleted, and the second time point cloud image is used as a benchmark to determine the similarity between the next time point cloud image adjacent to the second time point cloud image.

[0013] In one implementation, acquiring point cloud data and image data that fall within the same time period includes:

[0014] Acquire point cloud data and four sets of image data within the same time period;

[0015] as well as,

[0016] According to the second rule, frames are extracted from the four sets of image data in the time period to obtain four first frame images corresponding to the point cloud image at the first time point and four second frame images corresponding to the point cloud image at the second time point.

[0017] Feature vectors are extracted from the four first frame images and the four second frame images respectively;

[0018] The feature vectors of the four second frame images are compared with the feature vectors of the four corresponding first frame images using a preset method. If the similarity between the feature vectors of all the second frame images and the feature vectors of the corresponding first frame images does not meet the preset condition, the four second frame images corresponding to the second time point cloud image and the second time point cloud image are saved, and the second time point cloud image is used as a benchmark to determine the similarity with the next adjacent time point cloud image.

[0019] In one implementation, the step of comparing the feature vectors of the four second frame images with the feature vectors of the corresponding four first frame images using a preset method includes:

[0020] Get the user-set preset threshold;

[0021] The feature vectors of the four second frame images are compared with the feature vectors of the corresponding four first frame images.

[0022] If the similarity between the feature vectors of the four second frame images and the feature vectors of the corresponding first frame images is less than the preset threshold, then the four second frame images corresponding to the second time point cloud image and the second time point cloud image are saved, and the second time point cloud image is used as a benchmark to determine the similarity between it and the next adjacent time point cloud image.

[0023] In one implementation, if the similarity between the feature vector of at least one of the four second frame images and the feature vector of the corresponding first frame image is greater than or equal to the preset threshold, then the four second frame images corresponding to the second time point cloud image and the second time point cloud image are deleted, and the first time point cloud image is used as a benchmark to determine the similarity between the next time point cloud image adjacent to the second time point cloud image.

[0024] In one implementation, if the similarity between the feature vectors of the four second frame images and the feature vectors of the corresponding first frame images is less than the preset threshold, then the four second frame images corresponding to the second time point cloud image and the second time point cloud image are saved, and the second time point cloud image is used as a benchmark to determine the similarity between it and the adjacent next time point cloud image, including:

[0025] Using the law of cosines, the feature vectors of the four second frame images are compared with the feature vectors of the corresponding first frame images. If the similarity between the feature vector of at least one second frame image and the feature vector of the corresponding first frame image is greater than or equal to the preset threshold, then the four second frame images corresponding to the second time point cloud image and the second time point cloud image are deleted.

[0026] In one implementation, extracting feature vectors from the four first frame images and the four second frame images respectively includes:

[0027] The first and second frame-by-frame images are processed using a semantic segmentation model based on a convolutional neural network to obtain the feature vectors of the first and second frame-by-frame images respectively.

[0028] In one implementation, the step of extracting frames from the image data within the time period according to the second rule includes:

[0029] Acquire at least one image data at the same unit of time as the point cloud data, and generate at least one set of consecutive multi-frame extracted images;

[0030] Select at least one frame from a set such that the selected frame corresponds to the point cloud image at the first time point.

[0031] In one implementation, acquiring point cloud data and image data that fall within the same time period includes:

[0032] The number of lidars used to acquire the point cloud data includes at least one, and the point cloud data obtained from different lidars is fused to form the point cloud data.

[0033] The number of cameras used to acquire image data includes at least four in different directions, and the image data obtained from different cameras is used as the image data.

[0034] A second aspect of this application provides a storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method described above.

[0035] The technical solution provided in this application may include the following beneficial effects:

[0036] The point cloud data deduplication method and storage medium provided in this application extract feature vectors from a first and a corresponding second frame image and compare them, thereby transforming the deduplication of 3D point cloud data into 2D deduplication and saving it to a point cloud image that meets preset conditions, thus simplifying the data. This method, which transforms 3D point cloud data deduplication into 2D data deduplication, reduces the algorithmic difficulty during the deduplication process and improves the efficiency and accuracy of data deduplication. Furthermore, by using this one-to-many comparison method between 3D data and at least one set of 2D data, and performing deduplication processing according to preset conditions, it helps improve the accuracy of similarity judgment between adjacent time points and avoids misjudgments of point cloud data similarity under external conditions such as unidirectional occlusion, which could affect the accuracy of deduplication.

[0037] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0038] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0039] Figure 1 This is a schematic diagram of the first process of the point cloud data deduplication method shown in the embodiments of this application;

[0040] Figure 2This is a schematic diagram of the second process of the point cloud data deduplication method shown in the embodiments of this application;

[0041] Figure 3 This is a schematic diagram of the third process of the point cloud data deduplication method shown in the embodiments of this application;

[0042] Figure 4 This is a schematic diagram of the fourth process of the point cloud data deduplication method shown in the embodiments of this application;

[0043] Figure 5 This is a simplified structural diagram of the electronic device in the point cloud data deduplication method shown in the embodiments of this application. Detailed Implementation

[0044] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0045] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0046] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0047] In related technologies, when using monitoring systems to achieve vehicle-to-infrastructure (V2I) integration, point cloud data is typically collected using LiDAR and image data is collected by multiple cameras. Then, planar fitting is performed based on the point cloud data to achieve the recognition purpose. Point cloud data refers to a set of vectors in a three-dimensional coordinate system. Scanned data is recorded in the form of points, each containing three-dimensional coordinates, and some may contain color information (RGB) or reflectance intensity information. For monitoring systems, multiple frames within a certain time period may have high similarity, resulting in a large data storage space, which is detrimental to data storage and later retrieval.

[0048] In some cases, to achieve better recognition results through the combined use of LiDAR and cameras, surveillance systems are pre-trained before deployment to establish corresponding training models. The accuracy of the training data directly affects the performance of the surveillance system. Understandably, for the surveillance systems mentioned above, the types of targets to be identified will vary depending on the usage requirements in different environments. The recognition objectives of the surveillance system can be achieved by replacing different training data. For example, in intelligent driving, surveillance systems can be set up at intersections to monitor the entry and exit of different vehicles, or they can be set up on autonomous vehicles for obstacle avoidance, road recognition, and road sign identification, etc., without specific limitations. During the model building process, the selected time period contains a large amount of data from the surveillance system, and many of these data points have high similarity, resulting in significant time consumption during training. Therefore, the acquired data needs to be deduplicated beforehand.

[0049] To address the aforementioned issues, this application provides a first flowchart illustrating a method for verifying training model data.

[0050] See Figure 1 The method includes:

[0051] S101. Obtain point cloud data and image data that are within the same time period.

[0052] S102. Obtain the point cloud image of the first time point within the time period and the point cloud image of the next adjacent time point according to the first rule.

[0053] S103. According to the second rule, extract frames from the image data within the time period to obtain at least one first extracted frame image corresponding to the point cloud image at the first time point and at least one second extracted frame image corresponding to the point cloud image at the second time point.

[0054] S104. Extract the feature vectors of the first and second frame images respectively.

[0055] S105. Using a preset method, compare the feature vector of the second frame image with the feature vector of the first frame image. If the similarity between the feature vector of the second frame image and the feature vector of the first frame image does not meet the preset condition, save the second frame image and the second time point cloud image corresponding to the second time point cloud image, and use the second time point cloud image as a benchmark to judge the similarity with the next adjacent time point cloud data.

[0056] By acquiring point cloud data and image data from the same time period, point cloud images at different times are obtained based on the point cloud data. At least one first frame image, at least one second frame image, and so on, are obtained based on the time corresponding to the obtained point cloud images. Feature vectors are extracted from the first frame image and its corresponding second frame image and compared, thus transforming the deduplication of the 3D point cloud data into 2D deduplication. This is then saved to a point cloud image that meets preset conditions, thereby simplifying the data. This method of transforming 3D point cloud data deduplication into 2D data deduplication reduces the algorithmic difficulty in the deduplication process and improves the efficiency and accuracy of data deduplication. Furthermore, by using this one-to-many comparison method between 3D data and at least one set of 2D data, and performing deduplication processing according to preset conditions, the accuracy of similarity judgment between adjacent time points is improved. This avoids misjudgment of point cloud data similarity under external conditions such as unidirectional occlusion, which could affect the accuracy of deduplication.

[0057] In this embodiment, point cloud images can be acquired using a LiDAR within a certain time period to obtain point cloud data, and images can be acquired using a camera used in conjunction with the LiDAR to obtain image data. This method is not limited to applications such as data storage and model training. The point cloud data and image data acquired within this time period are saved separately and used as the required point cloud data and image data.

[0058] Optionally, the LiDAR and cameras can be installed on vehicles or at road intersections. To obtain target location and category information through the combined use of LiDAR and cameras, at least one camera is required. To facilitate more accurate acquisition of image data within the field of view, multiple cameras can be used, distributed with different shooting directions. For example, three cameras can be arranged circumferentially so that the image data captured by the cameras can cover the field of view through different shooting angles, and the shooting angles between the three cameras can be set to be evenly distributed circumferentially; or, when four circumferentially arranged cameras are used, the shooting angles between the four cameras can be set to be evenly distributed circumferentially.

[0059] Regardless of the number and arrangement of cameras, as long as the image data ultimately collected by the cameras can cover the LiDAR's field of view on the ground plane, it will avoid missing image data in the LiDAR's field of view and affecting the recognition accuracy within that field of view.

[0060] Understandably, when deploying LiDAR and accompanying cameras on vehicles or intersections, the number of LiDARs can be at least one. If two or more LiDARs are deployed, point cloud data acquisition requires separately acquiring point cloud data from different LiDARs within the same time period, then fusing these different point cloud data into a single point cloud dataset. Similarly, the number of cameras can also be one or more. With multiple cameras, each camera will capture images from different directions. For example, with four cameras, the images will be captured from four different directions, and the image data from the four cameras within the same time period will be saved separately and then combined as a single image dataset.

[0061] The first point cloud image and the second point cloud image of the adjacent next time moment are obtained according to the first rule. In this step, the point cloud data is a collection of multiple point cloud data. Obtaining the first point cloud image and the second point cloud image of the adjacent next time moment within the time period according to the first rule can specifically include: acquiring point cloud data per unit time, where a time period includes multiple unit time periods. The point cloud data per unit time period is used to obtain multiple frames of point cloud images through semantic segmentation, frame extraction, etc. For the multiple frames of point cloud images, multiple frames of point cloud images within the unit time period are selected at intervals as the first time point image, the second time point image, ..., the Nth time point image. Point cloud data from different unit time periods can be processed in this way. It can be understood that for the first time moment and the second time moment adjacent to it, multiple point cloud images are obtained through interval selection and then named in the order of first and second. When selecting multiple frames of point cloud images at intervals, the selection can be done at different intervals such as 1 frame or 3 frames, etc., without specific limitations.

[0062] Alternatively, it could include: acquiring point cloud data per unit time, generating multiple consecutive frames of point cloud images by frame extraction, selecting one frame from the multiple frames using a pre-defined algorithm as the first time-stamp point cloud image, and selecting point cloud data from different unit times using the same method and naming them according to the time sequence of the different unit times as the first time-stamp point cloud image, the second time-stamp point cloud image, ..., the Nth time-stamp point cloud image. It is understood that the pre-defined algorithm involved can be a formula or constraint that needs to be set according to different application scenarios or requirements. For example, selecting the middle frame from multiple frames within a unit time as the first time-stamp point cloud image, or selecting the 5th frame from multiple frames within a unit time as the first time-stamp point cloud image, is not specifically limited here.

[0063] See Figure 2 According to the second rule, the image data within the time period is frame-sampling to obtain at least one first frame image corresponding to the point cloud image at the first time point and at least one second frame image corresponding to the point cloud image at the second time point, including:

[0064] S1031. Acquire at least one image data at the same unit of time as the point cloud data, and generate at least one set of consecutive multi-frame extracted images.

[0065] S1032. Select at least one frame from the set, such that the selected frame corresponds to the point cloud image at the first time step.

[0066] To facilitate understanding, we will take the point cloud image at the first moment mentioned above as an example. This first moment can be obtained through different methods described above. A detailed explanation will be provided of the implementation method for acquiring at least one frame image corresponding to the point cloud image at the first moment. For the image data acquired within the time period, the number of cameras set can include multiple sets of image data, such as A, B, C…N.

[0067] Taking three sets of image data, A, B, and C, as examples, we process each set separately. Taking set A as an example, we extract frames from the image data within a unit of time, obtaining consecutive multi-frame images. The number of consecutive multi-frame images within this unit of time can be the same as or different from the number of point cloud images obtained within the same unit of time and the time interval. When both are the same, the multi-frame images can be selected using the same method as selecting point cloud data at a specific moment from the same unit of time. When both are different, the selection method is adaptively adjusted based on the number of extracted frames and the number of point cloud images obtained within the same unit of time, as long as a one-to-one correspondence between the extracted frames and the point cloud data at a specific moment is maintained. Thus, using this method, we can obtain the first, second, ..., Nth frame images corresponding to the number of camera groups based on the point cloud data at different times.

[0068] The feature vectors of the first and second frame images are extracted respectively. It is understandable that the extraction of feature vectors from two-dimensional images can be achieved through coordinate system transformation, which will not be explained in detail here.

[0069] A preset method is used to compare the feature vectors of the second extracted frame image with those of the first extracted frame image. If the similarity between the feature vectors of the second extracted frame image and the first extracted frame image does not meet a preset condition, the second extracted frame image corresponding to the second time-stamp point cloud image and the second time-stamp point cloud image are saved. The second time-stamp point cloud image is then used as a benchmark to determine its similarity to the next adjacent time-stamp point cloud image. By using the extracted frame images converted from the corresponding image data, the similarity of the feature vectors of the point cloud images is compared using a preset method, and the preset condition is used as a judgment benchmark to achieve deduplication of similar data in the point cloud data. This method reduces the difficulty of deduplication and ensures the accuracy of point cloud image deduplication, avoiding misjudgments that could affect recognition accuracy.

[0070] In one embodiment, if the similarity between the feature vector of at least one second-time frame image and the feature vector of the first-time frame image meets a preset condition, then the second-time frame image corresponding to the second-time point cloud image and the second-time point cloud image are deleted. Furthermore, the first-time point cloud image is used as a benchmark to determine the similarity between the second-time point cloud image and the next-time point cloud image adjacent to it. Within a unit of time, the differences between multiple point cloud images may not be significant. To achieve higher accuracy in deduplication, the first-time point cloud image is always used as the benchmark to determine similarity, provided the comparison of the next-time point cloud image meets the preset condition. This similarity determination can identify subtle differences between adjacent frames within a unit of time, thereby completing data deduplication according to the set preset conditions.

[0071] In one embodiment, if the similarity between the feature vector of at least one second-time frame image and the feature vector of the first-time frame image meets a preset condition, then the first-time frame image corresponding to the first-time point cloud image and the first-time point cloud image are deleted, and the second-time point cloud image is used as a benchmark to determine the similarity between the second-time point cloud image and the next-time point cloud image adjacent to it. When selecting one frame per unit time as the first-time point cloud data, the second-time point cloud data, ... the Nth-time point cloud data, the features of the images in adjacent times may be progressively related, making the similarity of each image the same or close. If the similarity only fluctuates within the range of the preset condition, the accuracy of always using the first-time point cloud image as the benchmark to determine the similarity will decrease when the comparison of the next-time point cloud image meets the preset condition, which is not conducive to the accuracy of deduplication.

[0072] As one specific implementation of this application, see [link to specific implementation]. Figure 3 The following example, which sets one LiDAR and four cameras, will be used to illustrate this application in detail.

[0073] S101a: Obtain point cloud data and 4 sets of image data within the same time period.

[0074] S101b: According to the second rule, extract frames from the four sets of image data in the time period to obtain four first extracted frame images corresponding to the point cloud image at the first time point and four second extracted frame images corresponding to the point cloud image at the second time point.

[0075] S101c: Extract feature vectors from the four first-frame images and the four second-frame images respectively.

[0076] S101d: Using a preset method, compare the feature vectors of the four second frame images with the feature vectors of the four corresponding first frame images. If the similarity between the feature vectors of all the second frame images and the feature vectors of the corresponding first frame images does not meet the preset condition, then save the four second frame images corresponding to the second time point cloud image and the second time point cloud image, and use the second time point cloud image as a benchmark to judge the similarity with the next adjacent time point cloud image.

[0077] Point cloud data and four sets of image data were acquired within the same time period using a set of LiDAR and four sets of cameras facing different directions. The point cloud data and the four sets of image data were processed as described above to obtain the corresponding first-time-point image and four first-frame images from different directions, the corresponding second-time-point image and four second-frame images from different directions, and so on, until the Nth-time-point image and four Nth-frame images from different directions. This simplifies the deduplication of 3D data to 2D data, which will not be elaborated further here. Furthermore, based on the obtained frame images from different times, feature vectors were extracted and similarity was compared. Since one time-point image corresponds to four frame images from four directions, the similarity of frame images from adjacent times corresponding to the same direction was compared to improve the accuracy of the similarity judgment between adjacent time-point images. That is, only when the comparison results of frame images from all four directions do not meet the preset conditions can they be judged as having high similarity, thus preserving the original data and process data, which is beneficial to the accuracy of deduplication.

[0078] In one embodiment, see Figure 4 The feature vectors of the four second-frame images are compared with the feature vectors of the corresponding four first-frame images using a preset method, including:

[0079] S1011d: Obtain the user-set preset threshold.

[0080] S1012d: Compare the feature vectors of the four second-frame images with the feature vectors of the corresponding four first-frame images.

[0081] S1013d. If the similarity between the feature vectors of the four second-time frame images and the feature vectors of the corresponding first-time frame images is less than a preset threshold, then save the four second-time frame images and the second-time point cloud image corresponding to the second-time point cloud image, and use the second-time point cloud image as a benchmark to judge the similarity between it and the next adjacent time-next point cloud image.

[0082] The specific methods employed may include: setting a variable preset threshold in the deduplication method, allowing users to adjust it according to different situations. For example, the preset threshold could be 90%, 80%, or a range of 80%-95%, without specific limitations. During comparison, if the similarity between the feature vectors of the four second-time frame images and the feature vectors of the corresponding first-time frame images is less than the preset threshold, it indicates that the two point cloud images have low similarity and do not need to be deleted. In this case, the four second-time frame images corresponding to the second-time point cloud image and the second-time point cloud image are saved, and the second-time point cloud image is used as a benchmark to determine the similarity with its adjacent next-time point cloud image. If the similarity between the feature vector of at least one of the four second-time frame images and the feature vector of the corresponding first-time frame image is greater than or equal to the preset threshold, the four second-time frame images corresponding to the second-time point cloud image and the second-time point cloud image are deleted, and the first-time point cloud image is used as a benchmark to determine the similarity with its adjacent next-time point cloud image. Only when the comparison results of all four extracted frames are less than the preset threshold is it considered that the similarity between the corresponding two point cloud images is not high. If at least one comparison result is greater than or equal to the preset threshold, one of the corresponding data must be deleted, and the other corresponding point cloud image is used as the benchmark for comparison. This comparison method can ensure that the similarity of the saved data meets the requirements, avoid misidentification and omission that affect the comparison results, and improve the accuracy of deduplication.

[0083] Optionally, when comparing the corresponding frame images, the cosine theorem can be used to compare the feature vectors of the four second frame images with the feature vectors of the corresponding first frame images. If the similarity between the feature vector of at least one second frame image and the feature vector of the corresponding first frame image is greater than or equal to a preset threshold, then the four second frame images corresponding to the second time point cloud image and the second time point cloud image are deleted.

[0084] It is understandable that extracting feature vectors from the four first-frame images and four second-frame images can involve using a semantic segmentation model of a convolutional neural network to process the first-frame images and the second-frame images respectively, obtaining the feature vectors of the first-frame images and the corresponding feature vectors of the second-frame images. This processing method can improve the precision and accuracy of the obtained feature vectors, which is beneficial to the accuracy of subsequent image similarity comparisons.

[0085] This application also provides an electronic device, see [link to document]. Figure 5 The electronic device 1000 includes a memory 1010 and a processor 1020.

[0086] The processor 1020 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0087] Memory 1010 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by the processor 1020 or other modules of the computer. Permanent storage devices may be read-write storage devices. Permanent storage devices may be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use mass storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices may be removable storage devices (e.g., floppy disks, optical drives). System memory may be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory may store some or all of the instructions and data required by the processor during operation. Furthermore, memory 1010 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some implementations, memory 1010 may include a removable storage device that is readable and / or writable, such as a laser disc (CD), a read-only digital multifunction optical disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-high density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.

[0088] The memory 1010 stores executable code, which, when processed by the processor 1020, can cause the processor 1020 to execute part or all of the methods described above.

[0089] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.

[0090] Alternatively, this application may be implemented as a non-transitory machine-readable storage medium (or computer-readable storage medium, or machine-readable storage medium) storing executable code (or computer program, or computer instruction code) thereon, which, when executed by a processor of an electronic device (or electronic device, server, etc.), causes the processor to perform part or all of the steps of the above-described method according to this application.

[0091] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for point cloud data deduplication, the method comprising: The method comprises: acquiring point cloud data and multiple sets of picture data in the same time period; each set of picture data comprises picture data collected by cameras in multiple different directions; obtaining a first time point cloud image and a second time point cloud image adjacent to the first time point cloud image in the time period according to a first rule; extracting feature vectors of the first and second extracted pictures respectively; comparing the feature vectors of the second extracted pictures with the feature vectors of the first extracted pictures by using a preset means, if the similarity of all the feature vectors of the second extracted pictures with all the feature vectors of the first extracted pictures does not meet a preset condition, saving the second extracted pictures corresponding to the second time point cloud image and the second time point cloud image, and taking the second time point cloud image as a reference to judge the similarity of the next time point cloud image adjacent to the second time point cloud image. if the similarity of at least one feature vector of the second extracted picture with at least one feature vector of the first extracted picture meets the preset condition, deleting the second extracted pictures corresponding to the second time point cloud image and the second time point cloud image, and taking the first time point cloud image as a reference to judge the similarity of the next time point cloud image adjacent to the second time point cloud image, or, if the similarity of at least one feature vector of the second extracted picture with at least one feature vector of the first extracted picture meets the preset condition, deleting the first extracted pictures corresponding to the first time point cloud image and the first time point cloud image, and taking the second time point cloud image as a reference to judge the similarity of the next time point cloud image adjacent to the second time point cloud image. The acquisition of point cloud data and picture data in the same time period comprises:

2. The method of claim 1, wherein, acquiring point cloud data and four sets of picture data in the same time period; and extracting feature vectors from the four first and second extracted pictures respectively; comparing the feature vectors of the four second extracted pictures with the feature vectors of the corresponding four first extracted pictures by using a preset means, if the similarity of all the feature vectors of the second extracted pictures with the feature vectors of the corresponding first extracted pictures does not meet a preset condition, saving the four second extracted pictures corresponding to the second time point cloud image and the second time point cloud image, and taking the second time point cloud image as a reference to judge the similarity of the next time point cloud image adjacent to the second time point cloud image. ​ ​ 3. The method of claim 2, wherein, The adopting preset means respectively compares the feature vectors of the four second frame pictures with the feature vectors of the corresponding four first frame pictures, including: Obtaining a preset threshold value set by a user; Respectively comparing the feature vectors of the four second frame pictures with the feature vectors of the corresponding four first frame pictures; If the similarity degrees of the feature vectors of the four second frame pictures and the feature vectors of the corresponding first frame pictures are all less than the preset threshold value, saving the four second frame pictures corresponding to the second time point cloud image and the second time point cloud image, and taking the second time point cloud image as a reference to judge the similarity degree of the next time point cloud image adjacent thereto.

4. The method of claim 3, wherein, If the similarity degrees of the feature vectors of at least one of the four second frame pictures and the feature vectors of the corresponding first frame pictures are greater than or equal to the preset threshold value, deleting the four second frame pictures corresponding to the second time point cloud image and the second time point cloud image, and taking the first time point cloud image as a reference to judge the similarity degree of the next time point cloud image adjacent to the second time point cloud image.

5. The method of claim 4, wherein, If the similarity degrees of the feature vectors of the four second frame pictures and the feature vectors of the corresponding first frame pictures are all less than the preset threshold value, saving the four second frame pictures corresponding to the second time point cloud image and the second time point cloud image, and taking the second time point cloud image as a reference to judge the similarity degree of the next time point cloud image adjacent thereto, including: Respectively comparing the feature vectors of the four second frame pictures with the feature vectors of the corresponding first frame pictures using the cosine theorem, and if the similarity degrees of the feature vectors of at least one of the second frame pictures and the feature vectors of the corresponding first frame pictures are greater than or equal to the preset threshold value, deleting the four second frame pictures corresponding to the second time point cloud image and the second time point cloud image.

6. The method of claim 2, wherein, The feature vectors of the four first frame pictures and the four second frame pictures are extracted respectively, including: The first frame pictures and the second frame pictures are processed respectively using a semantic segmentation model of a convolutional neural network to obtain the feature vectors of the first frame pictures and the corresponding feature vectors of the second frame pictures.

7. The method according to any one of claims 1 to 6, characterized in that, The picture data in the time period is frame extracted according to a second rule, including: Obtaining at least one picture data in the same unit time as the point cloud data, and generating at least one set of continuous multiple frame extracted pictures; Selecting at least one frame extracted picture in a set, so that the selected frame extracted picture corresponds to the first time point cloud image.

8. The method according to any one of claims 1-6, characterized in that, The point cloud data and the picture data in the same time period are obtained, including: The number of laser radars for obtaining the point cloud data includes at least one, and the point cloud data obtained from different laser radars is fused to form the point cloud data; The number of cameras for obtaining picture data includes at least four different directions, and the picture data obtained from different cameras is taken as the picture data.

9. A storage medium, characterized by having executable code stored thereon that, when executed by a processor of an electronic device, causes the processor to perform the method of any of claims 1-8.

Images