WO2022213992A1 - 一种数据处理方法及装置 - Google Patents

一种数据处理方法及装置 Download PDF

Info

Publication number
WO2022213992A1
WO2022213992A1 PCT/CN2022/085349 CN2022085349W WO2022213992A1 WO 2022213992 A1 WO2022213992 A1 WO 2022213992A1 CN 2022085349 W CN2022085349 W CN 2022085349W WO 2022213992 A1 WO2022213992 A1 WO 2022213992A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
layer
node
information
occupancy
Prior art date
Application number
PCT/CN2022/085349
Other languages
English (en)
French (fr)
Inventor
涂晨曦
蔡康颖
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2022213992A1 publication Critical patent/WO2022213992A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present application relates to the field of data processing, and in particular, to a data processing method and device.
  • a collection of points obtained is called a point cloud.
  • point clouds can be used to display digital and virtual objects; for example, in the field of autonomous driving, point clouds can be used to simulate reference objects to achieve Precise positioning and navigation of vehicles.
  • the data volume of point cloud is very large. In order to reduce the storage space occupied by the point cloud, the point cloud will be compressed and then stored.
  • the Octree Squeeze algorithm based on octree performs context prediction on the point cloud after octreeization, and compresses the point cloud according to the result of the context prediction.
  • the context prediction process adopted by the Oct Squeeze algorithm is to use a multi-layer perceptron (MLP) to perform feature fusion step by step to obtain the context prediction result.
  • MLP multi-layer perceptron
  • Each level of feature fusion will use at least one layer of MLP, which will occupy a lot of computing resources of point cloud compression equipment, and point cloud compression will take a long time. Therefore, how to reduce the model complexity of point cloud compression and the required computing resources is an urgent problem to be solved.
  • the present application provides a data processing method and device, which solve the problems of high complexity of the point cloud compression model in the prior art and large computational resource occupation.
  • the present application provides a data processing method, and the method can be applied to a sending end, or the method can be applied to an apparatus that can support a computing device to implement the method, for example, the apparatus includes a chip system, and the method includes: sending The terminal generates the data to be compressed in a tree-like structure according to the original data, and uses the recurrent network layer included in the data compression model to determine the data occupancy information in the tree-like structure.
  • the above-mentioned data occupation information is used to indicate the data distribution of the original data in the tree structure.
  • the sending end compresses the above-mentioned data to be compressed according to the data occupation information to obtain compressed data.
  • the present application uses a recurrent network layer to replace the multi-layer MLP network in the prior art for context prediction, which reduces the complexity of the data compression model and reduces the computing resources required for context prediction to obtain data occupancy information;
  • the data processing method provided by the present application does not need to store the features and the intermediate features, thereby reducing the storage space occupied by data compression.
  • the data processing method may further include: collecting raw data through a sensor, where the raw data is three-dimensional data.
  • the sensor includes at least one of lidar, millimeter wave radar, and sonar.
  • the data processing method may further include: collecting raw data through a camera, where the raw data is two-dimensional data.
  • the data processing method further includes: displaying at least one of a tree structure and data possession information.
  • displaying at least one of the tree structure and data possession information it is helpful for users to monitor the data compression process and locate errors in data compression (for example, data compression freezes or stops) s position.
  • using the data compression model to determine the data occupation information in the tree structure includes: the sender inputs the data occupation information of the i-1th layer node in the tree structure into the cyclic network layer, and obtains the ith The data possession information of the layer node, i is a positive integer, and the i-1 layer node is the parent node of the i layer node.
  • the data processing method provided by the present application can use the data of the parent node of the node to be encoded when the context prediction of the node to be encoded is performed.
  • Possession information is used to obtain the data possession information of the node to be encoded, and there is no need to perform feature fusion between the features of the node to be encoded and the features of its parent nodes, which reduces the use of the MLP network, reduces the complexity of the data compression model, and reduces the acquisition of context prediction.
  • the computing resources required for data to occupy information are used to obtain the data possession information of the node to be encoded, and there is no need to perform feature fusion between the features of the node to be encoded and the features of its parent nodes, which reduces the use of the MLP network, reduces the complexity of the data compression model, and reduces the acquisition of context prediction.
  • the computing resources required for data to occupy information is used to obtain the data possession information of the node to be encoded, and there is no need to perform feature fusion between the features of the node to be encoded and the features of its parent nodes, which reduces the use of the MLP network, reduces the complexity of the data compression model, and reduces the acquisition
  • using the data compression model to determine the data occupation information in the tree structure includes: the sending end converts the data occupation information of the i-1th layer node in the tree structure, the data occupation information of the i-1th layer node in the tree structure. At least one of the summary information and the characteristics of the i-th layer node is input to the recurrent network layer to obtain the data possession information of the i-th layer node, and the summary information of the i-1 layer node is used to describe the ancestor nodes of the i-1 layer node to All prediction information of the nodes at layer i-1.
  • the recurrent network layer does not need to use the intermediate features obtained by fusing the features of the node to be encoded and the features of its parent nodes, the sender does not need to store the features of the node to be encoded, the features of the parent node, and the intermediate features, which reduces the need for the sender to perform Storage space required for context prediction.
  • the context prediction of each node needs to start from the root node.
  • the recurrent network layer can utilize the characteristics of the node to be encoded and the node to be encoded.
  • the data occupancy information and summary information of the node to be encoded are obtained by extracting the summary information from the parent node of the node, so that the data processing method provided by the present application does not need to start from the root node, which reduces the computing resources required for context prediction.
  • the data processing method further includes: the sending end stores the data occupation information of the i-1th layer node, the summary information of the i-1th layer node and the characteristics of the i-th layer node in the tree structure At least one of them is input to the cyclic network layer, and the summary information of the nodes in the i-th layer is obtained, and the summary information of the nodes in the i-th layer is used to describe all the prediction information from the ancestor nodes of the nodes in the i-th layer to the nodes in the i-th layer.
  • the summary information of the nodes in the i-th layer can be obtained in an iterative manner.
  • the recurrent network layer includes at least one long short-term memory (LSTM) layer
  • the hyperbolic tangent function (tanh) of the LSTM layer ) and gate structure ( ⁇ function) can be used to selectively use the previous information of the node in the i-th layer to obtain the data occupation information and summary information of the node to be encoded.
  • the recurrent network layer in the data compression model can utilize all prediction information from the root node to the parent node of the node to be encoded, and these prediction information can be selectively memorized and stored by the recurrent network layer.
  • the recurrent network layer can use the information of all ancestor nodes of the node to be encoded, which improves the accuracy of context prediction, and uses the data possession information obtained in this application to treat the compressed data. Compression improves the data compression ratio.
  • the data compression model further includes a feature extraction layer
  • the data processing method further includes: the sender calculates the position, depth and sub-node number of the i-th layer node, as well as the occupancy status of the i-1-th layer node At least one of the bytes is input to the feature extraction layer, and the features of the nodes in the i-th layer are obtained.
  • the above-mentioned feature extraction layer includes at least one layer of MLP.
  • the sender inputs the position, depth and child node number of the i-th layer node, the occupancy status byte of the i-1th layer node, and the occupancy status byte of at least one sibling node of the i-th layer node.
  • Feature extraction layer to get the features of the i-th layer node.
  • the sibling nodes of the i-th layer node refer to other nodes that belong to the same parent node as the i-th layer node.
  • the sender can When the layer nodes perform context prediction, the prediction probability of the occupancy status bytes of some obviously impossible i-th layer nodes is reduced, and the prediction accuracy of the data occupancy information of the i-th layer nodes is improved.
  • the data compression model further includes a dimension adjustment layer
  • the data processing method further includes: the sender inputs the data occupancy information of the i-th layer node into the dimension adjustment layer to obtain an occupancy rate prediction table, and the occupancy rate The prediction table indicates the predicted probability of each occupancy byte of the i-th level node.
  • the dimension adjustment layer includes at least one layer of MLP, and the MLP can be used to adjust the output dimension of the data occupancy information to obtain a visualized prediction probability result.
  • the present application provides a data processing method, and the method can be applied to a receiving end, or the method can be applied to a computing device that can implement the method, for example, the computing device includes a chip system, and the method includes: the receiving end obtains The data is compressed, and the data occupancy information in the tree structure is determined by using the recurrent network layer included in the data compression model, and the data occupancy information is used to indicate the data distribution of the compressed data in the tree structure. The receiving end also decompresses the compressed data according to the data possession information to obtain decompressed data.
  • a recurrent network layer is used to replace the multi-layer MLP network in the prior art for context prediction, which reduces the computing resources required for context prediction to obtain data occupancy information; in addition, compared with the prior art, each node stores
  • the data processing method provided by the present application does not need to store the features and the intermediate features, thereby reducing the storage space occupied by data compression.
  • the recurrent network layer may include at least one LSTM layer.
  • using the data compression model to determine the data occupation information in the tree structure includes: the receiving end inputs the data occupation information of the i-1th layer node in the tree structure into the cyclic network layer, and obtains the ith The data possession information of the layer node, i is a positive integer, and the i-1 layer node is the parent node of the i layer node.
  • the receiving end uses the data compression model to determine the data occupation information in the tree structure, including: At least one of the summary information of the node and the characteristics of the node at layer i is input to the recurrent network layer to obtain the data occupation information of the node at layer i, and the summary information of the node at layer i-1 is used to describe the ancestor of the node at layer i-1 All prediction information from nodes to i-1 layer nodes.
  • the data processing method further includes: the receiving end stores the data occupation information of the i-1th layer node, the summary information of the i-1th layer node, and the characteristics of the i-th layer node in the tree structure At least one of them is input to the cyclic network layer, and the summary information of the nodes in the i-th layer is obtained, and the summary information of the nodes in the i-th layer is used to describe all the prediction information from the ancestor nodes of the nodes in the i-th layer to the nodes in the i-th layer.
  • the data compression model further includes a feature extraction layer
  • the data processing method further includes: the receiving end calculates the position, depth and sub-node number of the i-th layer node, as well as the occupancy status of the i-1th layer node At least one of the bytes is input to the feature extraction layer, and the features of the nodes in the i-th layer are obtained.
  • the feature extraction layer includes at least one layer of MLP.
  • the receiving end inputs at least one of the position, depth and child node number of the node at layer i, and the occupancy byte of the node at layer i-1 into the feature extraction layer, and obtains the value of the node at layer i.
  • Features including: the receiving end inputs the position, depth and child node number of the i-th layer node, the occupancy byte of the i-1 layer node, and the occupancy byte of at least one sibling node of the i-th layer node into the feature extraction layer , get the features of the i-th layer node.
  • the data compression model further includes a dimension adjustment layer
  • the data processing method further includes: the receiving end inputs the data occupancy information of the i-th layer node into the dimension adjustment layer to obtain an occupancy rate prediction table, and the occupancy rate prediction table is obtained.
  • the table indicates the predicted probability of each occupancy byte for the i-th tier node.
  • the dimension adjustment layer includes at least one layer of MLP.
  • the present application provides a data processing apparatus, and the beneficial effects can be found in the description of any aspect of the first aspect, which will not be repeated here.
  • the data processing apparatus has a function to implement the behavior in the method example of any one of the above-mentioned first aspects.
  • the functions can be implemented by hardware, or can be implemented by hardware executing corresponding software.
  • the hardware or software includes one or more modules corresponding to the above functions.
  • the data processing apparatus is applied to the sending end, and the data processing apparatus includes: a preprocessing unit, used for generating tree-structured data to be compressed according to the original data; a context prediction unit, used for utilizing a data compression model Determine the data occupation information in the tree structure, the data occupation information is used to indicate the data distribution of the original data in the tree structure, the data compression model includes a recurrent network layer, and the recurrent network layer is used to determine the data occupation information; coding unit, used for The to-be-compressed data is compressed according to the data occupation information to obtain compressed data.
  • the data processing device further includes: a collection unit for collecting raw data through sensors, where the raw data is three-dimensional data, and the sensors include lidar, millimeter wave At least one of radar and sonar.
  • the data processing apparatus further includes: a collection unit configured to collect raw data through a camera, where the original data is two-dimensional data.
  • the data processing apparatus further includes: a display unit, configured to display the tree structure and/or data occupation information.
  • the context prediction unit is specifically used to input the data occupancy information of the i-1th layer node in the tree structure into the cyclic network layer to obtain the ith layer node.
  • the data possession information of , i is a positive integer
  • the i-1 layer node is the parent node of the i layer node.
  • a recurrent network layer includes at least one LSTM layer.
  • the context prediction unit is specifically configured to use the data occupancy information of the i-1th layer node and the summary information of the i-1th layer node in the tree structure and at least one of the features of the nodes in the i-th layer are input into the recurrent network layer to obtain the data possession information of the nodes in the i-th layer, and the summary information of the nodes in the i-1 layer is used to describe the ancestor nodes of the nodes in the i-1 layer to the i-th layer.
  • the context prediction unit is also used to convert the data occupancy information of the i-1th layer node and the summary information of the i-1th layer node in the tree structure. And at least one of the features of the nodes in the i-th layer is input into the recurrent network layer, and the summary information of the nodes in the i-th layer is obtained. The summary information of the nodes in the i-th layer is used to describe all the predictions from the ancestors of the nodes in the i-th layer to the nodes in the i-th layer. information.
  • the data compression model further includes a feature extraction layer
  • the context prediction unit is also used to calculate the position, depth and sub-node number of the i-th layer node, and the number of the i-th layer node.
  • At least one of the occupancy status bytes of the nodes in the i-1 layer is input to the feature extraction layer, and the features of the nodes in the i-th layer are obtained.
  • the feature extraction layer includes at least one layer of MLP.
  • the context prediction unit is further used to convert the position, depth and child node number of the node at the i-th layer, the occupancy status byte of the node at the i-1-th layer, and at least one sibling node of the node at the i-th layer
  • the occupancy bytes are input into the feature extraction layer, and the features of the nodes in the i-th layer are obtained.
  • the data compression model further includes a dimension adjustment layer
  • the context prediction unit is further configured to input the data occupancy information of the i-th layer node into the dimension adjustment layer,
  • the occupancy rate prediction table is obtained, and the occupancy rate prediction table indicates the prediction probability of each occupancy condition byte of the node of the i-th layer.
  • the dimension adjustment layer includes at least one layer of MLP.
  • the present application provides a data processing apparatus, and the beneficial effects can be found in the description of any two aspects of the second aspect, and details are not repeated here.
  • the data processing apparatus has the function of implementing the behavior in the method instance of any two of the above-mentioned second aspects.
  • the functions can be implemented by hardware, or can be implemented by hardware executing corresponding software.
  • the hardware or software includes one or more modules corresponding to the above functions.
  • the data processing apparatus is applied to the receiving end, and the data processing apparatus includes: an acquisition unit, used for acquiring compressed data; a context prediction unit, used for determining data occupancy information in a tree structure by using a data compression model , the data occupation information is used to indicate the data distribution of the compressed data in the tree structure, the data compression model includes a cyclic network layer, and the cyclic network layer is used to determine the data occupation information; the decompression unit is used to decompress the compressed data according to the data occupation information, and obtain Unzip the data.
  • the context prediction unit is specifically used to input the data occupancy information of the i-1th layer node in the tree structure into the cyclic network layer to obtain the ith layer node.
  • the data possession information of , i is a positive integer
  • the i-1 layer node is the parent node of the i layer node.
  • a recurrent network layer includes at least one LSTM layer.
  • the context prediction unit is specifically used to convert the data occupation information of the i-1th layer node and the summary information of the i-1th layer node in the tree structure. and at least one of the features of the nodes in the i-th layer are input into the recurrent network layer to obtain the data possession information of the nodes in the i-th layer, and the summary information of the nodes in the i-1 layer is used to describe the ancestor nodes of the nodes in the i-1 layer to the i-th layer. - All prediction information for layer 1 nodes.
  • the context prediction unit is also used to convert the data occupation information of the nodes at the i-1th layer and the summary information of the nodes at the i-1th layer in the tree structure. And at least one of the features of the nodes in the i-th layer is input into the recurrent network layer, and the summary information of the nodes in the i-th layer is obtained. The summary information of the nodes in the i-th layer is used to describe all the predictions from the ancestors of the nodes in the i-th layer to the nodes in the i-th layer. information.
  • the data compression model further includes a feature extraction layer
  • the context prediction unit is also used to calculate the position, depth and sub-node number of the i-th layer node, and the number of the i-th layer node.
  • At least one of the occupancy status bytes of the nodes in the i-1 layer is input to the feature extraction layer, and the features of the nodes in the i-th layer are obtained.
  • the feature extraction layer includes at least one layer of MLP.
  • the context prediction unit is further used to convert the position, depth and child node number of the node at the i-th layer, the occupancy status byte of the node at the i-1-th layer, and at least one sibling node of the node at the i-th layer
  • the occupancy bytes are input into the feature extraction layer, and the features of the nodes in the i-th layer are obtained.
  • the data compression model further includes a dimension adjustment layer
  • the context prediction unit is further configured to input the data occupancy information of the i-th layer node into the dimension adjustment layer,
  • the occupancy rate prediction table is obtained, and the occupancy rate prediction table indicates the prediction probability of each occupancy condition byte of the node of the i-th layer.
  • the dimension adjustment layer includes at least one layer of MLP.
  • the present application provides a computing device, the computing device comprising at least one processor and a memory, where the memory is used to store a set of computer instructions; when the processor executes the set of computer instructions, the first aspect or the first aspect is executed. Any one of the possible implementations of the one aspect, or the operation steps of the data processing method in any of the second aspect and any of the possible implementations of the second aspect.
  • the present application provides a computer-readable storage medium, in which a computer program or instruction is stored, and when the computer program or instruction is executed by a computing device, any one of the first aspect and the first aspect may be implemented. manner, or operation steps of the method of the second aspect and any one of the possible implementation manners of the second aspect.
  • the present application provides a computer program product, which, when the computer program product runs on a computer, enables a computing device to execute any one of the possible implementations of the first aspect and the first aspect, or the second aspect and the second aspect Operation steps of the method of any of the possible implementations.
  • the present application provides a chip, including a memory and a processor, the memory is used to store computer instructions, and the processor is used to call and run the computer instructions from the memory to execute the above-mentioned first aspect and any possibility of the first aspect.
  • the method in the implementation manner of the second aspect, or the operation steps of the method in any possible implementation manner of the second aspect and the second aspect.
  • the present application may further combine to provide more implementation manners.
  • FIG. 1 is a schematic diagram of a scenario of a communication system provided by the present application.
  • Fig. 2 is a kind of system schematic diagram of point cloud compression and decompression provided by this application;
  • FIG. 3 is a schematic diagram of a point cloud compression and decompression process provided by the present application.
  • Fig. 4 is the schematic diagram of the context prediction method of a kind of Oct Squeeze algorithm provided by the prior art
  • FIG. 5 is a schematic flowchart of a data processing method provided by the present application.
  • FIG. 6 is a schematic diagram of a tree structure provided by the application.
  • FIG. 7 is a schematic diagram of a data compression model provided by the application.
  • FIG. 8 is a schematic flowchart of another data processing method provided by the present application.
  • FIG. 9 is a schematic diagram of the network structure of a LSTM provided by the application.
  • FIG. 10 is a schematic display diagram of a data processing provided by the application.
  • FIG. 11 is a schematic structural diagram of another data compression model provided by the application.
  • FIG. 12 is a schematic diagram of a data processing device provided by the application.
  • FIG. 13 is a schematic structural diagram of a computing device provided by the present application.
  • a point cloud is a data set of points.
  • the points in the point cloud can be represented by three-dimensional coordinates (X, Y, Z).
  • the points on the three-dimensional coordinates (X, Y, Z) can include color, classification value and intensity value. and other attribute information.
  • FIG. 1 is a schematic diagram of a scenario of a communication system provided by the present application.
  • the communication system includes at least one terminal (terminal 111 to terminal 113 as shown in FIG. 1 ), a network, and a data center 130 .
  • the terminals and the data center 130 may communicate through a network, which may be an internetwork.
  • a terminal may also be called a terminal device, a user equipment (UE), a mobile station (mobile station, MS), a mobile terminal (mobile terminal, MT), and the like.
  • UE user equipment
  • MS mobile station
  • MT mobile terminal
  • the terminal may be a mobile phone (the terminal 111 shown in FIG. 1 ), a tablet computer (the terminal 112 shown in FIG. 1 ), a computer with a wireless transceiving function (as shown in FIG. 1 ) shown terminal 113), virtual reality (Virtual Reality, VR) terminal device, augmented reality (Augmented Reality, AR) terminal device, wireless terminal in industrial control (industrial control), wireless terminal in self driving (self driving) Terminals (such as lidar integrated on vehicle 121 and vehicle 122 shown in FIG. 1 ), wireless terminals in transportation safety, wireless terminals in smart cities, and the like.
  • VR Virtual Reality
  • AR Augmented Reality
  • Terminals such as lidar integrated on vehicle 121 and vehicle 122 shown in FIG. 1
  • wireless terminals in transportation safety wireless terminals in smart cities, and the like.
  • the terminal may also be a smart home terminal, such as a smart screen, installed in a residence.
  • the terminal can also be a terminal set up in a hospital for remote medical surgery; for example, when performing non-invasive surgery, the terminal can be used to collect the internal information of the patient (such as the point cloud of the internal organs of the human body) .
  • the data center 130 may be a server cluster including at least one application server 131 , or may be a cloud data center constructed by the application server 131 .
  • multiple application servers can be independent and different physical devices, or the functions of multiple application servers can be integrated on the same physical device (for example, multiple application servers under the jurisdiction of a cloud service provider), and also It can be a physical device that integrates some application server functions.
  • the terminal is connected to the application server 131 in a wireless or wired manner.
  • Terminals can be fixed-position or movable.
  • the embodiments of the present application do not limit the number of terminals and application servers included in the communication system.
  • the terminal can implement functions such as collecting point clouds, compressing point clouds, or decompressing point clouds
  • the data center 130 can implement functions such as compressing point clouds or decompressing point clouds.
  • the vehicle-mounted terminal can collect point cloud and compressed point cloud
  • the data center can decompress the point cloud as an example.
  • the lidar collects point clouds of office buildings, residences and basketball courts on the roadside, and when turning right, the lidar collects the point clouds of the plants (trees shown in FIG. 1 ) on both sides of the roadside,
  • the processing device transmits the aforementioned point cloud to the data center 130 .
  • point cloud compression routes can be roughly divided into two categories: traditional point cloud compression algorithms and point cloud compression algorithms based on artificial intelligence (AI) technology.
  • AI artificial intelligence
  • traditional point cloud compression algorithms can be divided into two categories: the first is that point cloud compression equipment converts point clouds into two-dimensional signals, and then uses existing image compression algorithms to further compress, for example, dynamic images
  • the Planar Projection-based Compression Algorithm (MPEG V-PCC) provided by the Moving Pictures Experts Group (MPEG).
  • MPEG Moving Pictures Experts Group
  • the second category is the point cloud compression device that converts the point cloud into a tree structure for entropy coding, such as the kd-tree-based Draco algorithm and the octree-based geometry-based compression algorithm (MPEG G-PCC).
  • the point cloud compression device may be a device that compresses point clouds.
  • the point cloud compression device may be any one of the terminals 111 to 113 shown in FIG. 1 .
  • AI-based point cloud compression algorithms can be divided into two categories: one is a hybrid algorithm based on a traditional algorithm framework and uses an AI algorithm to replace the context encoding module, such as the Oct Squeeze algorithm.
  • Another class of algorithms are end-to-end AI (full AI) compression algorithms, such as autoencoder algorithms based on "Point net++" features.
  • FIG. 2 is a schematic diagram of a point cloud compression and decompression system provided by the present application.
  • the system includes a transmitter 210 and a receiver 220 .
  • the transmitter 210 establishes a communication connection with the receiver 220 through a communication channel 230 .
  • the above-mentioned sending end 210 can realize the function of data compression.
  • the sending end 210 can be any one of the terminal 111 to the terminal 113 , and the sending end 210 can also be a point cloud set on the vehicle 121 or the vehicle 122
  • a compression system eg, the point cloud compression system includes a lidar and a processing device in communication with the lidar.
  • the sending end 210 may include a data source 211 , a preprocessing module 212 , an encoder 213 and a communication interface 214 .
  • the data source 211 may comprise or be any type of electronic device for capturing point clouds, and/or any type of point data generating device, such as a computer graphics processor for generating computer animation scenes or any type for acquiring And/or devices that provide real-world point clouds, computer-generated point clouds.
  • the data source 211 may be any type of memory or storage that stores any point data in the above point cloud.
  • the point cloud 241 may also be referred to as raw data (or raw point cloud data) 241 .
  • the preprocessing module 212 is configured to receive the point cloud 241 and preprocess the point cloud 241 to obtain the preprocessed data 242 .
  • the preprocessing performed by the preprocessing module 212 may include color format conversion (eg, from RGB to YCbCr), octree structuring, and the like.
  • the encoder 213 is configured to receive the preprocessed data 242 and compress the preprocessed data 242 after performing context prediction to obtain point cloud compressed data 243 .
  • the communication interface 214 in the sender 210 can be used to: receive the point cloud compressed data 243, and send the point cloud compressed data 243 (or the point cloud compressed data 243) to another device such as the receiver 220 or any other device through the communication channel 230. any other processed version) for storage or direct reconstruction.
  • the above-mentioned receiving end 220 can realize the function of data decompression. As shown in FIG. 1 , the receiving end 220 can be any one or more servers in the data center 130 shown in FIG. Data decompression function.
  • the receiving end 220 may include a display device 221 , a post-processing module 222 , a decoder 223 and a communication interface 224 .
  • the communication interface 224 in the receiving end 220 is used to receive the point cloud compressed data 243 (or any other processed version) from the sending end 210 or from any other sending end such as a storage device.
  • the storage device is a point cloud data storage device
  • the point cloud compressed data 243 is supplied to the decoder 223 .
  • Communication interface 214 and communication interface 224 may be used to communicate through a direct communication link between sender 210 and receiver 220, such as a direct wired or wireless connection, etc., or through any type of network, such as a wired network, a wireless network, or any of them. Combination, any type of private network and public network, or any type of combination, send or receive point cloud compressed data 243.
  • the communication interface 214 may be used to encapsulate the point cloud compressed data 243 into a suitable format such as a message, and/or use any type of transfer encoding or processing to process the encoded point cloud compressed data 243 for communication in the communication chain transmission over the road or communication network.
  • Communication interface 224 corresponds to communication interface 214 , for example, and may be used to receive transmission data and process the transmission data using any type of corresponding transmission decoding or processing and/or decapsulation, resulting in point cloud compressed data 243 .
  • Both the communication interface 224 and the communication interface 214 can be configured as a one-way communication interface as indicated by the arrow of the corresponding communication channel 230 from the sending end 210 to the receiving end 220 in FIG. 2, or a two-way communication interface, and can be used to send and receive messages. etc. to establish a connection, acknowledge and exchange any other information related to a communication link and/or data transmission such as encoded compressed data transmission, etc.
  • the decoder 223 is configured to receive the point cloud compressed data 243 and obtain decoded data 244 after performing context prediction on the point cloud compressed data 243 .
  • the post-processing module 222 is configured to perform post-processing on the decoded decoded data 244 to obtain post-processing data 245 .
  • Post-processing performed by post-processing module 222 may include, for example, color format conversion (eg, from YCbCr to RGB), octree reconstruction, etc., or any other processing for generating data for display by display device 221 or the like.
  • Display device 221 is used to receive post-processed data 245 for display to a user or viewer or the like.
  • Display device 221 may be or include any type of display for representing the reconstructed image, eg, an integrated or external display screen or display.
  • the display screen may include a liquid crystal display (LCD), an organic light emitting diode (OLED) display, a plasma display, a projector, a micro LED display, a liquid crystal on silicon (LCoS) display ), digital light processor (DLP), or any other type of display.
  • the sending end 210 and the receiving end 220 may transmit the point cloud compressed data 243 through a data forwarding device.
  • the data forwarding device may be a router, a switch, or the like.
  • Fig. 3 is a schematic diagram of a point cloud compression and decompression process provided by the application, the data in the point cloud is stored in the form of coordinate points, as shown in Fig. 3 "[(x1, y1, z1), (x2 ,y2,z2)...]”.
  • the octreeization module 31 can realize the function of the preprocessing module 212 shown in FIG. 2 .
  • the octreeization module 31 performs octreeization on the point cloud to obtain the Compressed data, the data to be compressed includes a root node and 8 child nodes.
  • an octree is a tree-like data structure used to describe three-dimensional space.
  • Each node of the octree represents a volume element (voxel) of a cube, and each root node has eight child nodes ( The volume of the voxels represented by the eight child nodes is added together to equal the volume of the voxels represented by the parent node.
  • the octree module 31 may assign "1" or "0" to the child node according to whether there is data in the voxel corresponding to the child node. For example, “1” indicates that the voxel corresponding to the child node has data, and "0" indicates that there is no data in the voxel corresponding to the child node.
  • the octree module 31 integrates the values of the eight child nodes and represents them with 8 bits to obtain the occupancy byte of the root node, where the occupancy byte indicates the data distribution of the root node.
  • the black node indicates that there is data in the voxel corresponding to the child node
  • the white node indicates that there is no data in the voxel corresponding to the child node
  • the occupancy of the root node is: 1011 1100 bytes.
  • the context prediction device 32 and the entropy encoder 33 can implement the functions of the encoder 213 shown in FIG. 2 .
  • the entropy encoder 33 Compress the point cloud corresponding to the root node according to the occupancy status byte, so as to realize point cloud compression, and obtain the compressed data of the point cloud.
  • the entropy coding algorithm adopted by the entropy encoder 33 and the entropy decoder 35 may be any one or a combination of the following: Shannon coding, Huffman coding and arithmetic coding (arithmetic coding). For a specific implementation manner, please refer to the relevant descriptions in the prior art, which will not be repeated.
  • the context prediction device 34 and the entropy decoder 35 can implement the functions of the decoder 223 shown in FIG. 2 , for example, use the same context prediction module 34 as the point cloud compression to predict the occupancy of the root node in the compressed data Status byte, the entropy decoder 35 decompresses the compressed data corresponding to the node to be decoded according to the occupancy status byte of the root node, and the octree reconstruction module 36 reconstructs the decompressed compressed data corresponding to the node to be decoded , and after each layer of nodes is decompressed and reconstructed, the decompressed point cloud is obtained.
  • the entropy encoder will use the occupancy bytes predicted by the context prediction module.
  • the larger the compressed data compression ratio the ratio of the point cloud to the amount of compressed data, the better the point cloud compression effect.
  • the context prediction method provided by the Oct Squeeze algorithm is taken as an example.
  • the Oct Squeeze algorithm octrees the point cloud to obtain a multi-layer node. It is assumed that the multi-layer node includes at least 4 layers of nodes, as shown in the figure As shown in 4, Fig. 4 is a schematic diagram of the context prediction method of a kind of Oct Squeeze algorithm provided by the prior art, the point cloud after the octreeization includes great-grandfather node, grandfather node, parent node and node to be encoded, the depth of the aforementioned node.
  • the sequence is: i-3, i-2, i-1 and i, where i is a positive integer greater than or equal to 3.
  • the depth of the node represents the number of layers of the node from the root node to the current position.
  • the depth of the root node of the octree is 0, and the depth of the child nodes of the root node is 1.
  • the root node may be referred to as a layer 0 node
  • the child nodes may be referred to as a layer 1 node.
  • the context prediction method shown in FIG. 4 includes a feature extraction process.
  • an MLP network (such as 5-layer MLP layers) is used to perform feature extraction on each layer of nodes. As shown in FIG. 4 , the MLP network obtains each layer of nodes separately. , which is a high-dimensional feature vector.
  • the context prediction method shown in FIG. 4 also includes a feature fusion process.
  • an MLP network (such as 4-layer MLP layers) is used to perform a wave-net (WaveNet) step-by-step fusion of the features of each layer of nodes, as shown in FIG. 4
  • the MLP network fuses the features of the node to be encoded (such as the i-1 layer node) with the features of its parent node (such as the i-2 layer node), and after multi-level feature fusion, obtains the to-be-coded node.
  • the occupancy prediction result of the encoded node indicates the data distribution of the voxel midpoint corresponding to the node to be encoded.
  • WaveNet For the principle of WaveNet, reference may be made to the related elaboration in the prior art, which will not be repeated in this application.
  • the point cloud compression device will use the MLP network to repeatedly extract the features that have been extracted, which will increase the repeated calculation process; in order to reduce the computational resources required for context prediction, point cloud compression The compression device will store each feature and the fused features, however, this will take up a lot of storage space of the point cloud compression device.
  • the present application provides a data processing method, which includes: a sending end generates tree-structured data to be compressed according to original data, and uses a cyclic network layer included in a data compression model to determine In the data occupation information in the tree structure, the data occupation information is used to indicate the data distribution of the original data in the tree structure, and the sender compresses the above-mentioned to-be-compressed data according to the data occupation information to obtain compressed data.
  • the present application uses a recurrent network layer to replace the multi-layer MLP network in the prior art for context prediction, which reduces the complexity of the data compression model and reduces the computing resources required for context prediction to obtain data occupancy information; in addition, compared with In the prior art, the features of each node and the intermediate features obtained by fusing the features are stored.
  • the data processing method provided by the present application does not need to store the features and the intermediate features, thereby reducing the storage space occupied by data compression.
  • FIG. 5 is a schematic flowchart of a data processing method provided by this application.
  • the data processing method includes the following steps.
  • the sending end generates tree-structured data to be compressed according to the original data.
  • the sending end may be a mobile terminal (eg, a mobile phone, a tablet computer, etc.) deployed on the vehicle, and the sending end may also be a point cloud compression system mounted on the vehicle, for example, the point cloud compression system includes lidar and Processing equipment for LiDAR communication connections.
  • a mobile terminal eg, a mobile phone, a tablet computer, etc.
  • the sending end may also be a point cloud compression system mounted on the vehicle, for example, the point cloud compression system includes lidar and Processing equipment for LiDAR communication connections.
  • the raw data may be three-dimensional data, eg a point cloud.
  • the three-dimensional data may be collected by sensors.
  • the sensor may include at least one of lidar, millimeter-wave radar, and sonar, and the sensor may be integrated on the transmitter, or separately deployed on the same vehicle as the transmitter (vehicle 122 shown in FIG. 1 ). ) and establishes a communication connection with the sender.
  • the sensor may also be a sensor array, and each sensor in the sensor array may be respectively deployed on the vehicle 122 and the reference object to be collected (the tree shown in FIG. 1 ).
  • the above-mentioned point cloud may contain various kinds of information.
  • the point cloud is collected by a laser measurement device, it includes three-dimensional coordinates and laser reflection intensity.
  • the point cloud is collected by photogrammetry equipment, it may include three-dimensional coordinates and color information.
  • the point cloud is obtained after the laser measurement equipment and the photogrammetry equipment are jointly collected and merged, it may include three-dimensional coordinates, laser reflection intensity and color information.
  • the raw data may be two-dimensional data, eg image data.
  • the two-dimensional data may be collected by a camera.
  • the camera can be integrated on the sending end, or can be deployed on the vehicle 122, and establishes a communication connection with the sending end.
  • the camera may include one or more cameras, for example, the camera may be disposed at one or more positions of the front windshield, rear windshield, roof, and body of the vehicle 122 .
  • the above-mentioned tree structure may be an M-ary tree, where M is a positive integer greater than or equal to 2, for example, an octree, a quadtree, or a binary tree.
  • the sending end uses the data compression model to determine the data occupation information in the tree structure.
  • the above-mentioned data occupation information is used to indicate the data distribution of the original data in the tree structure.
  • the data occupation information can indicate the distribution of points in the voxel.
  • the data occupation information indicates the occupation of the root node in the data to be compressed.
  • the byte is "1000 1000 ' with a predicted probability of 99%.
  • the node with a depth of i-3 is the root node, as shown in (a) in Figure 6, the root node has 8 child nodes, the 8 The number of child nodes is 1 to 8.
  • the data occupancy information of the root node is shown in (b) of FIG. 6 , the data occupancy information indicates that the voxels corresponding to the two child nodes with serial numbers “1” and “5” in the voxels corresponding to the root node have data.
  • the context prediction can be performed on the node with the sequence number “1” and the depth i-2 shown in (b) in FIG. 6 , and the obtained data occupation information is shown in (c) in FIG. 6 .
  • the data occupancy information indicates that the voxels corresponding to two nodes of depth i-1 with serial numbers "1" and "5" in the voxels corresponding to the node with depth i-2 have data in them.
  • context prediction can be performed on the node with the sequence number “5” and the depth i-2 shown in (b) in FIG. 6 , and the obtained data occupation information is shown in (d) in FIG. 6 .
  • the data occupancy information indicates that the voxel corresponding to the node with the depth i-1 with the serial number "3" in the voxel corresponding to the node with the depth i-2 has data in the voxel.
  • context prediction can be performed on the node with the sequence number "5" and the depth i-1 shown in (c) in FIG. 6 , and the obtained data occupation information is shown in (e) in FIG. 6 .
  • the data occupancy information indicates that the voxels corresponding to the two nodes with the depth i and the sequence numbers "1" and "5" in the voxels corresponding to the node with the depth i have data in the voxels.
  • the data compression model described above includes a recurrent network layer for determining the data occupancy information.
  • the recurrent network layer may include at least one convolutional neural network layer, for example, a recurrent neural network (RNN), an LSTM layer, a gated recurrent unit (GRU), and the like.
  • RNN recurrent neural network
  • LSTM LSTM layer
  • GRU gated recurrent unit
  • the recurrent network layer may include any combination of the above multiple convolutional neural network layers.
  • the data compression model also contains a feature extraction layer.
  • the feature extraction layer is used to determine the feature of the node to be encoded according to at least one of the position, depth and child node number of the node to be encoded, and the occupancy status of the parent node of the node to be encoded, and the feature is a high-dimensional feature vector.
  • the number of layers of the feature extraction layer and the number of layers of the recurrent network layer can be determined according to the depth of the node.
  • FIG. 7 is a schematic diagram of a data compression model provided by the present application.
  • the data compression model 700 includes a feature extraction layer 710 and a recurrent network layer 720 .
  • the feature extraction layer 710 includes a first extraction layer 711 , a second extraction layer 712 , a third extraction layer 713 and a fourth extraction layer 714 .
  • the above-mentioned first extraction layer 711 to fourth extraction layer 714 may all be multiple layers MLP.
  • the above-mentioned first extraction layer 711 to fourth extraction layer 714 may be different MLP networks.
  • the number of MLP layers included in the feature extraction layer is different.
  • the second extraction layer 712 includes a 3-layer MLP
  • the third extraction layer 713 includes a 4-layer MLP.
  • the above-mentioned first extraction layer 711 to fourth extraction layer 714 may be the same MLP network.
  • the extraction layer can be reused.
  • the first extraction layer 711 to the fourth extraction layer 714 all include four layers of the same MLP.
  • the recurrent network layer 720 includes a first network layer 721 , a second network layer 722 , a third network layer 723 and a fourth network layer 724 .
  • the above-mentioned first network layer 721 to fourth network layer 724 may all be LSTMs.
  • the above-mentioned first network layer 721 to fourth network layer 724 all include three identical LSTM layers.
  • the sender may input the data occupancy information of the i-1th layer node in the tree structure into the recurrent network layer to obtain the data occupancy information of the i-th layer node.
  • i is a positive integer
  • the i-1 layer node is the parent node of the i layer node.
  • i ⁇ 3 is taken as an example for illustration.
  • the great-grandfather node 731 is the node with a depth of i-3 shown in FIG. 6
  • the grandfather node 732 is the first node with a depth of i-2 shown in FIG. 6
  • a black node, the parent node 733 is the second black node with a depth i-1 shown in FIG. 6
  • the node to be encoded 734 is the first black node with a depth i shown in FIG. 6 .
  • the data occupation information of the parent node 733 is shown in (e) of FIG. 6 .
  • the data occupation information indicates that the voxels corresponding to the parent node 733 have serial numbers “1” and “5”
  • the probability of having data in the voxels corresponding to the two child nodes of is 98%.
  • the data processing method provided by the present application can use the data of the parent node of the node to be encoded when the context prediction of the node to be encoded is performed.
  • Possession information is used to obtain the data occupancy information of the node to be encoded, and there is no need to perform feature fusion between the features of the node to be encoded and the features of its parent nodes, which reduces the use of the MLP network and reduces the complexity of the data compression model; in addition, due to the cyclic network
  • the layer does not need to use the intermediate features obtained by fusing the features of the node to be encoded and the features of its parent nodes, so the sender does not need to store the features of the node to be encoded, the features of the parent node, and the intermediate features, which reduces the amount of time required for the sender to perform context prediction. required storage space.
  • FIG. 8 is another data processing provided by this application.
  • a schematic flowchart of the method, S520 may include the following steps.
  • the sender inputs at least one of the position, depth and sub-node number of the node at the i-th layer, and the occupancy status of the node at the i-1-th layer into the feature extraction layer to obtain the feature of the node at the i-th layer.
  • the feature extraction layer 710 can obtain the feature of the node according to the explicit information of the node.
  • the fourth extraction layer 714 can obtain the feature of the node according to the position, depth and child node number of the node 734 to be encoded, and the occupancy of the parent node 733 In the case of bytes, the feature X t of the node 734 to be encoded is obtained.
  • This feature may also be called implicit feature or implicit information, etc.
  • the feature is a high-dimensional feature vector.
  • the feature of each layer of nodes is represented by X.
  • the feature of the node to be encoded 734 is X t
  • the feature of the parent node 733 is X t-1
  • the feature of the grandfather node 732 is X t-2
  • the great-grandfather node 731 is characterized by X t-3 .
  • the above-mentioned position indicates the position of the voxel corresponding to the node 734 to be encoded in the voxel corresponding to the parent node 733; the above-mentioned depth indicates that the depth of the node to be encoded 734 is i.
  • the above-mentioned child node number indicates the relative position of the voxel corresponding to the child node of the node to be encoded 734 in the voxel corresponding to the node to be encoded 734.
  • the relative position means that the child node is to be encoded Which of the eight child nodes of node 734 is specific.
  • the number of the child nodes may be 1 to 8, and the 8 numbers respectively represent eight different child nodes.
  • the numbers of the child nodes may also be 0 to 7 and so on.
  • the number of the child nodes can be adjusted according to the compression requirements of the point cloud, which is not limited in this application.
  • the above-mentioned S5201 specifically includes: the sender assigns the position, depth and sub-node number of the i-th layer node, the occupancy status byte of the i-1-th layer node, and at least one of the i-th layer nodes.
  • the occupancy status of sibling nodes is input into the feature extraction layer, and the features of the nodes in the i-th layer are obtained.
  • the sibling nodes of the i-th layer node refer to other nodes that belong to the same parent node (i-1-th layer node) as the i-th layer node.
  • the node to be encoded 734 is the first node of depth i shown in FIG. 6 .
  • the sibling node can be the second black node (right) at depth i.
  • the occupancy status byte of the sibling node may be determined by the sender using the data occupancy information of the sibling node to compress the data in the voxel corresponding to the sibling node.
  • the sender Since the characteristics of the i-th layer node include the occupancy status bytes of its parent node and sibling nodes, the sender reduces the occupancy status words of some obviously impossible i-th layer nodes when making context prediction for the i-th layer node. The prediction probability of the node is improved, and the prediction accuracy of the data occupancy information of the i-th layer node is improved.
  • the probability that the occupancy status byte of the i-th layer node obtained by the context prediction is "1111 1111" is 0.05%; if the characteristics of the i-th layer node include the occupancy status byte of its sibling node, and the occupancy status byte of the sibling node is "1111 1111", the occupancy status byte of the i-th layer node predicted by the context is The probability of "1111 1111" will be less than 0.05%, such as 0.01%.
  • the sender inputs at least one of the data occupancy information of the i-1 layer node in the tree structure, the summary information of the i-1 layer node, and the characteristics of the i layer node into the cyclic network layer to obtain the i layer node data possession information.
  • the summary information of the i-1 layer node is used to describe all prediction information from the ancestor node of the i-1 layer node to the i-1 layer node.
  • the summary information C t-1 of the parent node 733 is used to describe all prediction information from the ancestor nodes of the parent node 733 to the parent node 733 .
  • the ancestor node of the node at the i-1 level may be any node from the parent node to the root node of the node at the level i-1.
  • the ancestor node of the parent node 733 may be the grandfather node 732 or the great-grandfather node 731 shown in FIG. 7 , or any ancestor node of the great-grandfather node 731 , such as the octree where the parent node 733 is located. 's root node.
  • the fourth network layer 724 in the recurrent network layer 720 may occupy the data h t-1 of the parent node 733 and the summary information C t-1 of the parent node 733 according to the and the feature X t of the node to be encoded 734 to obtain the data occupation information h t of the node to be encoded 734 .
  • the sender inputs at least one of the data occupancy information of the i-1 layer node in the tree structure, the summary information of the i-1 layer node, and the characteristics of the i layer node into the cyclic network layer to obtain the i layer node summary information.
  • the summary information of the i-th layer node is used to describe all the prediction information from the ancestor node of the i-th layer node to the i-th layer node.
  • the summary information of the node to be encoded 734 can be represented by C t .
  • the above-mentioned summary information C t may be obtained in an iterative manner.
  • the cyclic network layer is an LSTM layer as an example for illustration, as shown in FIG. 9 , which is provided by this application.
  • a schematic diagram of the network structure of an LSTM which includes two hyperbolic tangent functions (tanh) and three gate structures ( ⁇ functions).
  • Both Tanh and ⁇ functions are used to selectively use the previous information of the i-th layer node to obtain the data occupation information and summary information of the node to be encoded.
  • the preceding information may include data occupancy information h t-1 of the parent node 733 , summary information C t-1 , and feature X t of the node to be encoded 734 .
  • the LSTM layer can use h t-1 , C t-1 and X t to obtain the summary information C t of the node to be encoded 734 and the data occupancy information h t .
  • the summary information C t of the i-th layer node can be used to predict the data occupancy information of the nodes in the next layer (i+1-th layer nodes), just as the summary information C t- of the i-1-th layer nodes 1 Participated in the prediction of data possession information of the i-th layer nodes, which will not be repeated here.
  • the relevant principles of the Tanh and ⁇ functions in the LSTM layer please refer to the relevant elaboration on LSTM in the prior art, and will not be repeated.
  • the context prediction of each node needs to start from the root node.
  • the recurrent network layer can use the characteristics of the node to be encoded and the parent of the node to be encoded.
  • the summary information (C t-1 ) extracted by the node is obtained to obtain the data occupation information and summary information (C t ) of the node to be encoded, so that the data processing method provided by this application does not need to start from the root node, which reduces the need for context prediction. computing resources.
  • the fourth network layer 724 receives the data h t-1 and C t-1 output by the third network layer, and predicts the data occupancy information h t of the node to be encoded 734 by using the feature X t of the node to be encoded 734 When , h t-1 , C t-1 and X t are integrated to obtain the summary information C t of the node 734 to be encoded.
  • the point cloud compression device in the prior art needs to store the features of each node and the features obtained by fusion, while this application uses a recurrent network layer to replace the multi-layer MLP network required for feature aggregation, reducing data compression.
  • the complexity of the model in addition, the recurrent network layer can extract and transmit the information of ancestor nodes, so that the sender does not need to store a large number of features, which saves the storage space of the sender.
  • the sender compresses the data to be compressed according to the data possession information to obtain compressed data.
  • the entropy encoder in the sending end obtains compressed data corresponding to the original data by performing multi-stage node-by-stage compression on the data to be compressed.
  • the great-grandfather node 731 with a depth of i-3 is the root node, and the point cloud compression process includes the following multi-level compression steps.
  • the first stage of compression uses the data occupancy information of the root node to compress the data in the voxel corresponding to the root node.
  • the data occupancy information of the root node may be obtained by using a data compression model for context prediction, or it may be preset data occupancy information for the data compression of the root node according to different data compression requirements (for example, the root node the probability distribution of the possession bytes).
  • the encoding method adopted by the entropy encoder is variable-length encoding as an example. If the tree structure is an octree, the data occupation information of the root node is shown in Table 1 below, and the occupation status byte of the root node is "0001 1000 The predicted probability of ” is 89%, and the compression of the data to be compressed includes the following steps: 1, sort the occupancy bytes according to the predicted probability from large to small; 2, group the occupancy bytes corresponding to the two minimum probabilities into a group It is divided into 2 branch fields and marked as "0" and "1" respectively.
  • Second-level compression the entropy encoder compresses the data in the voxels corresponding to the grandfather node 732 by using the data occupancy information of the grandfather node 732 .
  • Third-level compression the entropy encoder compresses the data in the voxel corresponding to the parent node 733 by using the data occupancy information of the parent node 733 .
  • the entropy encoder compresses the data in the voxel corresponding to the node to be encoded 734 by using the data occupancy information of the node to be encoded 734 .
  • the entropy encoding method used in the second-level compression to the fourth-level compression is the same as the entropy encoding method used in the first-level compression, which will not be repeated here.
  • the entropy coding method used in the first-stage compression to the fourth-stage compression is described by taking variable-length coding (Huffman coding) as an example, and the data compression process in the data processing method provided by the present application can also use Arithmetic coding mode, the present application does not limit the entropy coding mode used for data compression.
  • the point cloud is octreeized to obtain 4-level nodes as an example for illustration, but when the data volume of the point cloud is smaller, the point cloud can be compressed by using fewer layers of nodes; When the amount of data is larger, more layers of nodes can be used to compress the point cloud, which is not limited in this application.
  • the sender uses an entropy encoder to perform multi-stage compression on the data to be compressed according to the data occupation information to obtain compressed data.
  • the present application uses a recurrent network layer to replace the multi-layer MLP network required for feature aggregation, which reduces the complexity of the data compression model and reduces the data compression required.
  • Computational resources improve the efficiency of data compression; in addition, compared with the prior art, the transmitting end needs to store n-1 sets of features, in the data processing method provided by this application, the transmitting end only needs to save the data occupied by the node to be encoded. information, reducing the storage space consumption of the sender.
  • the sending end sends the compressed data to the receiving end.
  • the sender may send the compressed data to the receiver through a communication link.
  • the sender encapsulates the compressed data into a suitable format such as a message, and/or uses any type of transfer encoding or processing to process the compressed data for transmission over a communication link or communication network to transmit.
  • the data processing method provided by the present application does not need to store the features and the intermediate features, reduces the storage space occupied by data compression, and reduces the transmission end. and the amount of compressed data transmitted by the receiving end, reducing the delay in transmitting the point cloud.
  • the receiving end uses the data compression model to determine the data occupation information in the tree structure.
  • the data occupation information is used to indicate the data distribution of the compressed data in the tree structure.
  • the data compression model and the cyclic network layer please refer to the above-mentioned relevant description about S520, which will not be repeated here.
  • the sender can send the predicted data occupancy information together with the occupancy bytes of the parent node into the entropy encoder, reducing the number of bits required to record the data of the actual occupancy bytes, thereby reducing The space required to store the data to be compressed to achieve the effect of data compression.
  • the receiver in order to restore the tree-like structure from the compressed data, the receiver can use the same context prediction method as in the data compression process to achieve the effect of data decompression.
  • the process of context prediction is described by taking the transmitting end implementing data compression as an example.
  • the context prediction method in the data decompression process is the same as the data compression process, and will not be repeated here.
  • the receiving end decompresses the compressed data according to the data possession information to obtain decompressed data.
  • the entropy decoder in the receiving end obtains the decompressed data by decompressing the compressed data step by step with multi-level nodes.
  • the great-grandfather node 731 with a depth of i-3 is the root node, and the point cloud decompression process includes the following multi-level decompression steps.
  • the first stage of decompression the entropy decoder decompresses the compressed data in the voxel corresponding to the root node by using the data occupancy information of the root node.
  • the data occupancy information of the root node may be obtained by using a data compression model for context prediction, or it may be preset data occupancy information (eg, according to different data compression requirements) for data compression and decompression of the root node. Probability distribution of the occupancy bytes of the root node).
  • the encoding method used by the entropy decoder is variable-length encoding as an example. If the tree structure is an octree, the data occupation information of the root node is shown in Table 2 below, and the occupation status byte of the root node is "0001 1000 The predicted probability of ” is 89%, and the decompression of the compressed data includes the following steps: 1. Sort the occupancy bytes according to the predicted probability from large to small; 2. Group the occupancy bytes corresponding to the two minimum probabilities into a group It is divided into 2 branch fields and marked as "0" and "1" respectively.
  • Second-level decompression the entropy decoder decompresses the data in the voxel corresponding to the grandfather node 732 by using the data occupancy information of the grandfather node 732 .
  • the entropy decoder decompresses the data in the voxel corresponding to the parent node 733 by using the data occupancy information of the parent node 733 .
  • the entropy decoder decompresses the data in the voxel corresponding to the node to be encoded 734 by using the data occupancy information of the node to be encoded 734 .
  • the coding modes used in the second-stage decompression to the fourth-stage decompression are the same as the coding modes used in the first-stage decompression, and are not repeated here.
  • the coding methods used in the first-stage decompression to the fourth-stage decompression are described by taking variable-length coding (Huffman coding) as an example, and the data decompression process in the data processing method provided by the present application can also use arithmetic.
  • Encoding mode this application does not limit the encoding mode used for data decompression, but the encoding mode used for data decompression and data compression should be consistent.
  • Figures 6 and 7 take the example of obtaining 4-level nodes after the octree is reconstructed from compressed data. However, when the amount of compressed data is smaller, nodes with fewer layers can be used to decompress the compressed data; When the amount of data is larger, more layers of nodes can be used to decompress the compressed data, which is not limited in this application.
  • a recurrent network layer is used to replace the multi-layer MLP network in the prior art, which reduces the complexity of the data compression model and reduces the computing resources required for context prediction to obtain data occupancy information; in addition, compared with the existing In the technology, the features of each node and the intermediate features obtained by fusing the features are stored.
  • the data processing method provided by the present application does not need to store the features and the intermediate features, thereby reducing the storage space occupied by data decompression.
  • the data processing method further includes the following steps.
  • the sending end displays at least one of tree structure and data occupation information.
  • the above-mentioned display unit may be a display screen.
  • the display screen may be a touch screen.
  • the display screen may be a head-up display (HUD) disposed inside the vehicle close to the driver's side, or the display screen may be a projection screen disposed inside the vehicle. projector's projection area.
  • HUD head-up display
  • the sender can display the tree structure of the data to be compressed, which indicates the data distribution of the i-1 layer nodes and the i layer nodes.
  • the sender can display data occupancy information. As shown in (b) of Figure 10, it indicates the predicted probability of the occupancy status byte of the i-1 layer node, for example, the predicted probability that the occupancy status byte of the i-1 layer node is "00000000" is 0.1%, the predicted probability that the occupancy byte of the i-1 layer node is "1000 0100" is 99%, and the predicted probability that the occupancy byte of the i-1 layer node is "1111 1111" is 0.05%.
  • the sender can simultaneously display the tree structure of the data to be compressed and the data occupancy information of the nodes to be encoded (eg, nodes in the i-1 layer).
  • the data possession information may be a set of data that cannot be identified by operators.
  • the data compression model here also includes the dimension Taking the adjustment layer as an example for description, please continue to refer to FIG. 5 , the data processing method may further include the following steps.
  • the sender inputs the data occupancy information of the nodes in the i-th layer into the dimension adjustment layer to obtain an occupancy rate prediction table.
  • the above-mentioned dimension adjustment layer may include at least one layer of MLP, and the MLP is used to adjust the output dimension of the data occupancy information.
  • the output dimension of the MLP may be 256, so as to output the predicted probability of each occupancy condition byte corresponding to "0000 0000" to "1111 1111".
  • the output dimension of the MLP can also be 260.
  • 4 dimensions can also be reserved as options, which can indicate The device type of the sender (such as mobile phone, computer), this option can also be used as the storage address of the compressed data to indicate the storage location of the compressed data, etc.
  • the occupancy prediction table indicates the predicted probability of each occupancy byte of the i-th tier node.
  • the occupancy rate prediction table can be as shown in Table 3.
  • the prediction probability that the occupancy status byte of the i-th layer node is "00" is 5%
  • the occupancy status of the i-th layer node is 5%.
  • the predicted probability that the byte is "01" is 80%
  • the predicted probability that the occupancy byte of the i-th layer node is "10” is 12%
  • the predicted probability that the occupancy byte of the i-th layer node is "11” is 3%.
  • the dimension adjustment layer is used to adjust the dimension of data occupation information to obtain a visualized occupation rate prediction table, which is helpful for operators to monitor the process of data compression or decompression, and there are obvious errors in the context prediction.
  • operators can use the occupancy prediction table as a reference to quickly determine the problems in the data compression or decompression process, and improve the robustness of the data compression and decompression process.
  • the present application provides a possible specific implementation manner for the above-mentioned data compression model.
  • the feature extraction layer includes three layers of MLP.
  • the recurrent network layer includes 3 layers of LSTM
  • the dimension adjustment layer includes 1 layer of MLP as an example to illustrate, as shown in FIG. 11,
  • FIG. 11 is a schematic structural diagram of another data compression model provided by the application, and the data compression model 1100 includes Feature extraction layer 1110, recurrent network layer 1120 and dimension adjustment layer 1130.
  • Feature extraction layer 1110 includes first MLP 1111 (input dimension is m, output dimension is 128), second MLP 1112 (input dimension is 128, output dimension is 128), third MLP 1113 (input dimension is 128, output dimension is 128) 128).
  • the feature extraction layer 1110 can implement the function of the fourth extraction layer 714 shown in FIG. 7 .
  • the feature extraction layer 1110 is used to obtain the occupancy status of the nodes in the i-1th layer according to the position, depth and sub-node number of the nodes in the i-th layer. bytes, and at least one of the occupancy status bytes of at least one sibling node of the i-th layer node, to obtain the feature X t of the i-th layer node.
  • the characteristics of the i-th layer node and the characteristic X t please refer to the relevant description of S5201, which will not be repeated here.
  • the recurrent network layer 1120 includes a first LSTM 1121 (input dimension 128, output dimension 128, stride 1 ⁇ 1), a second LSTM 1122 (input dimension 128, output dimension 128, stride 1 ⁇ 1) , the third LSTM 1123 (input dimension 128, output dimension 128, stride 1 ⁇ 1).
  • the recurrent network layer 1120 may implement the function of the fourth network layer 724 shown in FIG. 7 .
  • the recurrent network layer 1120 may occupy information h t-1 and the Summarize the information C t-1 and the feature X t of the i-th layer node, obtain the data occupancy information h t of the i-th layer node, and the summary information C t of the i-th layer node.
  • the dimension adjustment layer 1130 includes a fourth MLP 1131 (input dimension 128, output dimension 256).
  • the dimension adjustment layer 1130 may adjust the output dimension of the data occupancy information.
  • the fourth MLP 1131 adjusts the output dimension of the data occupancy information h t of the i-th layer node to 256 to obtain an occupancy rate prediction table.
  • the occupancy rate prediction table may include prediction probabilities of 256 occupancy status bytes such as "0000 0000" to "1111 1111".
  • the data processing method provided by this application requires a total of 4 MLP layers and 3 LSTM layers to predict the data occupancy information of the node to be encoded, and stores the summary of the outputs of the 3 LSTM layers.
  • Information and data occupy information, and the amount of computation and storage space required for data compression and decompression are constant.
  • the data processing method provided by the present application adopts the cyclic network layer to replace the multi-layer MLP layer, which reduces the complexity of the data compression model, and the computing resources occupied by the data compression model are all constant, so that the depth of the node to be encoded is reduced.
  • the context prediction of the nodes to be encoded will not occupy more computing resources; in addition, each layer of the depth of the nodes to be encoded increases the space required to store summary information and data occupancy information. Fixed, which reduces the storage space required for data compression and decompression.
  • the Oct Squeeze algorithm when it performs context prediction, it can only use the information of fixed n-1 ancestor nodes, while in the data processing method provided in this application, the recurrent network layer in the data compression model can be used from the root node to the All the prediction information of the parent node of the node to be encoded, and these prediction information can be selectively memorized and retained by the recurrent network layer, so that when the data compression model predicts the data occupancy information of the node to be encoded, the recurrent network layer can use the node to be encoded.
  • the information of all ancestor nodes improves the accuracy of context prediction and improves the data compression ratio.
  • the computing device includes corresponding hardware structures and/or software modules for performing each function.
  • the units and method steps of each example described in conjunction with the embodiments disclosed in the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is performed by hardware or computer software-driven hardware depends on the specific application scenarios and design constraints of the technical solution.
  • FIG. 12 is a schematic diagram of a data processing apparatus provided by this application. These data processing apparatuses can be used to implement the functions of the sending end and the receiving end in the above method embodiments, and thus can also achieve the beneficial effects of the above method embodiments.
  • the data processing apparatus may be the terminals 111 to 113 or the data center 130 as shown in FIG. 1 , and may also be a module (eg, a chip) applied to the application server 131 .
  • the structures and functions of the first data processing device 1210 and the second data processing device 1220 are described below with reference to FIG. 12 .
  • the first data processing device 1210 can implement the function of the sending end shown in FIG. 2
  • the second data processing device 1220 can The function of the receiving end shown in FIG. 2 is realized. It should be understood that this embodiment only exemplarily divides the structures and functional modules of the first data processing apparatus 1210 and the second data processing apparatus 1220, and this application does not make any specific divisions.
  • the first data processing device 1210 establishes a communication connection with the second data processing device 1220 through a communication channel 1230.
  • the communication channel 1230 can transmit the compressed data sent by the sender to the receiver, and the communication channel 1230 can also transmit and receive the data. Other information sent by the terminal to the sender.
  • the first data processing apparatus 1210 includes a collection unit 1211 , a display unit 1212 , a preprocessing unit 1213 , a context prediction unit 1214 , an encoding unit 1215 and a transceiver unit 1216 , and the above units can be used to implement the above-mentioned FIG. 5 or FIG. In the method embodiment shown in 8, the method corresponding to each operation step performed by the transmitting end.
  • the display unit 1212 is used to implement S570
  • the preprocessing unit 1213 is used to implement S510
  • the context prediction unit 1214 is used to implement S520 and S580
  • the encoding unit 1215 is used to implement S530
  • the transceiver unit 1216 is used to implement S540.
  • the context prediction unit 1214 is also used to implement S5201 to S5203.
  • the acquisition unit 1211 may implement the functions implemented by the sensors (such as lidar, millimeter-wave radar, and sonar, etc.) and cameras provided in the foregoing embodiments.
  • the display unit 1212 may include a display screen.
  • the display screen may be a touch screen.
  • the display screen may be a HUD or the like.
  • the second data processing apparatus 1220 includes an acquisition unit 1221 , a context prediction unit 1222 and a decompression unit 1223 , and the above units can be used to implement the method performed by the receiving end in the method embodiment shown in FIG. 5 or FIG. 8 . Methods corresponding to each operation step.
  • the obtaining unit 1221 is used to obtain compressed data
  • the context prediction unit 1222 is used to execute S550
  • the decompression unit 1223 is used to execute S550. S560.
  • first data processing apparatus 1210 and the second data processing apparatus 1220 can be obtained directly by referring to the relevant descriptions in the method embodiments shown in FIG. 5 or FIG. 8 , and details are not repeated here.
  • FIG. 13 is a schematic structural diagram of a computing device provided by this application.
  • the computing device 1300 includes a processor 1310 and a communication interface 1320 .
  • the processor 1310 and the communication interface 1320 are coupled to each other.
  • the communication interface 1320 may be a transceiver or an input-output interface.
  • the computing device 1300 may further include a memory 1330 for storing instructions executed by the processor 1310 or input data required by the processor 1310 to execute the instructions or data generated after the processor 1310 executes the instructions.
  • the processor 1310 may generate data to be compressed in a tree-like structure according to the original data, and use the recurrent network layer included in the data compression model to determine the data occupancy information in the tree-like structure.
  • the data occupation information is used to indicate the data distribution of the original data in the tree structure. Further, the processor 1310 compresses the above-mentioned data to be compressed according to the data occupation information to obtain compressed data.
  • the processor 1310 , the communication interface 1320 and the memory 1330 may also cooperate to implement various operation steps in the data processing method performed by the sender and the receiver.
  • the computing device 1300 may also perform the functions of the first data processing apparatus 1210 and the second data processing apparatus 1220 shown in FIG. 12 , which will not be repeated here.
  • the specific connection medium between the communication interface 1320 , the processor 1310 , and the memory 1330 is not limited in the embodiments of the present application.
  • the communication interface 1320, the processor 1310, and the memory 1330 are connected through a bus 1340 in FIG. 13.
  • the bus is represented by a thick line in FIG. 13, and the connection between other components is only for schematic illustration. , is not limited.
  • the bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of presentation, only one thick line is used in FIG. 13, but it does not mean that there is only one bus or one type of bus.
  • the memory 1330 can be used to store software programs and modules, such as program instructions/modules corresponding to the data processing methods provided in the embodiments of the present application.
  • the processor 1310 executes various functional applications by executing the software programs and modules stored in the memory 1330. and data processing.
  • the communication interface 1320 can be used for signaling or data communication with other devices.
  • the computing device 1300 may have multiple communication interfaces 1320 in this application.
  • the above-mentioned memory may be, but not limited to, random access memory (Random Access Memory, RAM), read-only memory (Read Only Memory, ROM), programmable read-only memory (Programmable Read-Only Memory, PROM), can Erasable Read-Only Memory (Erasable Programmable Read-Only Memory, EPROM), Electrical Erasable Programmable Read-Only Memory (EEPROM), etc.
  • RAM Random Access Memory
  • ROM Read Only Memory
  • PROM programmable read-only memory
  • EPROM Erasable Programmable Read-Only Memory
  • EEPROM Electrical Erasable Programmable Read-Only Memory
  • the above-mentioned processor may be an integrated circuit chip with signal processing capability.
  • the processor may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; it may also be a digital signal processor (Digital Signal Processing, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • CPU Central Processing Unit
  • NP Network Processor
  • DSP Digital Signal Processing
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • the method steps in the embodiments of the present application may be implemented in a hardware manner, or may be implemented in a manner in which a processor executes software instructions.
  • the software instructions can be composed of corresponding software modules, and the software modules can be stored in RAM, flash memory, ROM, PROM, EPROM, EEPROM, registers, hard disk, removable hard disk, CD-ROM or any other form of storage medium well known in the art .
  • An exemplary storage medium is coupled to the processor, such that the processor can read information from, and write information to, the storage medium.
  • the storage medium can also be an integral part of the processor.
  • the processor and storage medium may reside in an ASIC. Additionally, the ASIC may reside in a computing device.
  • the processor and storage medium may also exist in the computing device as discrete components.
  • the above-mentioned embodiments it may be implemented in whole or in part by software, hardware, firmware or any combination thereof.
  • software it can be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer programs or instructions.
  • the processes or functions described in the embodiments of the present application are executed in whole or in part.
  • the computer may be a general purpose computer, special purpose computer, computer network, communication device, user equipment, or other programmable device.
  • the computer program or instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer program or instructions may be downloaded from a website site, computer, A server or data center transmits by wire or wireless to another website site, computer, server or data center.
  • the computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server, data center, or the like that integrates one or more available media.
  • the usable medium can be a magnetic medium, such as a floppy disk, a hard disk, and a magnetic tape; it can also be an optical medium, such as a digital video disc (DVD); it can also be a semiconductor medium, such as a solid state drive (solid state drive). , SSD).
  • a magnetic medium such as a floppy disk, a hard disk, and a magnetic tape
  • an optical medium such as a digital video disc (DVD)
  • DVD digital video disc
  • it can also be a semiconductor medium, such as a solid state drive (solid state drive). , SSD).
  • “at least one” means one or more, and “plurality” means two or more.
  • “And/or”, which describes the relationship of the associated objects, indicates that there can be three kinds of relationships, for example, A and/or B, it can indicate that A exists alone, A and B exist at the same time, and B exists alone, where A, B can be singular or plural.
  • the character “/” generally indicates that the related objects are a kind of "or” relationship; in the formula of this application, the character "/” indicates that the related objects are a kind of "division” Relationship.
  • words such as “exemplary” or “for example” are used to represent examples, illustrations or illustrations. Any embodiments or designs described in the embodiments of the present application as “exemplary” or “such as” should not be construed as preferred or advantageous over other embodiments or designs. Rather, use of words such as “exemplary” or “such as” is intended to present the related concepts in a specific manner.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本申请提供了一种数据处理方法及装置，涉及数据处理领域。该方法包括：发送端根据原始数据生成树状结构的待压缩数据，并利用数据压缩模型包含的循环网络层确定原始数据在树状结构的数据占有信息。上述的数据占有信息用于指示原始数据在树状结构中的数据分布。进而，发送端根据该数据占有信息压缩上述待压缩数据得到压缩数据。接收端利用数据压缩模型包含的循环网络层确定压缩数据在树状结构的数据占有信息。根据数据占有信息解压压缩数据，得到解压数据。本申请采用一个循环网络层替代了现有技术中多层MLP网络进行上下文预测，降低了数据压缩模型的复杂度，减少了上下文预测获取数据占有信息时所需的计算资源。

Description

一种数据处理方法及装置
本申请要求于2021年04月09日提交国家知识产权局、申请号为202110384626.7、申请名称为“一种数据处理方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及数据处理领域，尤其涉及一种数据处理方法及装置。
背景技术
在获取物体表面每个采样点的空间坐标后，得到的一个点的集合，称之为点云(Point Cloud)。例如，在虚拟现实/增强现实(virtual reality/augmented reality，VR/AR)领域，可以利用点云展现数字人和虚拟物体；又如，在自动驾驶领域，利用点云可以模拟参考物，以实现车辆的精准定位和导航。通常，点云的数据量很大，为减少存储点云所占据的存储空间，会对点云进行压缩后再进行存储。
通常，基于八叉树的压缩(Octree Squeeze)算法对点云进行八叉树化后进行上下文预测，并依据上下文预测的结果对点云进行压缩。然而，Oct Squeeze算法采用的上下文预测过程是：利用多层感知机(multi-layer perceptron，MLP)进行特征的逐级融合，得到上下文预测结果。每级特征融合都会使用至少一层MLP，这会占用点云压缩设备大量的计算资源，且点云压缩耗时较长。因此，如何降低点云压缩的模型复杂度以及所需的计算资源是目前亟需解决的问题。
发明内容
本申请提供一种数据处理方法及装置，解决了现有技术中点云压缩的模型复杂度较高以及占用较大的计算资源的问题。
为达到上述目的，本申请采用如下技术方案：
第一方面，本申请提供了一种数据处理方法，该方法可应用于发送端，或者该方法可应用于可以支持计算设备实现该方法的装置，例如该装置包括芯片系统，该方法包括：发送端根据原始数据生成树状结构的待压缩数据，并利用数据压缩模型包含的循环网络层确定在树状结构的数据占有信息。上述的数据占有信息用于指示原始数据在树状结构中的数据分布。进而，发送端根据该数据占有信息压缩上述的待压缩数据得到压缩数据。如此，本申请采用一个循环网络层替代了现有技术中多层MLP网络进行上下文预测，降低了数据压缩模型的复杂度，减少了上下文预测获取数据占有信息时所需的计算资源；另外，相较于现有技术中会存储每个节点的特征以及融合特征得到的中间特征，本申请提供的数据处理方法无需存储特征及中间特征，减少了数据压缩所占据的存储空间。
在一种示例中，在利用数据压缩模型确定在树状结构的数据占有信息之前，该数据处理方法还可以包括：通过传感器采集原始数据，原始数据为三维数据。例如，该传感器包括激光雷达、毫米波雷达和声呐中至少一种。
在另一种示例中，在利用数据压缩模型确定在树状结构的数据占有信息之前，该数据处理方法还可以包括：通过摄像头采集原始数据，原始数据为二维数据。
作为一种可选的实现方式，该数据处理方法还包括：显示树状结构和数据占有信息中至少一种。在数据压缩的过程比较长的情况下，通过对树状结构和数据占有信息中至少一种进行显示，有利于用户监控数据压缩过程，定位数据压缩出现错误(如数据压缩出现卡顿或停止)的位置。
作为一种可选的实现方式，利用数据压缩模型确定在树状结构的数据占有信息，包括：发送端将树状结构中第i-1层节点的数据占有信息输入循环网络层，得到第i层节点的数据占有信息，i为正整数，第i-1层节点是第i层节点的父节点。相较于现有技术中点云压缩设备会对每个特征以及融合得到的特征进行存储，本申请提供的数据处理方法在对待编码节点进行上下文预测时，可以利用待编码节点的父节点的数据占有信息来获取待编码节点的数据占有信息，无需对待编码节点的特征和其父节点的特征进行特征融合，这减少了MLP网络的使用，降低了数据压缩模型的复杂度，减少了上下文预测获取数据占有信息时所需的计算资源。
作为一种可选的实现方式，利用数据压缩模型确定在树状结构的数据占有信息，包括：发送端将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的数据占有信息，第i-1层节点的总结信息用于描述第i-1层节点的祖先节点到第i-1层节点的所有预测信息。由于循环网络层不需要使用对待编码节点的特征和其父节点的特征进行融合得到的中间特征，因此发送端无需存储待编码节点的特征、父节点的特征以及中间特征，这减少了发送端进行上下文预测所需的存储空间。此外，相较于现有技术中点云压缩设备对每个节点进行上下文预测都需要从根节点开始，本申请提供的数据处理方法中，循环网络层可以利用待编码节点的特征以及待编码节点的父节点提炼的总结信息，得到待编码节点的数据占有信息以及总结信息，使得本申请提供的数据处理方法无需从根节点开始，这减少了上下文预测所需的计算资源。
作为一种可选的实现方式，该数据处理方法还包括：发送端将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的总结信息，第i层节点的总结信息用于描述第i层节点的祖先节点到第i层节点的所有预测信息。第i层节点的总结信息可以是通过迭代的方式获取的，例如，若循环网络层包括至少一层长短期记忆网络(long short-term memory，LSTM)层，LSTM层的双曲正切函数(tanh)和门结构(σ函数)可以用于选择性的利用第i层节点的前文信息，得到待编码节点的数据占有信息和总结信息。在本申请提供的数据处理方法中，数据压缩模型中的循环网络层可以利用从根节点开始到待编码节点的父节点的所有预测信息，且这些预测信息可以被循环网络层选择性的记忆和保留，使得在数据压缩模型预测待编码节点的数据占有信息时，循环网络层可以利用待编码节点的所有祖先节点的信息，提高了上下文预测的精度，利用本申请得到的数据占有信息对待压缩数据进行压缩，提高了数据压缩比。
作为一种可选的实现方式，数据压缩模型还包括特征提取层，该数据处理方法还包括：发送端将第i层节点的位置、深度和子节点编号，以及第i-1层节点的占有情况字节中至少一种输入特征提取层，得到第i层节点的特征。例如，上述的特征提取层包括至少一层MLP。
作为一种可能的示例，发送端将第i层节点的位置、深度和子节点编号，第i-1层节点 的占有情况字节，以及第i层节点的至少一个兄弟节点的占有情况字节输入特征提取层，得到第i层节点的特征。第i层节点的兄弟节点是指与第i层节点同属一个父节点的其他节点，由于第i层节点的特征包括其父节点和兄弟节点的占有情况字节，使得发送端在对该第i层节点进行上下文预测时，降低了一些明显不可能的第i层节点的占有情况字节的预测概率，提高了第i层节点的数据占有信息的预测准确率。
作为一种可选的实现方式，数据压缩模型还包括维度调整层，该数据处理方法还包括：发送端将第i层节点的数据占有信息输入维度调整层，得到占有率预测表，该占有率预测表指示第i层节点的每个占有情况字节的预测概率。例如，维度调整层包括至少一层MLP，该MLP可以用于调整数据占有信息的输出维度，得到可视化的预测概率结果。
第二方面，本申请提供一种数据处理方法，该方法可应用于接收端，或者该方法可应用于可以实现该方法的计算设备，例如该计算设备包括芯片系统，该方法包括：接收端获取压缩数据，并利用数据压缩模型包含的循环网络层确定在树状结构的数据占有信息，该数据占有信息用于指示压缩数据在树状结构中的数据分布。接收端还根据数据占有信息解压压缩数据，得到解压数据。本申请采用一个循环网络层替代了现有技术中多层MLP网络进行上下文预测，减少了上下文预测获取数据占有信息时所需的计算资源；另外，相较于现有技术中会存储每个节点的特征以及融合特征得到的中间特征，本申请提供的数据处理方法无需存储特征及中间特征，减少了数据压缩所占据的存储空间。
在一种可能的示例中，循环网络层可以包括至少一层LSTM层。
作为一种可选的实现方式，利用数据压缩模型确定在树状结构的数据占有信息，包括：接收端将树状结构中第i-1层节点的数据占有信息输入循环网络层，得到第i层节点的数据占有信息，i为正整数，第i-1层节点是第i层节点的父节点。
作为一种可选的实现方式，接收端利用数据压缩模型确定在树状结构的数据占有信息，包括：接收端将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的数据占有信息，第i-1层节点的总结信息用于描述第i-1层节点的祖先节点到第i-1层节点的所有预测信息。
作为一种可选的实现方式，该数据处理方法还包括：接收端将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的总结信息，第i层节点的总结信息用于描述第i层节点的祖先节点到第i层节点的所有预测信息。
作为一种可选的实现方式，数据压缩模型还包括特征提取层，该数据处理方法还包括：接收端将第i层节点的位置、深度和子节点编号，以及第i-1层节点的占有情况字节中至少一种输入特征提取层，得到第i层节点的特征。例如，特征提取层包括至少一层MLP。
在一种可能的示例中，接收端将第i层节点的位置、深度和子节点编号，以及第i-1层节点的占有情况字节中至少一种输入特征提取层，得到第i层节点的特征，包括：接收端将第i层节点的位置、深度和子节点编号，第i-1层节点的占有情况字节，以及第i层节点的至少一个兄弟节点的占有情况字节输入特征提取层，得到第i层节点的特征。
作为一种可选的实现方式，数据压缩模型还包括维度调整层，该数据处理方法还包括：接收端将第i层节点的数据占有信息输入维度调整层，得到占有率预测表，占有率预测表指示第i层节点的每个占有情况字节的预测概率。例如，维度调整层包括至少一层MLP。
第三方面，本申请提供一种数据处理装置，有益效果可以参见第一方面中任一方面的描述，此处不再赘述。所述数据处理装置具有实现上述第一方面中任一方面的方法实例中行为的功能。所述功能可以通过硬件实现，也可以通过硬件执行相应的软件实现。所述硬件或软件包括一个或多个与上述功能相对应的模块。在一个可能的设计中，该数据处理装置应用于发送端，该数据处理装置包括：预处理单元，用于根据原始数据生成树状结构的待压缩数据；上下文预测单元，用于利用数据压缩模型确定在树状结构的数据占有信息，数据占有信息用于指示原始数据在树状结构中的数据分布，数据压缩模型包含循环网络层，循环网络层用于确定数据占有信息；编码单元，用于根据数据占有信息压缩待压缩数据，得到压缩数据。
结合第三方面提供的数据处理装置，作为一种可选的实现方式，该数据处理装置还包括：采集单元，用于通过传感器采集原始数据，原始数据为三维数据，传感器包括激光雷达、毫米波雷达和声呐中至少一种。
结合第三方面提供的数据处理装置，作为一种可选的实现方式，该数据处理装置还包括：采集单元，用于通过摄像头采集原始数据，原始数据为二维数据。
结合第三方面提供的数据处理装置，作为一种可选的实现方式，该数据处理装置还包括：显示单元，用于显示树状结构和/或数据占有信息。
结合第三方面提供的数据处理装置，作为一种可选的实现方式，上下文预测单元具体用于将树状结构中第i-1层节点的数据占有信息输入循环网络层，得到第i层节点的数据占有信息，i为正整数，第i-1层节点是第i层节点的父节点。例如，循环网络层包括至少一层LSTM层。
结合第三方面提供的数据处理装置，作为一种可选的实现方式，上下文预测单元具体用于将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的数据占有信息，第i-1层节点的总结信息用于描述第i-1层节点的祖先节点到第i-1层节点的所有预测信息。
结合第三方面提供的数据处理装置，作为一种可选的实现方式，上下文预测单元还用于将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的总结信息，第i层节点的总结信息用于描述第i层节点的祖先节点到第i层节点的所有预测信息。
结合第三方面提供的数据处理装置，作为一种可选的实现方式，数据压缩模型还包括特征提取层，该上下文预测单元还用于将第i层节点的位置、深度和子节点编号，以及第i-1层节点的占有情况字节中至少一种输入特征提取层，得到第i层节点的特征。例如，特征提取层包括至少一层MLP。
在一种可能的示例中，该上下文预测单元还用于将第i层节点的位置、深度和子节点编号，第i-1层节点的占有情况字节，以及第i层节点的至少一个兄弟节点的占有情况字节输入特征提取层，得到第i层节点的特征。
结合第三方面提供的数据处理装置，作为一种可选的实现方式，数据压缩模型还包括维度调整层，该上下文预测单元，还用于将第i层节点的数据占有信息输入维度调整层，得到占有率预测表，占有率预测表指示第i层节点的每个占有情况字节的预测概率。例如，维度调整层包括至少一层MLP。
第四方面，本申请提供一种数据处理装置，有益效果可以参见第二方面中任二方面的描述，此处不再赘述。所述数据处理装置具有实现上述第二方面中任二方面的方法实例中行为的功能。所述功能可以通过硬件实现，也可以通过硬件执行相应的软件实现。所述硬件或软件包括一个或多个与上述功能相对应的模块。在一个可能的设计中，该数据处理装置应用于接收端，该数据处理装置包括：获取单元，用于获取压缩数据；上下文预测单元，用于利用数据压缩模型确定在树状结构的数据占有信息，数据占有信息用于指示压缩数据在树状结构中的数据分布，数据压缩模型包含循环网络层，循环网络层用于确定数据占有信息；解压单元，用于根据数据占有信息解压压缩数据，得到解压数据。
结合第四方面提供的数据处理装置，作为一种可选的实现方式，上下文预测单元具体用于将树状结构中第i-1层节点的数据占有信息输入循环网络层，得到第i层节点的数据占有信息，i为正整数，第i-1层节点是第i层节点的父节点。例如，循环网络层包括至少一层LSTM层。
结合第四方面提供的数据处理装置，作为一种可选的实现方式，上下文预测单元具体用于将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的数据占有信息，第i-1层节点的总结信息用于描述第i-1层节点的祖先节点到第i-1层节点的所有预测信息。
结合第四方面提供的数据处理装置，作为一种可选的实现方式，上下文预测单元还用于将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的总结信息，第i层节点的总结信息用于描述第i层节点的祖先节点到第i层节点的所有预测信息。
结合第四方面提供的数据处理装置，作为一种可选的实现方式，数据压缩模型还包括特征提取层，该上下文预测单元还用于将第i层节点的位置、深度和子节点编号，以及第i-1层节点的占有情况字节中至少一种输入特征提取层，得到第i层节点的特征。例如，特征提取层包括至少一层MLP。
在一种可能的示例中，该上下文预测单元还用于将第i层节点的位置、深度和子节点编号，第i-1层节点的占有情况字节，以及第i层节点的至少一个兄弟节点的占有情况字节输入特征提取层，得到第i层节点的特征。
结合第四方面提供的数据处理装置，作为一种可选的实现方式，数据压缩模型还包括维度调整层，该上下文预测单元，还用于将第i层节点的数据占有信息输入维度调整层，得到占有率预测表，占有率预测表指示第i层节点的每个占有情况字节的预测概率。例如，维度调整层包括至少一层MLP。
第五方面，本申请提供一种计算设备，该计算设备包括至少一个处理器和存储器，存储器用于存储一组计算机指令；当处理器执行所述一组计算机指令时，执行第一方面或第一方面任一种可能实现方式，或第二方面和第二方面中任一种可能实现方式中的数据处理方法的操作步骤。
第六方面，本申请提供一种计算机可读存储介质，存储介质中存储有计算机程序或指令，当计算机程序或指令被计算设备执行时，实现第一方面和第一方面中任一种可能实现方式，或第二方面和第二方面中任一种可能实现方式的方法的操作步骤。
第七方面，本申请提供一种计算机程序产品，当计算机程序产品在计算机上运行时， 使得计算设备执行第一方面和第一方面中任一种可能实现方式，或第二方面和第二方面中任一种可能实现方式的方法的操作步骤。
第八方面，本申请提供一种芯片，包括存储器和处理器，存储器用于存储计算机指令，处理器用于从存储器中调用并运行该计算机指令，以执行上述第一方面及其第一方面任意可能的实现方式中的方法，或第二方面和第二方面中任一种可能实现方式的方法的操作步骤。
本申请在上述各方面提供的实现方式的基础上，还可以进行进一步组合以提供更多实现方式。
附图说明
图1为本申请提供的一种通信系统的场景示意图；
图2为本申请提供的一种点云压缩和解压的系统示意图；
图3为本申请提供的一种点云压缩和解压过程的示意图；
图4为现有技术提供的一种Oct Squeeze算法的上下文预测方法的示意图；
图5为本申请提供的一种数据处理方法的流程示意图；
图6为本申请提供的一种树状结构的示意图；
图7为本申请提供的一种数据压缩模型的示意图；
图8为本申请提供的另一种数据处理方法的流程示意图；
图9为本申请提供的一种LSTM的网络结构示意图；
图10为本申请提供的一种数据处理的显示示意图；
图11为本申请提供的另一种数据压缩模型的结构示意图；
图12为本申请提供的一种数据处理装置的示意图；
图13为本申请提供的一种计算设备的结构示意图。
具体实施方式
为了下述各实施例的描述清楚简洁，首先给出相关技术的简要介绍。
点云是点的数据集，点云中的点可以由三维坐标(X，Y，Z)进行表示位置，处于三维坐标(X，Y，Z)上的点可以包括颜色、分类值和强度值等属性信息。
通常，点云的数据量很大，存储点云会占据较大的存储空间。为了解决该问题，会对点云进行压缩后存储。请参见图1，图1为本申请提供的一种通信系统的场景示意图，该通信系统包括至少一个终端(如图1所示的终端111至终端113)、网络和数据中心130。终端和数据中心130可以通过网络进行通信，该网络可以是互联网络。
终端(terminal)也可以称为终端设备、用户设备(user equipment，UE)、移动台(mobile station，MS)、移动终端(mobile terminal，MT)等。
在一些实施例中，终端可以是手机(如图1中所示出的终端111)、平板电脑(如图1中所示出的终端112)、带无线收发功能的电脑(如图1中所示出的终端113)、虚拟现实(Virtual Reality，VR)终端设备、增强现实(Augmented Reality，AR)终端设备、工业控制(industrial control)中的无线终端、无人驾驶(self driving)中的无线终端(如集成在图1中所示出的车辆121和车辆122上的激光雷达)、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端等等。
在另一些实施例中，终端还可以是设置在住宅中的智慧家庭(smart home)终端，如智慧屏。终端还可以是设置在医院中，用于进行远程手术(remote medical surgery)的终端；例如，在进行无创手术时，该终端可以用于采集患者的身体内部信息(如人体内部器官的点云)。
数据中心130可以是包括至少一个应用服务器131的服务器集群，也可以是由应用服务器131构建的云数据中心。例如，多个应用服务器可以是独立的不同的物理设备，也可以是将多个应用服务器的功能集成在同一个物理设备上(如：云服务提供商管辖范围内的多个应用服务器)，还可以是一个物理设备上集成了部分应用服务器的功能。
终端通过无线或有线方式与应用服务器131连接。终端可以是固定位置的，也可以是可移动的。本申请的实施例对该通信系统中包括的终端和应用服务器的数量不做限定。
终端可以实现采集点云、压缩点云或解压点云等功能，数据中心130可以实现压缩点云或解压点云等功能。如图1所示，这里以车载终端可以采集点云和压缩点云，数据中心可以解压点云为例，车载终端可以包括安装在车辆122上的激光雷达以及与激光雷达通信连接的处理设备，例如，车辆122在直行时激光雷达采集路边的办公楼、住宅和篮球场的点云，在右转弯时激光雷达采集路旁的两侧植物(图1所示出的树)的点云，处理设备将上述的点云传输至数据中心130。然而，现在主流的64线激光雷达一小时收集的点云的数据量在100吉字节(GB)以上，这样的数据量对处理设备的存储和传输点云都是巨大的挑战。为了能降低传输点云的时延和存储点云所需的存储容量，高效的压缩算法不可或缺。
在目前的技术方案中，点云压缩路线可以大致分为两大类：传统点云压缩算法以及基于人工智能(artificial intelligence，AI)技术的点云压缩算法。
在一种示例中，传统点云压缩算法可以划分为两类：第一类是点云压缩设备将点云转化为二维信号，然后利用现有的图像压缩算法来进一步压缩，例如，动态图像专家组(Moving Pictures Experts Group，MPEG)提供的基于平面投影的压缩算法(MPEG V-PCC)。第二类是点云压缩设备将点云转化为树结构再进行熵编码，例如，基于kd-tree的Draco算法以及基于八叉树的基于几何的压缩算法(MPEG G-PCC)。
在本文中，点云压缩设备可以是对点云进行压缩的设备。例如，点云压缩设备可以是如图1所示的终端111至终端113中的任意一个。
在另一种示例中，基于AI的点云压缩算法可以划分为两类：一类是基于传统算法框架，并采用AI算法替代上下文编码模块的混合算法，如Oct Squeeze算法。另一类算法是端到端AI(全AI)压缩算法，如基于“Point net++”特征的自编码器算法。
总体上说，无论是传统的压缩算法，还是基于AI技术的压缩算法，都是点云压缩设备基于树结构对点云进行分解并压缩，得到点云压缩数据。如图2所示，图2为本申请提供的一种点云压缩和解压的系统示意图，该系统包括发送端210和接收端220，发送端210通过通信信道230与接收端220建立通信连接。
上述的发送端210可以实现数据压缩的功能，如图1所示，发送端210可以是终端111～终端113中的任意一个，发送端210还可以是设置在车辆121或车辆122上的点云压缩系统，例如，该点云压缩系统包括激光雷达以及与激光雷达通信连接的处理设备。
发送端210可以包括数据源211、预处理模块212、编码器213和通信接口214。
数据源211可包括或可以为任意类型的用于捕获点云的电子设备，和/或任意类型的点 数据生成设备，例如用于生成计算机动画场景的计算机图形处理器或任意类型的用于获取和/或提供现实世界点云、计算机生成点云的设备。数据源211可以为存储上述点云中的任意点数据的任意类型的内存或存储器。
为了区分预处理模块212执行的处理，点云241也可称为原始数据(或原始点云数据)241。
预处理模块212用于接收点云241，并对点云241进行预处理，得到预处理数据242。例如，预处理模块212执行的预处理可包括颜色格式转换(例如从RGB转换为YCbCr)、八叉树结构化等。
编码器213用于接收预处理数据242，并在进行上下文预测后压缩预处理数据242，得到点云压缩数据243。
发送端210中的通信接口214可用于：接收点云压缩数据243，并通过通信信道230向接收端220等另一设备或任何其它设备发送点云压缩数据243(或将该点云压缩数据243进行其它任意处理后的版本)，以便存储或直接重建。
上述的接收端220可以实现数据解压的功能，如图1所示，接收端220可以是图1所示出的数据中心130中任意一个或多个服务器，例如，应用服务器131实现接收端220的数据解压功能。
接收端220可以包括显示设备221、后处理模块222、解码器223和通信接口224。
接收端220中的通信接口224用于从发送端210或从存储设备等任意其它发送端接收点云压缩数据243(或其它任意处理后的版本)，例如，存储设备为点云数据存储设备，并将点云压缩数据243提供给解码器223。
通信接口214和通信接口224可用于通过发送端210和接收端220之间的直连通信链路，例如直接有线或无线连接等，或者通过任意类型的网络，例如有线网络、无线网络或其任意组合、任意类型的私网和公网或其任意类型的组合，发送或接收点云压缩数据243。
例如，通信接口214可用于将点云压缩数据243封装为报文等合适的格式，和/或使用任意类型的传输编码或处理来处理所述编码后的点云压缩数据243，以便在通信链路或通信网络上进行传输。
通信接口224与通信接口214对应，例如，可用于接收传输数据，并使用任意类型的对应传输解码或处理和/或解封装对传输数据进行处理，得到点云压缩数据243。
通信接口224与通信接口214均可配置为如图2中从发送端210指向接收端220的对应通信信道230的箭头所指示的单向通信接口，或双向通信接口，并且可用于发送和接收消息等，以建立连接，确认并交换与通信链路和/或例如编码后的压缩数据传输等数据传输相关的任何其它信息，等等。
解码器223用于接收点云压缩数据243，并对点云压缩数据243进行上下文预测后，得到解码数据244。
后处理模块222用于对解码后的解码数据244进行后处理，得到后处理数据245。后处理模块222执行的后处理可以包括例如颜色格式转换(例如从YCbCr转换为RGB)、八叉树重构等，或者用于产生供显示设备221等显示的数据等任何其它处理。
显示设备221用于接收后处理数据245，以向用户或观看者等显示。显示设备221可以为或包括任意类型的用于表示重建后图像的显示器，例如，集成或外部显示屏或显示器。 例如，显示屏可包括液晶显示器(liquid crystal display，LCD)、有机发光二极管(organic light emitting diode，OLED)显示器、等离子显示器、投影仪、微型LED显示器、硅基液晶显示器(liquid crystal on silicon，LCoS)、数字光处理器(digital light processor，DLP)或任意类型的其它显示屏。
作为一种可选的实施方式，发送端210和接收端220可以通过数据转发设备进行点云压缩数据243的传输。例如，数据转发设备可以是路由器或交换机等。
图3为本申请提供的一种点云压缩和解压过程的示意图，点云中的数据以坐标点的形式进行存储，如图3所示出的“[(x1,y1,z1),(x2,y2,z2)…]”。
在点云压缩过程中，八叉树化模块31可以实现图2所示出的预处理模块212的功能，例如，八叉树化模块31将点云进行八叉树(octree)化，得到待压缩数据，该待压缩数据包括根节点和8个子节点。如图3所示，八叉树是一种用于描述三维空间的树状数据结构，八叉树的每个节点表示一个正方体的体积元素(体素)，每个根节点有八个子节点(通过对体素进行前后，左右，上下的划分得到)，将八个子节点所表示的体素的体积加在一起就等于父节点所表示的体素的体积。
在使用八叉树表示点云的时候，八叉树化模块31可以依据子节点对应的体素中是否有数据，为子节点赋值“1”或“0”。例如，“1”指示子节点对应的体素中具有数据，“0”指示子节点对应的体素中无数据。八叉树化模块31整合八个子节点的值，并用8个比特表示，得到根节点的占有情况字节(occupancy byte)，该占有情况字节指示根节点的数据分布情况。如图3所示，黑色的节点指示该子节点对应的体素中有数据，白色的节点指示该子节点对应的体素中无数据，则根节点的占有情况字节为：1011 1100。
点云压缩过程中，上下文预测装置32和熵编码器33可以实现图2所示出的编码器213的功能，例如，在上下文预测装置32预测得到根节点的占用情况字节后，熵编码器33根据该占有情况字节对该根节点对应的点云进行压缩，以实现点云压缩，得到点云的压缩数据。熵编码器33和熵解码器35采用的熵编码算法可以是以下的任意一种或组合：香农(Shannon)编码、哈夫曼(Huffman)编码和算术编码(arithmetic coding)等，关于熵编码的具体实现方式请参考现有技术的相关阐述，不予赘述。
在点云解压过程中，上下文预测装置34和熵解码器35可以实现图2所示出的解码器223的功能，例如，使用与点云压缩相同的上下文预测模块34预测压缩数据中根节点的占有情况字节，熵解码器35和根据该根节点的占有情况字节解压该待解码节点对应的压缩数据，八叉树重构模块36对解压后的该待解码节点对应的压缩数据进行重构，待每一层节点都进行解压和重构之后，得到解压后的点云。
在点云压缩的过程中，熵编码器都会用到上下文预测模块预测得到的占有情况字节，该占有情况字节的预测结果越接近点云的真实数据分布情况，则熵编码器进行点云压缩的数据压缩比(点云与压缩数据的数据量之比)越大，点云压缩的效果越好。
在目前的技术方案中，这里以Oct Squeeze算法提供的上下文预测方法为例，Oct Squeeze算法将点云进行八叉树化后得到多层节点，假设该多层节点包括至少4层节点，如图4所示，图4为现有技术提供的一种Oct Squeeze算法的上下文预测方法的示意图，八叉树化后的点云包括曾祖父节点、祖父节点、父节点和待编码节点，前述节点的深度依次为：i-3、i-2、i-1和i，i为大于或等于3的正整数。
节点的深度表征节点从根节点到当前位置的层数，示例的，八叉树的根节点的深度为0，该根节点的子节点的深度为1。如图3所示，根节点可以称为第0层节点，子节点可以称为第1层节点。
图4所示出的上下文预测方法包括特征提取过程，示例的，利用MLP网络(如5层MLP layers)对每一层节点进行特征提取，如图4所示，MLP网络分别获取每一层节点的特征，该特征为高维度的特征向量。
图4所示出的上下文预测方法还包括特征融合过程，示例的，利用MLP网络(如4层MLP layers)对每一层节点的特征进行波网式(WaveNet)的逐级融合，如图4所示，MLP网络将待编码节点(如第i-1层节点)的特征和其父节点(如第i-2层节点)的特征进行融合，在经过多级的特征融合后，获得该待编码节点的占有率预测结果。该占有率预测结果指示待编码节点对应的体素中点的数据分布情况。关于WaveNet的原理可以参考现有技术的相关阐述，本申请不予赘述。
因此，由于每一层节点的特征提取都需要设置一个MLP网络，每两个相邻特征的融合也需要单独设置一个MLP网络，导致占据点云压缩设备利用大量的计算资源进行MLP网络的相关计算，浪费了计算资源。此外，在待编码节点的深度增加的情况下，点云压缩设备会使用MLP网络对已经提取的特征进行重复提取，这会增加重复的计算过程；为了减少上下文预测所需的计算资源，点云压缩设备会对每个特征以及融合得到的特征进行存储，然而，这又会占用点云压缩设备大量的存储空间。
为了减少上下文预测所占据的计算资源和存储空间，本申请提供一种数据处理方法，其包括：发送端根据原始数据生成树状结构的待压缩数据，并利用数据压缩模型包括的循环网络层确定在树状结构的数据占有信息，该数据占有信息用于指示原始数据在树状结构中的数据分布，发送端根据该数据占有信息压缩上述的待压缩数据得到压缩数据。本申请采用一个循环网络层替代了现有技术中多层MLP网络进行上下文预测，降低了数据压缩模型的复杂度，减少了上下文预测获取数据占有信息时所需的计算资源；另外，相较于现有技术中会存储每个节点的特征以及融合特征得到的中间特征，本申请提供的数据处理方法无需存储特征及中间特征，减少了数据压缩所占据的存储空间。
这里以图2所示出的发送端210可以实现数据压缩，接收端220可以实现数据解压为例进行说明，如图5所示，图5为本申请提供的一种数据处理方法的流程示意图，该数据处理方法包括以下步骤。
S510、发送端根据原始数据生成树状结构的待压缩数据。
该发送端可以是部署在车辆上的移动终端(例如，手机、平板电脑等)，该发送端还可以是搭载在车辆上的点云压缩系统，例如，该点云压缩系统包括激光雷达以及与激光雷达通信连接的处理设备。
在第一种可能的情形中，该原始数据可以是三维数据，例如，点云。
该三维数据可以是传感器采集的。例如，该传感器可以包括激光雷达、毫米波雷达和声呐中至少一种，该传感器可以集成在发送端上，也可以单独部署在与发送端同属的载具(如图1所示出的车辆122)上，并与发送端建立了通信连接。又如，该传感器还可以是传感器阵列，该传感器阵列中的各个传感器可以分别部署在车辆122和待采集的参考物上(如图1所示出的树)。
上述的点云可以包括多种信息。例如，若点云是激光测量设备采集的，其包括三维坐标和激光反射强度。又如，若点云是摄影测量设备采集的，其可以包括三维坐标和颜色信息。又如，若点云是激光测量设备和摄影测量设备共同采集并合并后得到的，其可以包括三维坐标、激光反射强度和颜色信息。
在第二种可能的情形中，该原始数据可以是二维数据，例如，图像数据。
该二维数据可以是摄像头采集的。该摄像头可以集成在发送端上，也可以部署在车辆122上，并与发送端建立了通信连接。示例的，该摄像头可以包括一个或多个相机，例如，该摄像头可以设置在车辆122的前挡风玻璃、后挡风玻璃、车顶、和车身等位置中的一个或多个位置。
上述的树状结构可以是M叉树，M为大于或等于2的正整数，例如，八叉树，四叉树或二叉树等。
S520、发送端利用数据压缩模型确定在树状结构的数据占有信息。
上述的数据占有信息用于指示原始数据在树状结构中的数据分布。如图2所示出的点云对应的体素，数据占有信息可以指示该体素中的点的分布情况，例如，该数据占有信息指示待压缩数据中根节点的占有情况字节为“1000 1000”的预测概率为99％。
在第一种示例中，如图6所示，假设i＝3，则深度为i-3的节点为根节点，如图6中的(a)所示，根节点具有8个子节点，该8个子节点的编号为1～8。根节点的数据占有信息如图6中的(b)所示，该数据占有信息指示根节点对应的体素中序号为“1”和“5”的两个子节点对应的体素中具有数据。
在第二种示例中，对图6中的(b)所示出的序号为“1”的深度为i-2的节点可以进行上下文预测，得到的数据占有信息如图6中的(c)所示，该数据占有信息指示该深度为i-2的节点对应的体素中序号为“1”和“5”的两个深度为i-1的节点对应的体素中具有数据。
在第三种示例中，对图6中的(b)所示出的序号为“5”的深度为i-2的节点可以进行上下文预测，得到的数据占有信息如图6中的(d)所示，该数据占有信息指示该深度为i-2的节点对应的体素中序号为“3”的深度为i-1的节点对应的体素中具有数据。
在第四种示例中，对图6中的(c)所示出的序号为“5”的深度为i-1的节点可以进行上下文预测，得到的数据占有信息如图6中的(e)所示，该数据占有信息指示该深度为i的节点对应的体素中序号为“1”和“5”的两个深度为i的节点对应的体素中具有数据。
上述的数据压缩模型包含循环网络层，该循环网络层用于确定该数据占有信息。该循环网络层可以包括至少一层卷积神经网络层，例如，循环神经网络(recurrent neural network，RNN)，LSTM层，门控循环单元(gated recurrent unit，GRU)等。作为一种可选的实施方式，该循环网络层可以包括以上多种卷积神经网络层中的任意组合。
数据压缩模型还包含特征提取层。特征提取层用于依据待编码节点的位置、深度和子节点编号，以及待编码节点的父节点的占有情况字节中至少一种，确定待编码节点的特征，该特征为高维度的特征向量。特征提取层的层数和循环网络层的层数可以是依据节点的深度决定的。
在一种示例中，图7为本申请提供的一种数据压缩模型的示意图，该数据压缩模型700包括特征提取层710和循环网络层720。
特征提取层710包括第一提取层711、第二提取层712、第三提取层713和第四提取层 714，示例的，上述的第一提取层711～第四提取层714均可以是多层MLP。
在一种可能的情形中，上述的第一提取层711～第四提取层714可以是不同的MLP网络，示例的，针对于深度不同的节点，特征提取层包括的MLP的层数不同。例如，第二提取层712包括3层MLP，第三提取层713包括4层MLP。
在另一种可能的情形中，上述的第一提取层711～第四提取层714可以是相同的MLP网络。示例的，针对于深度不同的节点，提取层是可以复用的。例如，第一提取层711～第四提取层714均包括4层相同的MLP。
循环网络层720包括第一网络层721、第二网络层722、第三网络层723和第四网络层724。示例的，上述的第一网络层721～第四网络层724均可以是LSTM。例如，上述的第一网络层721～第四网络层724均包括3层相同的LSTM层。
作为一种可选的实现方式，发送端可以将树状结构中第i-1层节点的数据占有信息输入循环网络层，得到第i层节点的数据占有信息。
其中，i为正整数，第i-1层节点是第i层节点的父节点。
这里以i≥3为例进行说明，在本文中，曾祖父节点731是图6所示出的深度为i-3的节点，祖父节点732是图6示出的深度为i-2的第一个黑色节点，父节点733是图6示出的深度为i-1的第二个黑色节点，待编码节点734是图6示出的深度为i的第一个黑色节点。
在一种可能的示例中，父节点733的数据占有信息如图6中的(e)所示，例如，该数据占有信息指示父节点733对应的体素中序号为“1”和“5”的两个子节点对应的体素中具有数据的概率为98％。
相较于现有技术中点云压缩设备会对每个特征以及融合得到的特征进行存储，本申请提供的数据处理方法在对待编码节点进行上下文预测时，可以利用待编码节点的父节点的数据占有信息来获取待编码节点的数据占有信息，无需对待编码节点的特征和其父节点的特征进行特征融合，这减少了MLP网络的使用，降低了数据压缩模型的复杂度；此外，由于循环网络层不需要使用对待编码节点的特征和其父节点的特征进行融合得到的中间特征，因此发送端无需存储待编码节点的特征、父节点的特征以及中间特征，这减少了发送端进行上下文预测所需的存储空间。
针对于上述确定第i层节点的数据占有信息的过程，这里以图7所示出的数据压缩模型700为例进行说明，如图8所示，图8为本申请提供的另一种数据处理方法的流程示意图，S520可以包括以下步骤。
S5201、发送端将第i层节点的位置、深度和子节点编号，以及第i-1层节点的占有情况字节中至少一种输入特征提取层，得到第i层节点的特征。
上述第i层节点的位置、深度和子节点编号，以及第i-1层节点的占有情况字节也可以称为该第i层节点的显性信息。如图7所示，特征提取层710可以根据节点的显性信息获取该节点的特征，示例的，第四提取层714根据待编码节点734的位置、深度和子节点编号，以及父节点733的占有情况字节，获取待编码节点734的特征X t，该特征还可以被称为隐含特征或隐含信息等，该特征为高维度的特征向量。例如，每层节点的特征使用X来表示，如图7所示，在t时刻，待编码节点734的特征为X t，父节点733的特征为X t-1，祖父节点732的特征为X t-2，曾祖父节点731的特征为X t-3。
在一种示例中，上述的位置指示待编码节点734对应的体素在父节点733对应的体素 中的位置；上述的深度指示待编码节点734的深度为i。
在另一种示例中，上述的子节点编号指示待编码节点734的子节点对应的体素在待编码节点734对应的体素中的相对位置，例如，该相对位置是指子节点为待编码节点734的八个子节点中的具体哪一个。例如，子节点编号可以为1～8，8个编号分别表示八个不同的子节点。又如，子节点的编号还可以是0～7等。子节点的编号可以根据点云的压缩需求进行调整，本申请不予限定。
作为一种可选的实施方式，上述的S5201具体包括：发送端将第i层节点的位置、深度和子节点编号，第i-1层节点的占有情况字节，以及第i层节点的至少一个兄弟节点的占有情况字节输入特征提取层，得到第i层节点的特征。第i层节点的兄弟节点是指与第i层节点同属一个父节点(第i-1层节点)的其他节点，示例的，待编码节点734是图6所示出的深度为i的第一个黑色节点(左侧)，该兄弟节点可以是深度为i的第二个黑色节点(右侧)。例如，该兄弟节点的占有情况字节可以是发送端利用该兄弟节点的数据占有信息，将该兄弟节点对应的体素中的数据进行压缩后确定的。
由于第i层节点的特征包括其父节点和兄弟节点的占有情况字节，使得发送端在对该第i层节点进行上下文预测时，降低了一些明显不可能的第i层节点的占有情况字节的预测概率，提高了第i层节点的数据占有信息的预测准确率。例如，如图6中(b)所示，若第i层节点的特征不包括其兄弟节点的占有情况字节，上下文预测得到的第i层节点的占有情况字节是“1111 1111”的概率为0.05％；若第i层节点的特征包括其兄弟节点的占有情况字节，且该兄弟节点的占有情况字节为“1111 1111”，上下文预测得到的第i层节点的占有情况字节是“1111 1111”的概率会小于0.05％，如0.01％。
S5202、发送端将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的数据占有信息。
该第i-1层节点的总结信息用于描述第i-1层节点的祖先节点到第i-1层节点的所有预测信息。如图7所示，父节点733的总结信息C t-1用于描述父节点733的祖先节点到父节点733的所有预测信息。
在一种示例中，第i-1层节点的祖先节点可以是第i-1层节点的父节点至根节点中的任意一个节点。如图7所示，父节点733的祖先节点可以是图7所示出的祖父节点732或曾祖父节点731，也可以是曾祖父节点731的任意一个祖先节点，如父节点733所处的八叉树的根节点。
在一种可能的实现方式中，如图7所示，循环网络层720中的第四网络层724可以根据父节点733的数据占有信息h t-1、父节点733的总结信息C t-1和待编码节点734的特征X t获取待编码节点734的数据占有信息h t。
S5203、发送端将树状结构中第i-1层节点的数据占有信息、第i-1层节点的总结信息和第i层节点的特征中至少一种输入循环网络层，得到第i层节点的总结信息。
该第i层节点的总结信息用于描述第i层节点的祖先节点到第i层节点的所有预测信息。如图7所示，在t时刻，待编码节点734(第i层节点)的总结信息可以用C t表示。
作为一种可选的实施方式，上述的总结信息C t可以是通过迭代的方式获取的，这里以循环网络层是LSTM层为例进行说明，如图9所示，图9为本申请提供的一种LSTM的网络结构示意图，该LSTM包括两个双曲正切函数(tanh)和三个门结构(σ函数)。
Tanh和σ函数均用于选择性的利用第i层节点的前文信息，得到待编码节点的数据占有信息和总结信息。如图7所示，该前文信息可以包括父节点733的数据占有信息h t-1、总结信息C t-1和待编码节点734的特征X t。LSTM层可以利用h t-1、C t-1和X t来获取待编码节点734的总结信息C t，以及数据占有信息h t。
在LSTM网络中，该第i层节点的总结信息C t可以用于参与下一层节点(第i+1层节点)的数据占有信息预测，正如第i-1层节点的总结信息C t-1参与了第i层节点的数据占有信息预测，此处不予赘述。关于LSTM层中的Tanh和σ函数的相关原理请参考现有技术中关于LSTM的相关阐述，不予赘述。
作为一种可选的实现方式，图9所示出的
Figure PCTCN2022085349-appb-000001
和
Figure PCTCN2022085349-appb-000002
运算可以用卷积来代替。
相较于现有技术中点云压缩设备对每个节点进行上下文预测都需要从根节点开始，本申请提供的数据处理方法中，循环网络层可以利用待编码节点的特征以及待编码节点的父节点提炼的总结信息(C t-1)，得到待编码节点的数据占有信息以及总结信息(C t)，使得本申请提供的数据处理方法无需从根节点开始，这减少了上下文预测所需的计算资源。
请参见图7，第四网络层724接收第三网络层输出的用h t-1和C t-1，并在利用待编码节点734的特征X t预测待编码节点734的数据占有信息h t时，将h t-1、C t-1和X t进行整合，得到待编码节点734的总结信息C t。
值得注意的是，现有技术中点云压缩设备需要存储每个节点的特征以及融合得到的特征，而本申请采用循环网络层替代了进行特征聚合所需的多层MLP网络，减少了数据压缩模型的复杂度；此外，循环网络层可以提炼和传输祖先节点的信息，使得发送端无需存储大量的特征，这节省了发送端的存储空间。
S530、发送端根据该数据占有信息压缩待压缩数据，得到压缩数据。
在一种示例中，发送端中的熵编码器通过对待压缩数据进行多级节点的逐级压缩，得到原始数据对应的压缩数据。请继续参见图7，假设i＝3，则深度为i-3的曾祖父节点731为根节点，点云压缩的过程包括以下多级压缩的步骤。
第一级压缩：熵编码器利用根节点的数据占有信息对根节点所对应体素中的数据进行压缩。例如，根节点的数据占有信息可以是使用数据压缩模型进行上下文预测获得的，也可以是根据不同的数据压缩需求，为根节点的数据压缩给出了预设的数据占有信息(如，根节点的占有情况字节的概率分布)。
这里以熵编码器采用的编码方式是变长编码为例进行说明，若树状结构为八叉树，根节点的数据占有信息如下表1所示，根节点的占有情况字节是“0001 1000”的预测概率为89％，对待压缩数据进行压缩包括以下步骤：①，根据预测概率从大到小将各占有情况字节进行排序；②，将2个最小概率对应的占有情况字节组成1组并划分为2个分支域，并分别标记为“0”和“1”，如占有情况字节是“0000 0000”的情况标记为“0”，占有情况字节是“1111 1111”的情况标记为“1”，并将占有情况字节是“0000 0000”和“1111 1111”的2个分支域合并为1个分支域，其概率为两种占有情况字节的和(0.01％+0.003％＝0.013％)；③，将所有的占有情况字节按照和②类似的方式进行组合，直到得到的分支域的概率为1；④，查询概率和为1的分支域到各占有情况字节的路径，并记录各路径从右到左各分支域的“0”和“1”所对应的占有情况字节，得到各占有情况字节对应的码字(二进制文件)，完成根节点的数据压缩。
表1
占有情况字节 预测概率
0000 0000 0.01％
··· ···
0001 1000 89％
··· ···
1111 1111 0.003％
第二级压缩：熵编码器利用祖父节点732的数据占有信息对祖父节点732所对应体素中的数据进行压缩。
第三级压缩：熵编码器利用父节点733的数据占有信息对父节点733所对应体素中的数据进行压缩。
第四级压缩：熵编码器利用待编码节点734的数据占有信息对待编码节点734所对应体素中的数据进行压缩。
值得注意的是，第二级压缩～第四级压缩所采用的熵编码方式和第一级压缩采用的熵编码方式是相同的，此处不予赘述。此外，第一级压缩～第四级压缩所采用的熵编码方式是以变长编码(霍夫曼编码)为例进行说明，而本申请提供的数据处理方法中的数据压缩过程也可以是采用算术编码方式，本申请对数据压缩所采用的熵编码方式不予限定。
图6和图7中是以点云八叉树化后获得4级节点为例进行示意，但在点云的数据量更小时，可以使用更少层的节点对点云进行压缩；在点云的数据量更大时，可以使用更多层的节点对点云进行压缩，本申请对此不予限定。
发送端利用熵编码器，根据数据占有信息对待压缩数据进行多级压缩，得到压缩数据。相较于现有技术采用多层的MLP进行上下文预测，本申请采用循环网络层替代了进行特征聚合所需的多层MLP网络，减少了数据压缩模型的复杂度，减少了数据压缩所需的计算资源，提高了数据压缩的效率；此外，相较于现有技术中发送端需要存储n-1组特征，在本申请提供的数据处理方法中，发送端仅需保存待编码节点的数据占有信息，减少了发送端的存储空间消耗。
S540、发送端向接收端发送压缩数据。
在一种示例中，发送端可以通过通信链路向接收端发送该压缩数据。例如，若压缩数据为二进制文件，发送端将压缩数据封装为报文等合适的格式，和/或使用任意类型的传输编码或处理来处理所述压缩数据，以便在通信链路或通信网络上进行传输。
相较于现有技术中会存储每个节点的特征以及融合特征得到的中间特征，本申请提供的数据处理方法无需存储特征及中间特征，减少了数据压缩所占据的存储空间，减少了发送端和接收端传输的压缩数据的数据量，降低了传输点云的时延。
S550、接收端利用数据压缩模型确定在树状结构的数据占有信息。
该数据占有信息用于指示压缩数据在树状结构中的数据分布。关于数据压缩模型和循环网络层的具体实现方式请参考上述关于S520的相关阐述，此处不予赘述。
在数据压缩的过程中，发送端可以将预测得到的数据占有信息和父节点的占有情况字节一同送入熵编码器中，减少记录实际占有情况字节的数据所需的比特数，进而减少存储 待压缩数据所需的空间，达到数据压缩的效果。反之亦然，在数据解压的过程中，接收端为了从压缩数据中恢复树状结构，可以使用和数据压缩过程中一致的上下文预测方法，达到数据解压的效果。
在本申请的上述实施例中，关于上下文预测的过程均以实现数据压缩的发送端为例进行说明，数据解压过程中的上下文预测方法和数据压缩过程相同，此处不予赘述。
S560、接收端根据数据占有信息解压压缩数据，得到解压数据。
在一种示例中，接收端中的熵解码器通过对压缩数据进行多级节点的逐级解压，得到解压数据。请继续参见图7，假设i＝3，则深度为i-3的曾祖父节点731为根节点，点云解压的过程包括以下多级解压的步骤。
第一级解压：熵解码器利用根节点的数据占有信息对根节点所对应体素中的压缩数据进行解压。例如，根节点的数据占有信息可以是使用数据压缩模型进行上下文预测获得的，也可以是根据不同的数据压缩需求，为根节点的数据压缩和解压给出了预设的数据占有信息(如，根节点的占有情况字节的概率分布)。
这里以熵解码器采用的编码方式是变长编码为例进行说明，若树状结构为八叉树，根节点的数据占有信息如下表2所示，根节点的占有情况字节是“0001 1000”的预测概率为89％，对压缩数据进行解压包括以下步骤：①，根据预测概率从大到小将各占有情况字节进行排序；②，将2个最小概率对应的占有情况字节组成1组并划分为2个分支域，并分别标记为“0”和“1”，如占有情况字节是“0000 0000”的情况标记为“0”，占有情况字节是“1111 1111”的情况标记为“1”，并将占有情况字节是“0000 0000”和“1111 1111”的2个分支域合并为1个分支域，其概率为两种占有情况字节的和(0.01％+0.003％＝0.013％)；③，将所有的占有情况字节按照和②类似的方式进行组合，直到得到的分支域的概率为1；④，查询概率和为1的分支域到各占有情况字节的路径，并记录各路径从右到左各分支域的“0”和“1”所对应的占有情况字节，得到各占有情况字节对应的码字(二进制文件)，完成根节点对应的压缩数据的解压。
表2
占有情况字节 预测概率
0000 0000 0.01％
··· ···
0001 1000 89％
··· ···
1111 1111 0.003％
第二级解压：熵解码器利用祖父节点732的数据占有信息对祖父节点732所对应体素中的数据进行解压。
第三级解压：熵解码器利用父节点733的数据占有信息对父节点733所对应体素中的数据进行解压。
第四级解压：熵解码器利用待编码节点734的数据占有信息对待编码节点734所对应体素中的数据进行解压。
值得注意的是，第二级解压～第四级解压所采用的编码方式和第一级解压采用的编码方式是相同的，此处不予赘述。此外，第一级解压～第四级解压所采用的编码方式是以变 长编码(霍夫曼编码)为例进行说明，而本申请提供的数据处理方法中的数据解压过程也可以是采用算术编码方式，本申请对数据解压所采用的编码方式不予限定，但是数据解压和数据压缩采用的编码方式应是一致的。
图6和图7中是以压缩数据重构八叉树后获得4级节点为例进行示意，但在压缩数据的数据量更小时，可以使用更少层的节点对压缩数据进行解压；在压缩数据的数据量更大时，可以使用更多层的节点对压缩数据进行解压，本申请对此不予限定。
本申请采用一个循环网络层替代了现有技术中多层的MLP网络，降低了数据压缩模型的复杂度，减少了上下文预测获取数据占有信息时所需的计算资源；另外，相较于现有技术中会存储每个节点的特征以及融合特征得到的中间特征，本申请提供的数据处理方法无需存储特征及中间特征，减少了数据解压所占据的存储空间。
通常，点云的数据量很大，数据压缩的过程也会比较长，数据压缩的过程也会出现各种突发情况(如数据压缩出现卡顿或停止)，为了便于监控数据压缩过程，这里以图5所示出的发送端具有显示单元为例进行说明，该数据处理方法还包括以下步骤。
S570、发送端显示树状结构和数据占有信息中至少一种。
上述的显示单元可以是显示屏，例如，若发送端为手机，则该显示屏可以是触摸屏。又如，若发送端为自动驾驶系统的控制设备，该显示屏可以是设置于车辆内部靠近驾驶员侧的平视显示器(head up display，HUD)，该显示屏还可以是设置于车辆内部的投影仪的投影区域。
在第一种可能的设计中，如图10中的(a)所示，发送端可以显示待压缩数据的树状结构，其指示第i-1层节点和第i层节点的数据分布情况。
在第二种可能的设计中，发送端可以显示数据占有信息。如图10中的(b)所示，其指示了第i-1层节点的占有情况字节的预测概率，如，第i-1层节点的占有情况字节为“00000000”的预测概率为0.1％，第i-1层节点的占有情况字节为“1000 0100”的预测概率为99％，第i-1层节点的占有情况字节为“1111 1111”的预测概率为0.05％。
在第三种可能的设计中，如图10中的(c)所示，发送端可以同时显示待压缩数据的树状结构和待编码节点(如第i-1层节点)的数据占有信息。
在数据压缩的过程比较长的情况下，通过对树状结构和数据占有信息中至少一种进行显示，有利于用户监控数据压缩过程，定位数据压缩出现错误(如数据压缩出现卡顿或停止)的位置，提高数据压缩的准确性。
数据占有信息可能是运营人员无法识别的一组数据，在数据压缩或解压时，若上下文预测出现明显的错误，难以确定上下文预测出现的问题，为了解决上述问题，这里以数据压缩模型还包括维度调整层为例进行说明，请继续参见图5，该数据处理方法还可以包括以下步骤。
S580、发送端将第i层节点的数据占有信息输入维度调整层，得到占有率预测表。
上述的维度调整层可以包括至少一层MLP，该MLP用于调整数据占有信息的输出维度。例如，若树状结构为八叉树，则该MLP的输出维度可以是256，以将“0000 0000”～“1111 1111”对应的每一种占有情况字节的预测概率进行输出。又如，若树状结构为八叉树，则该MLP的输出维度还可以是260，除了输出每种占有情况字节的概率，还可以预留4个维度作为可选项，该可选项可以指示发送端的设备类型(如手机、电脑)，该可选项还可以作 为压缩数据的存储地址，以指示压缩数据的存储位置等。
该占有率预测表指示第i层节点的每个占有情况字节的预测概率。例如，若树状结构为两叉树，该占有率预测表可以如表3所示，第i层节点的占有情况字节是“00”的预测概率为5％，第i层节点的占有情况字节是“01”的预测概率为80％，第i层节点的占有情况字节是“10”的预测概率为12％，第i层节点的占有情况字节是“11”的预测概率为3％。
表3
序号 占有情况字节 预测概率
1 00 5％
2 01 80％
3 10 12％
4 11 3％
在数据压缩或解压的过程中，利用维度调整层对数据占有信息进行维度调整，得到可视化的占有率预测表，有利于运营人员对数据压缩或解压的过程进行监控，在上下文预测出现明显的错误的情况下，运营人员可以将该占有率预测表作为参考，以便快速确定数据压缩或解压过程中出现的问题，提高数据压缩和解压过程的鲁棒性。
作为一种可选的实施方式，若待压缩数据的树状结构为八叉树，针对于上述的数据压缩模型，本申请提供一种可能的具体实现方式，这里以特征提取层包括3层MLP、循环网络层包括3层LSTM、维度调整层包括1层MLP为例进行说明，如图11所示，图11为本申请提供的另一种数据压缩模型的结构示意图，该数据压缩模型1100包括特征提取层1110、循环网络层1120和维度调整层1130。
特征提取层1110包括第一MLP 1111(输入维度为m、输出维度为128)、第二MLP 1112(输入维度为128、输出维度为128)、第三MLP 1113(输入维度为128、输出维度为128)。特征提取层1110可以实现图7所示出的第四提取层714的功能，例如，特征提取层1110用于根据第i层节点的位置、深度和子节点编号，第i-1层节点的占有情况字节，以及第i层节点的至少一个兄弟节点的占有情况字节中至少一种，得到第i层节点的特征X t。关于第i层节点的特征和特征X t请参考S5201的相关阐述，此处不予赘述。
循环网络层1120包括第一LSTM 1121(输入维度为128、输出维度为128、步长为1×1)、第二LSTM 1122(输入维度为128、输出维度为128、步长为1×1)、第三LSTM 1123(输入维度为128、输出维度为128、步长为1×1)。循环网络层1120可以实现图7所示出的第四网络层724的功能，例如，循环网络层1120可以根据第i-1层节点的数据占有信息h t-1、第i-1层节点的总结信息C t-1和第i层节点的特征X t，获取第i层节点的数据占有信息h t，以及第i层节点的总结信息C t。
维度调整层1130包括第四MLP 1131(输入维度为128、输出维度为256)。维度调整层1130可以调整数据占有信息的输出维度，例如，第四MLP 1131将第i层节点的数据占有信息h t的输出维度调整为256，得到占有率预测表。示例的，占有率预测表可以包括“0000 0000”～“1111 1111”等256种占有情况字节的预测概率。
以图11所示出的数据压缩模型为例，本申请提供的数据处理方法，预测待编码节点的数据占有信息总共需要4个MLP层和3个LSTM层，以及储存3个LSTM层输出的总结信息和数据占有信息，数据压缩和解压所需的计算量和存储空间均为常数。
相比之下，在图4所提供的上下文预测方法中，无论待编码节点属于哪一层，Oct Squeeze算法最多也只能利用到n-1个祖先节点的信息(如，n＝4)，导致在待编码节点的深度i值(n≤i)较大时，对待编码节点进行上下文预测得到的占有率预测结果的精度较低。例如，预测一个节点的数据占有信息需要计算4n+1个MLP层以及存储n-1组特征，也就是说，随着n值的增大，上下文预测的计算量和特征所需的存储空间会线性增加。
因此，本申请提供的数据处理方法采用循环网络层替代了多层的MLP层，降低了数据压缩模型的复杂度，且数据压缩模型所占用的计算资源均为常数，使得在待编码节点的深度增加或祖先节点的数量增加的情况下，对待编码节点进行上下文预测不会占用更多的计算资源；此外，待编码节点的深度每增加一层，存储总结信息和数据占有信息所需的空间也是固定的，这减少了数据压缩和解压所需的存储空间。
此外，Oct Squeeze算法进行上下文预测时，只能采用固定的n-1个祖先节点的信息，而在本申请提供的数据处理方法中，数据压缩模型中的循环网络层可以利用从根节点开始到待编码节点的父节点的所有预测信息，且这些预测信息可以被循环网络层选择性的记忆和保留，使得在数据压缩模型预测待编码节点的数据占有信息时，循环网络层可以利用待编码节点的所有祖先节点的信息，提高了上下文预测的精度，提高了数据压缩比。
可以理解的是，为了实现上述实施例中的功能，计算设备包括了执行各个功能相应的硬件结构和/或软件模块。本领域技术人员应该很容易意识到，结合本申请中所公开的实施例描述的各示例的单元及方法步骤，本申请能够以硬件或硬件和计算机软件相结合的形式来实现。某个功能究竟以硬件还是计算机软件驱动硬件的方式来执行，取决于技术方案的特定应用场景和设计约束条件。
上文中结合图1至图11，详细描述了根据本实施例所提供的数据处理方法，下面将结合图12和图13，描述根据本实施例所提供的数据处理装置和计算设备。
图12为本申请提供的一种数据处理装置的示意图。这些数据处理装置可以用于实现上述方法实施例中发送端和接收端的功能，因此也能实现上述方法实施例所具备的有益效果。在本实施例中，该数据处理装置可以是如图1所示的终端111～终端113或数据中心130，还可以是应用于应用服务器131的模块(如芯片)。
下面结合图12对第一数据处理装置1210与第二数据处理装置1220的结构和功能进行介绍，第一数据处理装置1210可以实现图2所示出的发送端的功能，第二数据处理装置1220可以实现图2所示出的接收端的功能。应理解，本实施例仅对第一数据处理装置1210与第二数据处理装置1220的结构和功能模块进行示例性划分，本申请并不对其具体划分做任何限定。
如图12所示，第一数据处理装置1210通过通信信道1230与第二数据处理装置1220建立通信连接，通信信道1230可以传输上述发送端向接收端发送的压缩数据，通信信道1230还可以传输接收端向发送端发送的其他信息。
如图12所示，第一数据处理装置1210包括采集单元1211、显示单元1212、预处理单元1213、上下文预测单元1214、编码单元1215和收发单元1216，上述单元可以用于实现上述图5或图8中所示的方法实施例中发送端执行的各个操作步骤对应的方法。
当第一数据处理装置1210用于实现图5所示的方法实施例中的功能时，显示单元1212用于执行S570，预处理单元1213用于实现S510，上下文预测单元1214用于实现S520和 S580，编码单元1215用于实现S530，收发单元1216用于实现S540。
可选的，当第一数据处理装置1210用于实现图8所示的方法实施例中的功能时，上下文预测单元1214还用于实现S5201～S5203。
例如，采集单元1211可以实现上述实施例提供的传感器(如激光雷达、毫米波雷达和声呐等)和摄像头实现的功能。显示单元1212可以是包括显示屏，例如，若第一数据处理装置1210为手机，则该显示屏可以是触摸屏。又如，若第一数据处理装置1210为自动驾驶系统的控制设备，该显示屏可以是HUD等。
如图12所示，第二数据处理装置1220包括获取单元1221、上下文预测单元1222和解压单元1223，上述单元可以用于实现上述图5或图8中所示的方法实施例中接收端执行的各个操作步骤对应的方法。
当第二数据处理装置1220用于实现图5或图8所示的方法实施例中的功能时，获取单元1221用于获取压缩数据，上下文预测单元1222用于执行S550，解压单元1223用于执行S560。
有关上述第一数据处理装置1210与第二数据处理装置1220更详细的描述可以直接参考上述图5或图8所示的方法实施例中相关描述直接得到，这里不加赘述。
图13为本申请提供的一种计算设备的结构示意图，该计算设备1300包括处理器1310和通信接口1320。处理器1310和通信接口1320之间相互耦合。可以理解的是，通信接口1320可以为收发器或输入输出接口。可选的，计算设备1300还可以包括存储器1330，用于存储处理器1310执行的指令或存储处理器1310运行指令所需要的输入数据或存储处理器1310运行指令后产生的数据。
作为一种可能的实现方式，处理器1310可以根据原始数据生成树状结构的待压缩数据，并利用数据压缩模型包含的循环网络层确定在树状结构的数据占有信息。该数据占有信息用于指示原始数据在树状结构中的数据分布。进而，处理器1310根据该数据占有信息压缩上述的待压缩数据得到压缩数据。
当计算设备1300用于实现图5或图8所示的方法时，处理器1310、通信接口1320和存储器1330还可以协同实现发送端和接收端执行的数据处理方法中的各个操作步骤。计算设备1300还可以执行图12所示出的第一数据处理装置1210和第二数据处理装置1220的功能，此处不予赘述。
本申请实施例中不限定上述通信接口1320、处理器1310以及存储器1330之间的具体连接介质。本申请实施例在图13中以通信接口1320、处理器1310以及存储器1330之间通过总线1340连接，总线在图13中以粗线表示，其它部件之间的连接方式，仅是进行示意性说明，并不引以为限。所述总线可以分为地址总线、数据总线、控制总线等。为便于表示，图13中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
存储器1330可用于存储软件程序及模块，如本申请实施例所提供的数据处理方法对应的程序指令/模块，处理器1310通过执行存储在存储器1330内的软件程序及模块，从而执行各种功能应用以及数据处理。该通信接口1320可用于与其他设备进行信令或数据的通信。在本申请中该计算设备1300可以具有多个通信接口1320。
其中，上述的存储器可以是但不限于，随机存取存储器(Random Access Memory，RAM)，只读存储器(Read Only Memory，ROM)，可编程只读存储器(Programmable  Read-Only Memory，PROM)，可擦除只读存储器(Erasable Programmable Read-Only Memory，EPROM)，电可擦除只读存储器(Electric Erasable Programmable Read-Only Memory，EEPROM)等。
上述的处理器可以是一种集成电路芯片，具有信号处理能力。该处理器可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、网络处理器(Network Processor，NP)等；还可以是数字信号处理器(Digital Signal Processing，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field－Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。
本申请的实施例中的方法步骤可以通过硬件的方式来实现，也可以由处理器执行软件指令的方式来实现。软件指令可以由相应的软件模块组成，软件模块可以被存放于RAM、闪存、ROM、PROM、EPROM、EEPROM、寄存器、硬盘、移动硬盘、CD-ROM或者本领域熟知的任何其它形式的存储介质中。一种示例性的存储介质耦合至处理器，从而使处理器能够从该存储介质读取信息，且可向该存储介质写入信息。当然，存储介质也可以是处理器的组成部分。处理器和存储介质可以位于ASIC中。另外，该ASIC可以位于计算设备中。当然，处理器和存储介质也可以作为分立组件存在于计算设备中。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机程序或指令。在计算机上加载和执行所述计算机程序或指令时，全部或部分地执行本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、通信装置、用户设备或者其它可编程装置。所述计算机程序或指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机程序或指令可以从一个网站站点、计算机、服务器或数据中心通过有线或无线方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是集成一个或多个可用介质的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质，例如，软盘、硬盘、磁带；也可以是光介质，例如，数字视频光盘(digital video disc，DVD)；还可以是半导体介质，例如，固态硬盘(solid state drive，SSD)。
在本申请的各个实施例中，如果没有特殊说明以及逻辑冲突，不同的实施例之间的术语和/或描述具有一致性、且可以相互引用，不同的实施例中的技术特征根据其内在的逻辑关系可以组合形成新的实施例。
本申请中，“至少一个”是指一个或者多个，“多个”是指两个或两个以上。“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B的情况，其中A,B可以是单数或者复数。在本申请的文字描述中，字符“/”，一般表示前后关联对象是一种“或”的关系；在本申请的公式中，字符“/”，表示前后关联对象是一种“相除”的关系。
本申请说明书和权利要求书及上述附图中的术语“第一”、“第二”和“第三”等是用于区别不同对象，而不是用于限定特定顺序。
在本申请实施例中，“示例性的”或者“例如”等词用于表示作例子、例证或说明。本申 请实施例中被描述为“示例性的”或者“例如”的任何实施例或设计方案不应被解释为比其它实施例或设计方案更优选或更具优势。确切而言，使用“示例性的”或者“例如”等词旨在以具体方式呈现相关概念。
可以理解的是，在本申请的实施例中涉及的各种数字编号仅为描述方便进行的区分，并不用来限制本申请的实施例的范围。上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定。

Claims (30)

  1. 一种数据处理方法，其特征在于，包括：
    根据原始数据生成树状结构的待压缩数据；
    利用数据压缩模型确定在所述树状结构的数据占有信息，所述数据占有信息用于指示所述原始数据在所述树状结构中的数据分布，所述数据压缩模型包含循环网络层，所述循环网络层用于确定所述数据占有信息；
    根据所述数据占有信息压缩所述待压缩数据，得到压缩数据。
  2. 根据权利要求1所述的方法，其特征在于，在所述利用数据压缩模型确定在所述树状结构的数据占有信息之前，所述方法还包括：
    通过传感器采集所述原始数据，所述原始数据为三维数据，所述传感器包括激光雷达、毫米波雷达和声呐中至少一种。
  3. 根据权利要求1所述的方法，其特征在于，在所述利用数据压缩模型确定在所述树状结构的数据占有信息之前，所述方法还包括：
    通过摄像头采集所述原始数据，所述原始数据为二维数据。
  4. 根据权利要求1-3中任一项所述的方法，其特征在于，所述方法还包括：
    显示所述树状结构和/或所述数据占有信息。
  5. 根据权利要求1-4中任一项所述的方法，其特征在于，所述利用数据压缩模型确定在所述树状结构的数据占有信息，包括：
    将树状结构中第i-1层节点的数据占有信息输入所述循环网络层，得到第i层节点的数据占有信息，i为正整数，所述第i-1层节点是所述第i层节点的父节点。
  6. 根据权利要求1-4中任一项所述的方法，其特征在于，所述利用数据压缩模型确定在所述树状结构的数据占有信息，包括：
    将所述树状结构中第i-1层节点的数据占有信息、所述第i-1层节点的总结信息和所述第i层节点的特征中至少一种输入所述循环网络层，得到所述第i层节点的数据占有信息，所述第i-1层节点的总结信息用于描述所述第i-1层节点的祖先节点到所述第i-1层节点的所有预测信息。
  7. 根据权利要求6所述的方法，其特征在于，所述方法还包括：
    将所述树状结构中第i-1层节点的数据占有信息、所述第i-1层节点的总结信息和所述第i层节点的特征中至少一种输入所述循环网络层，得到所述第i层节点的总结信息，所述第i层节点的总结信息用于描述所述第i层节点的祖先节点到所述第i层节点的所有预测信息。
  8. 根据权利要求6或7所述的方法，其特征在于，所述数据压缩模型还包括特征提取层，所述方法还包括：
    将所述第i层节点的位置、深度和子节点编号，以及所述第i-1层节点的占有情况字节中至少一种输入所述特征提取层，得到所述第i层节点的特征。
  9. 根据权利要求8所述的方法，其特征在于，将所述第i层节点的位置、深度和子节点编号，以及所述第i-1层节点的占有情况字节中至少一种输入所述特征提取层，得到所述第i层节点的特征，包括：
    将所述第i层节点的位置、深度和子节点编号，所述第i-1层节点的占有情况字节，以 及所述第i层节点的至少一个兄弟节点的占有情况字节输入所述特征提取层，得到所述第i层节点的特征。
  10. 根据权利要求5-9中任一项所述的方法，其特征在于，所述数据压缩模型还包括维度调整层，所述方法还包括：
    将所述第i层节点的数据占有信息输入所述维度调整层，得到占有率预测表，所述占有率预测表指示所述第i层节点的每个占有情况字节的预测概率。
  11. 根据权利要求10所述的方法，其特征在于，所述维度调整层包括至少一层多层感知机MLP。
  12. 根据权利要求8或9所述的方法，其特征在于，所述特征提取层包括至少一层MLP。
  13. 根据权利要求1-12中任一项所述的方法，其特征在于，所述循环网络层包括至少一层长短期记忆网络LSTM层。
  14. 一种数据处理方法，其特征在于，包括：
    获取压缩数据；
    利用数据压缩模型确定在树状结构的数据占有信息，所述数据占有信息用于指示所述压缩数据在所述树状结构中的数据分布，所述数据压缩模型包含循环网络层，所述循环网络层用于确定所述数据占有信息；
    根据所述数据占有信息解压所述压缩数据，得到解压数据。
  15. 根据权利要求14所述的方法，其特征在于，所述利用数据压缩模型确定在所述树状结构的数据占有信息，包括：
    将树状结构中第i-1层节点的数据占有信息输入所述循环网络层，得到第i层节点的数据占有信息，i为正整数，所述第i-1层节点是所述第i层节点的父节点。
  16. 根据权利要求14所述的方法，其特征在于，所述利用数据压缩模型确定在所述树状结构的数据占有信息，包括：
    将所述树状结构中第i-1层节点的数据占有信息、所述第i-1层节点的总结信息和所述第i层节点的特征中至少一种输入所述循环网络层，得到所述第i层节点的数据占有信息，所述第i-1层节点的总结信息用于描述所述第i-1层节点的祖先节点到所述第i-1层节点的所有预测信息。
  17. 根据权利要求16所述的方法，其特征在于，所述方法还包括：
    将所述树状结构中第i-1层节点的数据占有信息、所述第i-1层节点的总结信息和所述第i层节点的特征中至少一种输入所述循环网络层，得到所述第i层节点的总结信息，所述第i层节点的总结信息用于描述所述第i层节点的祖先节点到所述第i层节点的所有预测信息。
  18. 根据权利要求16或17所述的方法，其特征在于，所述数据压缩模型还包括特征提取层，所述方法还包括：
    将所述第i层节点的位置、深度和子节点编号，以及所述第i-1层节点的占有情况字节中至少一种输入所述特征提取层，得到所述第i层节点的特征。
  19. 根据权利要求18所述的方法，其特征在于，将所述第i层节点的位置、深度和子节点编号，以及所述第i-1层节点的占有情况字节中至少一种输入所述特征提取层，得到所述第i层节点的特征，包括：
    将所述第i层节点的位置、深度和子节点编号，所述第i-1层节点的占有情况字节，以及所述第i层节点的至少一个兄弟节点的占有情况字节输入所述特征提取层，得到所述第i层节点的特征。
  20. 根据权利要求15-19中任一项所述的方法，其特征在于，所述数据压缩模型还包括维度调整层，所述方法还包括：
    将所述第i层节点的数据占有信息输入所述维度调整层，得到占有率预测表，所述占有率预测表指示所述第i层节点的每个占有情况字节的预测概率。
  21. 根据权利要求20所述的方法，其特征在于，所述维度调整层包括至少一层多层感知机MLP。
  22. 根据权利要求18或19所述的方法，其特征在于，所述特征提取层包括至少一层MLP。
  23. 根据权利要求14-22中任一项所述的方法，其特征在于，所述循环网络层包括至少一层长短期记忆网络LSTM层。
  24. 一种数据处理装置，其特征在于，包括：
    预处理单元，用于根据原始数据生成树状结构的待压缩数据；
    上下文预测单元，用于利用数据压缩模型确定在所述树状结构的数据占有信息，所述数据占有信息用于指示所述原始数据在所述树状结构中的数据分布，所述数据压缩模型包含循环网络层，所述循环网络层用于确定所述数据占有信息；
    编码单元，用于根据所述数据占有信息压缩所述待压缩数据，得到压缩数据。
  25. 根据权利要求24所述的装置，其特征在于，所述上下文预测单元，具体用于将所述树状结构中第i-1层节点的数据占有信息、所述第i-1层节点的总结信息和所述第i层节点的特征中至少一种输入所述循环网络层，得到所述第i层节点的数据占有信息，所述第i-1层节点的总结信息用于描述所述第i-1层节点的祖先节点到所述第i-1层节点的所有预测信息。
  26. 根据权利要求25所述的装置，其特征在于，所述上下文预测单元，还用于将所述树状结构中第i-1层节点的数据占有信息、所述第i-1层节点的总结信息和所述第i层节点的特征中至少一种输入所述循环网络层，得到所述第i层节点的总结信息，所述第i层节点的总结信息用于描述所述第i层节点的祖先节点到所述第i层节点的所有预测信息。
  27. 根据权利要求24-26中任一项所述的装置，其特征在于，所述循环网络层包括至少一层长短期记忆网络LSTM层。
  28. 一种数据处理装置，其特征在于，包括：
    获取单元，用于获取压缩数据；
    上下文预测单元，用于利用数据压缩模型确定在树状结构的数据占有信息，所述数据占有信息用于指示所述压缩数据在所述树状结构中的数据分布，所述数据压缩模型包含循环网络层，所述循环网络层用于确定所述数据占有信息；
    解压单元，用于根据所述数据占有信息解压所述压缩数据，得到解压数据。
  29. 一种计算设备，其特征在于，包括存储器和处理器，所述存储器用于存储一组计算机指令；当所述处理器执行所述一组计算机指令时，执行上述权利要求1至13中任一项所述的方法的操作步骤，或权利要求14-23中任一项所述的方法的操作步骤。
  30. 一种计算机可读存储介质，其特征在于，所述存储介质中存储有计算机程序或指令，当所述计算机程序或指令被计算设备执行时，实现如权利要求1至13中任一项所述的方法，或权利要求14-23中任一项所述的方法。
PCT/CN2022/085349 2021-04-09 2022-04-06 一种数据处理方法及装置 WO2022213992A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110384626.7A CN115249033A (zh) 2021-04-09 2021-04-09 一种数据处理方法及装置
CN202110384626.7 2021-04-09

Publications (1)

Publication Number Publication Date
WO2022213992A1 true WO2022213992A1 (zh) 2022-10-13

Family

ID=83545989

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/085349 WO2022213992A1 (zh) 2021-04-09 2022-04-06 一种数据处理方法及装置

Country Status (2)

Country Link
CN (1) CN115249033A (zh)
WO (1) WO2022213992A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115857836A (zh) * 2023-02-10 2023-03-28 中南大学湘雅医院 基于大数据的信息储存方法及装置

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115421161B (zh) * 2022-11-03 2023-04-07 上海伯镭智能科技有限公司 基于激光雷达测距的无人驾驶矿车控制方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106911930A (zh) * 2017-03-03 2017-06-30 深圳市唯特视科技有限公司 一种基于递归卷积神经网络进行压缩感知视频重建的方法
US10839530B1 (en) * 2018-09-04 2020-11-17 Apple Inc. Moving point detection
CN112438049A (zh) * 2020-06-24 2021-03-02 北京小米移动软件有限公司 编码和解码方法、编码器、解码器以及软件
CN112581552A (zh) * 2020-12-14 2021-03-30 深圳大学 自适应分块的基于体素的点云压缩方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106911930A (zh) * 2017-03-03 2017-06-30 深圳市唯特视科技有限公司 一种基于递归卷积神经网络进行压缩感知视频重建的方法
US10839530B1 (en) * 2018-09-04 2020-11-17 Apple Inc. Moving point detection
CN112438049A (zh) * 2020-06-24 2021-03-02 北京小米移动软件有限公司 编码和解码方法、编码器、解码器以及软件
CN112581552A (zh) * 2020-12-14 2021-03-30 深圳大学 自适应分块的基于体素的点云压缩方法及装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
TU CHENXI; TAKEUCHI EIJIRO; CARBALLO ALEXANDER; TAKEDA KAZUYA: "Point Cloud Compression for 3D LiDAR Sensor using Recurrent Neural Network with Residual Blocks", 2019 INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION (ICRA), 20 May 2019 (2019-05-20), pages 3274 - 3280, XP033594196, DOI: 10.1109/ICRA.2019.8794264 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115857836A (zh) * 2023-02-10 2023-03-28 中南大学湘雅医院 基于大数据的信息储存方法及装置

Also Published As

Publication number Publication date
CN115249033A (zh) 2022-10-28

Similar Documents

Publication Publication Date Title
US11920934B2 (en) Path planning using sparse volumetric data
WO2022213992A1 (zh) 一种数据处理方法及装置
US10511696B2 (en) System and method for aggregation, archiving and compression of internet of things wireless sensor data
WO2024060558A1 (zh) 可行域预测方法、装置、系统和存储介质
US11533514B2 (en) Encoding method, decoding method, information processing method, encoding device, decoding device, and information processing system
Varischio et al. Hybrid point cloud semantic compression for automotive sensors: A performance evaluation
Xing et al. MABNet: a lightweight stereo network based on multibranch adjustable bottleneck module
US20230281881A1 (en) Video Frame Compression Method, Video Frame Decompression Method, and Apparatus
CN114598883A (zh) 点云属性的预测方法、编码器、解码器及存储介质
WO2023174256A1 (zh) 一种数据压缩方法以及相关设备
WO2023116567A1 (zh) 数据编解码方法、装置和设备
Wu et al. Hybrid mobile vision for emerging applications
CN113870271A (zh) 3d点云的压缩方法、装置、设备及存储介质
Shao et al. Point cloud in the air
WO2023024840A1 (zh) 点云编解码方法、编码器、解码器及存储介质
WO2022213843A1 (zh) 一种图像处理方法、训练方法及装置
WO2024011381A1 (zh) 点云编解码方法、装置、设备及存储介质
WO2024026712A1 (zh) 点云编解码方法、装置、设备及存储介质
WO2023024842A1 (zh) 点云编解码方法、装置、设备及存储介质
WO2022257528A1 (zh) 点云属性的预测方法、装置及相关设备
CN116781916B (zh) 车辆图像存储方法、装置、电子设备和计算机可读介质
US20240187648A1 (en) Coding Vertex Information of a Point Cloud
WO2022188582A1 (zh) 点云中邻居点的选择方法、装置及编解码器
WO2023173238A1 (zh) 编解码方法、码流、编码器、解码器以及存储介质
WO2024086154A1 (en) Tree-based deep entropy model for point cloud compression

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22784058

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22784058

Country of ref document: EP

Kind code of ref document: A1