Point cloud processing system and method
By designing a point cloud processing system and utilizing hash filtering and gradient calculation modules for pipelined parallel processing, the problems of high computational load and high energy consumption in existing technologies are solved, achieving efficient point cloud data processing that is suitable for embedded environments such as automobiles.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YAOYAO
- Filing Date
- 2022-07-18
- Publication Date
- 2026-06-02
AI Technical Summary
Existing autonomous driving algorithms suffer from complex data structures, large computational load, and high power and energy consumption when computed on CPUs or GPUs, making them unsuitable for direct application in embedded environments such as automobiles.
Design a point cloud processing system, including a PL end and a PS end. Utilize a hash filtering module and a gradient calculation module for pipelined parallel processing. Combine with a data transport module and an application processor to achieve parallel processing and on-chip storage of point cloud data, reducing external DDR memory access operations.
It improves data throughput, reduces latency, and can be directly applied to embedded environments such as automobiles, thereby enhancing the performance and energy efficiency of the algorithm.
Smart Images
Figure CN115202758B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a point cloud processing system and method. Background Technology
[0002] Most existing autonomous driving-related algorithms are computed using a central processing unit (CPU) or a graphics processing unit (GPU). Computation based on existing CPUs or GPUs generally suffers from complex data structures, huge computational and data volumes, and high power and energy consumption, making them unsuitable for direct application in embedded environments such as automobiles. Summary of the Invention
[0003] To address the aforementioned technical problems, embodiments of the present invention provide a point cloud processing system and method.
[0004] In a first aspect, embodiments of the present invention provide a point cloud processing system, comprising:
[0005] PL terminal and PS terminal, wherein the PL terminal and the PS terminal are electrically connected;
[0006] The PL terminal includes a hash filtering module, a gradient calculation module, and a data transport module. The hash filtering module is electrically connected to the data transport module and the gradient calculation module, respectively. The gradient calculation module is also electrically connected to the data transport module. The data transport module is also electrically connected to the application processor of the PS terminal.
[0007] The PS terminal is used to send point cloud data calculation commands to the PL terminal;
[0008] The hash filtering module is used to generate a point cloud data reading command according to the point cloud data calculation command, send the point cloud data reading command to the data transport module, receive initial point cloud data from the data transport module, and process the initial point cloud data in real time using a pipeline to obtain processed point cloud data.
[0009] The gradient calculation module is used to perform gradient calculation on the processed point cloud data in a data stream parallel manner through multiple calculation sub-modules connected by FIFO or ping-pong buffer to obtain multiple gradient vectors.
[0010] The application processor on the PS side is used to calculate the descent direction of multiple gradient vectors and determine whether the descent direction converges. If the descent direction converges, the termination point is cloud computing.
[0011] According to one specific embodiment of this disclosure, the application processor at the PS end is further configured to initiate point cloud iterative calculation if the descent direction does not converge.
[0012] According to one specific embodiment of this disclosure, the PS terminal is further configured to receive parameter configuration commands and start commands for the AXI4-Lite interface, configure the AXI4-Lite interface according to the parameter configuration commands, and generate the point cloud data calculation commands according to the start commands.
[0013] According to one specific embodiment of this disclosure, the hash filtering module is further configured to perform pipelined real-time noise reduction processing on the initial point cloud data to obtain downsampled point cloud data, and store the downsampled point cloud data in an on-chip cache.
[0014] According to one specific embodiment of this disclosure, the gradient calculation module is further configured to pipeline the reading of the downsampled point cloud data, perform coordinate transformation processing on the downsampled point cloud data to obtain coordinate-transformed point cloud data, acquire adjacent grid data of the coordinate-transformed point cloud data, and calculate the current pose score and gradient vector based on the coordinate-transformed point cloud data and the adjacent grid data.
[0015] According to one specific embodiment of this disclosure, the gradient calculation module is further configured to obtain the point cloud data storage address of the adjacent grid based on radius search, and read the adjacent grid data from the point cloud data storage address of the adjacent grid through the AXI4-Full interface.
[0016] According to one specific embodiment of this disclosure, the hash filtering module is further configured to store the processed point cloud data into an on-chip cache.
[0017] According to one specific embodiment of this disclosure, the gradient calculation module is further configured to write a plurality of the gradient vectors into an on-chip cache.
[0018] Secondly, embodiments of the present invention provide a point cloud processing method, applied to the point cloud point location processing system provided in the first aspect, the method comprising:
[0019] The PS terminal sends a point cloud data calculation command to the PL terminal;
[0020] The hash filtering module generates a point cloud data reading command based on the point cloud data calculation command, sends the point cloud data reading command to the data transport module, receives the initial point cloud data from the data transport module, and processes the initial point cloud data in real time using a pipeline to obtain the processed point cloud data.
[0021] The gradient calculation module performs gradient calculations on the processed point cloud data in a parallel data stream manner through multiple calculation sub-modules connected by FIFO or ping-pong buffers to obtain multiple gradient vectors.
[0022] The application processor on the PS side calculates the descent direction of multiple gradient vectors and determines whether the descent direction converges. If the descent direction converges, the termination point is cloud computing.
[0023] According to one specific embodiment of this disclosure, the method further includes:
[0024] If the descent direction does not converge, the application processor at the PS end initiates point cloud iterative calculation.
[0025] The point cloud processing system and method provided in this application fully utilize pipeline technology to achieve parallel processing of point cloud data, greatly improving data throughput. They optimize the data storage structure of the reference map and utilize on-chip storage for critical data, reducing access to external DDR memory. By employing a hardware-software co-processing approach for point cloud localization, computationally intensive tasks are accelerated using hardware, while control-related logic is implemented in software. This significantly improves the actual performance of the algorithm, greatly reduces latency, and allows for direct application in embedded environments such as automotive. Attached Figure Description
[0026] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope of protection of the present invention. In the various drawings, similar components are numbered similarly.
[0027] Figure 1 This diagram illustrates a flow chart of a CPU-based cloud computing process in the prior art.
[0028] Figure 2 A schematic diagram of the point cloud processing system provided in an embodiment of this application is shown;
[0029] Figure 3 A schematic diagram of the point cloud processing system provided in an embodiment of this application is shown;
[0030] Figure 4 A flowchart of a point cloud processing method provided in an embodiment of this application is shown. Detailed Implementation
[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0032] The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0033] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of the invention, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.
[0034] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.
[0035] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.
[0036] Most existing autonomous driving algorithms rely on CPUs or GPUs for computation, which suffers from complex data structures, massive computational loads, huge data volumes, and high power and energy consumption, making them unsuitable for direct application in embedded environments such as automobiles. Therefore, designing high-efficiency and low-power dedicated hardware accelerators is of great significance.
[0037] In practical applications, excessive processing latency or high computational and storage resource consumption can significantly limit the application areas of LiDAR point cloud processing. For example, the application requirements for point cloud target detection and point cloud localization in the field of autonomous driving technology not only involve processing large amounts of data but also have high real-time requirements. Therefore, research on how to optimize point cloud data processing to achieve high throughput and low latency is very urgent.
[0038] Typically, point cloud localization algorithms run on the CPU. See also Figure 1The point cloud localization algorithm process is as follows: First, the relevant parameters are initialized, mainly including the initial pose initialization. The current frame point cloud is then filtered, specifically using voxel filtering. Next, point cloud registration is performed. Specifically, point cloud registration includes: downsampling the real-time scanned point cloud data, discarding invalid data, performing point cloud coordinate transformation (TransformPointCloud) and gradient calculation, calculating the descent direction using Newton's method, normalizing delta_p, and performing step size search (computeStepLength). For example, with a step size of a_t, the transformation matrix p is calculated according to the formula p_t = p + delta_p * a_t. The convergence condition is then determined, and this process is iterated until the descent direction converges, outputting the optimal pose p. Finally, the obtained pose, combined with the reference map, enables point cloud localization.
[0039] In the existing solutions mentioned above, the CPU can only process point cloud data serially and cannot perform pipelined parallel processing. Therefore, when processing large-scale point cloud data, the traditional point cloud processing method on the CPU will have a large delay, which often becomes a bottleneck in system performance and cannot meet the requirements of autonomous driving applications.
[0040] Example 1
[0041] This application provides a point cloud processing system.
[0042] Please see Figure 2Point cloud processing systems include programmable logic (PL) terminals and programmable systems (PSS) terminals. The System (PS) terminal is electrically connected to the PL terminal. The PL terminal includes a hash filtering module, a gradient calculation module, and a data transport module. The hash filtering module is electrically connected to both the data transport module and the gradient calculation module. The gradient calculation module is also electrically connected to the data transport module, and the data transport module is also electrically connected to the application processor of the PS terminal. The PS terminal is used to send point cloud data calculation commands to the PL terminal. The hash filtering module is used to generate point cloud data read commands based on the point cloud data calculation commands, send the point cloud data read commands to the data transport module, receive initial point cloud data from the data transport module, and process the initial point cloud data in real time using a pipeline to obtain processed point cloud data. The gradient calculation module is used to perform gradient calculations on the processed point cloud data in parallel using a data stream through multiple calculation sub-modules connected by FIFO or ping-pong buffers to obtain multiple gradient vectors. The application processor of the PS terminal is used to calculate the descent direction of the multiple gradient vectors and determine whether the descent direction converges. If the descent direction converges, the point cloud computing is terminated.
[0043] It's important to note that a FIFO is a memory where data operates on a first-in, first-out (FIFO) basis. A ping-pong operation typically involves two memory locations, with the data stream switching between them via a selector to ensure uninterrupted data flow. Both FIFOs and ping-pong operations generally serve as data buffers between different modules to match their data processing speeds and prevent data loss. FIFOs or ping-pong operations are often used as data buffers to connect different modules. For example, data might flow from module A to the FIFO and then be output from the FIFO to module B.
[0044] To further clarify, the application processor (APU) on the PS side can also be considered an ARM CPU.
[0045] In one embodiment, the application processor at the PS end is further configured to initiate point cloud iterative calculation if the descent direction does not converge.
[0046] It should be noted that starting point cloud iterative computation can be understood as resending the point cloud data computation command, controlling the hash filtering module, gradient calculation module, and data transfer module to repeatedly perform the point cloud data positioning and computation process.
[0047] In one embodiment, the PS terminal is further configured to receive parameter configuration commands and start commands from the AXI4-Lite interface, configure the AXI4-Lite interface according to the parameter configuration commands, and generate the point cloud data calculation commands according to the start commands.
[0048] It's worth noting that the AXI4 interface has three types: AXI4-Lite, AXI4-Full, and AXI4-Stream. AXI4-Lite is a subset of AXI4-Full and is typically used for register configuration. AXI4-Full supports burst transfers and is generally used for large-scale data transfers. The AXI4-Stream interface is divided into mm2s and s2mm based on direction. mm2s indicates data is mapped from memory to a stream; typically, data is read from DDR and then output to the stream. s2mm indicates data is mapped from a stream to memory; typically, data is written to DDR from an external streaming interface.
[0049] In one embodiment, the hash filtering module is further configured to perform pipelined real-time noise reduction processing on the initial point cloud data to obtain downsampled point cloud data, and store the downsampled point cloud data in an on-chip cache.
[0050] It should be noted that the initial point cloud data volume is very large, and the hash filtering module can reduce the data volume. The hash filtering module makes full use of pipelined parallel processing in its calculation process. The hash filtering module processes the real-time scanned point cloud in a pipeline and stores the results in the on-chip cache, avoiding subsequent processing from accessing external DDR memory.
[0051] In one embodiment, the gradient calculation module is further configured to pipeline the reading of the downsampled point cloud data, perform coordinate transformation processing on the downsampled point cloud data to obtain coordinate-transformed point cloud data, acquire adjacent grid data of the coordinate-transformed point cloud data, and calculate the current pose score and gradient vector based on the coordinate-transformed point cloud data and the adjacent grid data.
[0052] It should be noted that by using multiple computational submodules connected by FIFO or ping-pong buffers to perform gradient calculations on the processed point cloud data in a parallel data flow manner, multiple gradient vectors are obtained, which greatly improves the data throughput.
[0053] In one embodiment, the gradient calculation module is further configured to obtain the point cloud data storage address of the adjacent grid based on radius search, and read the adjacent grid data from the point cloud data storage address of the adjacent grid through the AXI4-Full interface.
[0054] In one embodiment, the hash filtering module is further configured to store the processed point cloud data into an on-chip cache.
[0055] In one embodiment, the gradient calculation module is further configured to write a plurality of the gradient vectors into an on-chip cache.
[0056] It should be noted that storing the processed point cloud data and the gradient calculation results after connection into the on-chip cache reduces memory access operations to external DDR.
[0057] Please see Figure 3 The point cloud processing system operates as follows: The gradient computation (Derivatives) module is configured via the AXI4-Lite interface. This configuration includes the source data address, destination address, and the parameters used in the computation. The application processor (APU) on the PS side sends a start command via the AXI4-Lite interface to initiate the hash-filter module. It should be noted that the hash-filter module can also be called a hash-filter accelerator, and the gradient computation module can also be called a Derivatives accelerator.
[0058] The hash filtering module begins operation, sending a command to the datamover module via the `axis_mm2s_cmd` interface to read point cloud data from Double Rate Synchronous Dynamic Random Access Memory (DDR). Specifically, the command to read point cloud data is the MM2S command. The point cloud data received via the `axis_mm2s` interface is downsampled, and the downsampled point cloud is stored in an on-chip cache (BRAM) for use by the gradient calculation module (Derivatives accelerator). The BRAM can also be referred to as block memory. The datamover module then transmits the AXI4-Stream. The specific downsampling process can involve calculating the average value of point cloud data within the same voxel using a hash table, performing downsampling, and obtaining downsampled point cloud data, thus reducing the size of the point cloud data.
[0059] The application processor on the PS side sends a start command via the AXI4-Lite port to initiate the gradient calculation module (Derivatives accelerator). The gradient calculation module (Derivatives accelerator) continuously reads point cloud data from the on-chip cache and performs coordinate transformation. It then performs a radius search on the transformed points, reads the PDF (Predicted Gradient) of the adjacent grid cells from the DDR via the AXI4 interface, performs gradient calculation, and finally writes the result back to the DDR via the AXI4 interface. For details, see [link to documentation]. Figure 3 The gradient calculation module transforms a point (x, y, z) into a point (x', y', z') through matrix-vector multiplication, performs radius search, calculates the grid search address, determines the offset address of adjacent grids, and reads the PDF within the grid based on the offset address of adjacent grids. Specifically, it reads the PDF within the adjacent grids of the point from DDR via the AXI4 interface. The point gradient (x, y, z) is calculated using matrix multiplication. The parameters used in the matrix multiplication process are J_H_AngMat23×3;Cnt, where J_H_AngMat23×3;Cnt are the parameters required to calculate the Jacobian matrix and Hessian matrix. The point gradient is combined with the Jacobian matrix and Hessian matrix to calculate the Hessian gradient function and return the score. The Gaussian parameters are adjusted, and the gradient calculation is updated. The application processor (APU) on the PS side calculates the descent direction and determines whether the descent direction converges. If it does not converge, the process is iterated until convergence occurs.
[0060] It should be noted that, in Figure 3 In the middle, the PS end also includes a high performance (HP) (Slave-AXI, S-AXI) interface and a master-advanced eXtensible interface (M-AXI). The HPS-AXI interface and the AXI interface are electrically connected, and the M-AXI interface is electrically connected to the BRAM.
[0061] It should be noted that the PS has both M-AXI (Master-AXI) and S-AXI (Slave-AXI) interfaces. When the PS is acting as the master device, it uses the M-AXI interface to read data from the slave device. When the PS is acting as the slave device, the external device reads data from the PS through the S-AXI interface.
[0062] exist Figure 3 In this configuration, the HP S-AXI interface and the AXI interface are electrically connected to form a corresponding data path, indicating that the PS is currently acting as a Salve, and the accelerator reads data from the PS through this interface. The MAXI interface is electrically connected to the BRAM to form a corresponding data path, indicating that the PS is currently acting as a Master, and needs to read data stored in the BRAM.
[0063] The point cloud processing system provided in this application is a high-efficiency hardware architecture capable of point cloud localization. This hardware architecture fully utilizes pipeline technology to achieve parallel processing of point cloud data, greatly improving data throughput. The hardware architecture optimizes the data storage structure of the reference map and uses on-chip storage for key data, reducing access to external DDR memory. By using a hardware-software co-processing approach to achieve point cloud localization, computationally intensive tasks are accelerated using hardware, while control-related logic is implemented in software. This significantly improves the actual performance of the algorithm and greatly reduces latency, making it directly applicable to embedded environments such as automotive.
[0064] Example 2
[0065] This application also provides a point cloud processing method, which is applied to the point cloud processing system provided in the embodiment.
[0066] See Figure 4 The point cloud processing method includes:
[0067] Step S101: The PS terminal sends a point cloud data calculation command to the PL terminal;
[0068] In step S102, the hash filtering module generates a point cloud data reading command based on the point cloud data calculation command, sends the point cloud data reading command to the data transport module, receives the initial point cloud data from the data transport module, and processes the initial point cloud data in real time using a pipeline to obtain the processed point cloud data.
[0069] In step S103, the gradient calculation module performs gradient calculation on the processed point cloud data in a parallel data flow manner through multiple calculation sub-modules connected by FIFO or ping-pong buffer to obtain multiple gradient vectors.
[0070] In step S104, the application processor on the PS side calculates the descent direction of the multiple gradient vectors and determines whether the descent direction converges. If the descent direction converges, the termination point cloud computing is established.
[0071] In one embodiment, the method further includes:
[0072] If the descent direction does not converge, the application processor at the PS end initiates point cloud iterative calculation.
[0073] In one embodiment, the method further includes:
[0074] The PS terminal receives parameter configuration commands and start commands from the AXI4-Lite interface, configures the AXI4-Lite interface according to the parameter configuration commands, and generates the point cloud data calculation commands according to the start commands.
[0075] In one embodiment, the method further includes:
[0076] The hash filtering module uses a pipeline to perform real-time noise reduction processing on the initial point cloud data to obtain downsampled point cloud data, and stores the downsampled point cloud data in the on-chip cache.
[0077] In one embodiment, the method further includes:
[0078] The gradient calculation module pipeline reads the downsampled point cloud data, performs coordinate transformation on the downsampled point cloud data to obtain coordinate-transformed point cloud data, obtains the adjacent grid data of the coordinate-transformed point cloud data, and calculates the current pose score and gradient vector based on the coordinate-transformed point cloud data and the adjacent grid data.
[0079] In one embodiment, the method further includes:
[0080] The gradient calculation module obtains the point cloud data storage address of the adjacent grid based on radius search, and reads the adjacent grid data from the point cloud data storage address of the adjacent grid through the AXI4-Full interface.
[0081] In one embodiment, the method further includes:
[0082] The hash filtering module stores the processed point cloud data into the on-chip cache.
[0083] In one embodiment, the method further includes:
[0084] The gradient calculation module writes multiple gradient vectors into the on-chip cache.
[0085] The point cloud processing method provided in this application embodiment is applied to the point cloud processing system provided in embodiment 1, and can realize the corresponding functions of the point cloud processing system. To avoid repetition, it will not be described in detail here.
[0086] The point cloud processing method provided in this application fully utilizes pipeline technology to achieve parallel processing of point cloud data, greatly improving data throughput. The data storage structure of the reference map is optimized, and key data is stored on-chip, reducing access to external DDR memory. Point cloud localization is achieved through a hardware-software co-processing approach, accelerating computationally intensive tasks with hardware implementation and controlling related logic with software implementation. This significantly improves the actual performance of the algorithm, greatly reduces latency, and can be directly applied to embedded environments such as automotive.
[0087] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A point cloud processing system, characterized in that, It includes a PL terminal and a PS terminal, wherein the PL terminal and the PS terminal are electrically connected; The PL terminal includes a hash filtering module, a gradient calculation module, and a data transfer module. The hash filtering module is electrically connected to the data transfer module and the gradient calculation module, respectively. The data transfer module is also electrically connected to the application processor of the PS terminal. The PS terminal is used to send point cloud data calculation commands to the PL terminal; The hash filtering module is used to generate a point cloud data read command according to the point cloud data calculation command, send the point cloud data read command to the data transport module, receive initial point cloud data from the data transport module, process the initial point cloud data in real time using a pipeline to obtain processed point cloud data; it is also used to process the initial point cloud data in real time using a pipeline to obtain downsampled point cloud data, and store the downsampled point cloud data in an on-chip cache. The gradient calculation module is used to perform gradient calculations on the processed point cloud data in parallel data flow using multiple calculation sub-modules connected by FIFO or ping-pong buffers to obtain multiple gradient vectors; it is also used to pipeline the reading of the downsampled point cloud data, perform coordinate transformation processing on the downsampled point cloud data to obtain coordinate-transformed point cloud data, obtain the adjacent grid data of the coordinate-transformed point cloud data, and calculate the current pose score and gradient vector based on the coordinate-transformed point cloud data and the adjacent grid data. The application processor on the PS side is used to calculate the descent direction of multiple gradient vectors and determine whether the descent direction converges. If the descent direction converges, the termination point is cloud computing.
2. The system according to claim 1, characterized in that, The application processor at the PS end is also used to initiate point cloud iterative calculation if the descent direction does not converge.
3. The system according to claim 1, characterized in that, The PS terminal is also used to receive parameter configuration commands and start commands for the AXI4-Lite interface, configure the AXI4-Lite interface according to the parameter configuration commands, and generate the point cloud data calculation commands according to the start commands.
4. The system according to claim 1, characterized in that, The hash filtering module is also used to process the initial point cloud data in real time using a pipeline to obtain downsampled point cloud data, and store the downsampled point cloud data in an on-chip cache.
5. The system according to claim 1, characterized in that, The gradient calculation module is also used to obtain the point cloud data storage address of the adjacent grid based on the radius search, and read the adjacent grid data from the point cloud data storage address of the adjacent grid through the AXI4-Full interface.
6. The system according to claim 1, characterized in that, The hash filtering module is also used to store the processed point cloud data into an on-chip cache.
7. The system according to claim 1, characterized in that, The gradient calculation module is also used to write multiple gradient vectors into an on-chip cache.
8. A point cloud processing method, characterized in that, The method, applied to the point cloud processing system according to any one of claims 1-7, comprises: The PS terminal sends a point cloud data calculation command to the PL terminal; The hash filtering module generates a point cloud data reading command based on the point cloud data calculation command, sends the point cloud data reading command to the data transport module, receives the initial point cloud data from the data transport module, and processes the initial point cloud data in real time using a pipeline to obtain the processed point cloud data. The gradient calculation module performs gradient calculations on the processed point cloud data in a parallel data stream manner through multiple calculation sub-modules connected by FIFO or ping-pong buffers to obtain multiple gradient vectors. The application processor on the PS side calculates the descent direction of multiple gradient vectors and determines whether the descent direction converges. If the descent direction converges, the termination point is cloud computing.
9. The method according to claim 8, characterized in that, The method further includes: If the descent direction does not converge, the application processor at the PS end initiates point cloud iterative calculation.