Adaptive quantization method considering distance range for lidar-based point cloud compression

Abstract

An adaptive quantization method considering a distance range for LiDAR-based point cloud compression is provided. According to an embodiment of the present invention, the adaptive quantization method classifies point cloud data on the basis of a distance from an origin, and quantizes the classified point cloud data using different quantization schemes. Accordingly, by performing adaptive quantization that takes the distance range of point cloud data into consideration, rounding errors and clipping errors during quantization are significantly reduced, thereby enabling data compression and integer conversion while minimizing data loss and accuracy degradation.

Description

Adaptive quantization method considering distance range for LiDAR-based point cloud compression

[0001] The present invention relates to data quantization, and more specifically, to a method for quantizing point cloud data acquired from a LiDAR sensor to lighten deep learning computations.

[0002] Point cloud data acquired by a LiDAR sensor typically has an FP32 data type (4 Bytes). However, in embedded hardware, floating-point operations are more complex and require a larger amount of computation compared to integer operations, and require four times the memory space compared to INT8 data (1 Byte). Accordingly, data compression and integer conversion through quantization are essential.

[0003] In addition, unlike 2D image data, point cloud data generated by a LiDAR sensor contains geometric distance information, such as the x, y, and z coordinates of surrounding objects, rather than visual information in the range of 0 to 255. Therefore, traditional quantization methods used in 2D image processing can cause significant performance degradation when applied to 3D point cloud data.

[0004] In particular, for point cloud data of longer distances rather than shorter distances, significant loss occurs due to clipping errors during quantization.

[0005] The present invention has been devised to solve the above-mentioned problems, and the objective of the present invention is to provide an adaptive quantization method and apparatus that considers a distance range as a means to minimize data loss and accuracy loss during the quantization of point cloud data.

[0006] An adaptive quantization method according to an embodiment of the present invention for achieving the above objective comprises: a step of acquiring point cloud data; a step of calculating a distance from an origin for the acquired point cloud data; a step of classifying the point cloud data based on the calculated distance; a step of quantizing the classified point cloud data in different ways; and a step of integrating the quantized point cloud data.

[0007] The quantization step may involve quantizing the classified point cloud data by applying different quantization parameters (scale, zero_point).

[0008] The classification step may involve classifying point cloud data based on whether the calculated distance is below a threshold.

[0009] The quantization step may involve quantizing point cloud data where the calculated distance is greater than or equal to a threshold value by subtracting specific values ​​from the coordinate values.

[0010] The acquisition phase may involve acquiring point cloud data from the LiDAR sensor.

[0011] The adaptive quantization method according to the present invention further comprises the step of grouping acquired point cloud data into a grid shape; and the calculation step, classification step, quantization step, and integration step may be performed on the grouped point cloud data.

[0012] The grid shape can be a voxel or a filler.

[0013] The adaptive quantization method according to the present invention further comprises the step of passing a feature map, composed of point cloud data grouped into a quantized grid shape, through a convolutional layer; and the integration step may be to integrate the feature maps that have passed through the convolutional layer.

[0014] The integrated feature map can be used as input for a 3D object recognition model.

[0015] According to another aspect of the present invention, an adaptive quantization device is provided, characterized by comprising: an acquisition unit for acquiring point cloud data; a classification unit for the acquired point cloud data, which calculates a distance from an origin and classifies the point cloud data based on the calculated distance; a quantization unit for the classified point cloud data, which quantizes them in different ways; and an integration unit for integrating the quantized point cloud data.

[0016] According to another aspect of the present invention, an object recognition method is provided, characterized by comprising: a step of classifying point cloud data based on distance from an origin; a step of quantizing the classified point cloud data in different ways; a step of integrating the quantized point cloud data; and a step of recognizing an object based on the integrated point cloud data.

[0017] According to another aspect of the present invention, an object recognition device is provided, characterized by comprising: a classification unit that classifies point cloud data based on distance from an origin; a quantization unit that quantizes the classified point cloud data in different ways; an integration unit that integrates the quantized point cloud data; and a recognition unit that recognizes an object based on the integrated point cloud data.

[0018] As explained above, according to the embodiments of the present invention, through adaptive quantization considering distance ranges for point cloud data, rounding error and clipping error during quantization are drastically reduced, making it possible to compress and convert data into integers while minimizing data loss and accuracy loss.

[0019] FIG. 1 is an adaptive quantization device according to one embodiment of the present invention,

[0020] FIG. 2 is a distance-based adaptive quantization method according to another embodiment of the present invention,

[0021] Figure 3 is an example of distance-based classification of point cloud data,

[0022] Figure 4 is a quantization formula for point cloud data,

[0023] FIG. 5 is an adaptive quantization device according to another embodiment of the present invention,

[0024] FIG. 6 is a distance-based adaptive quantization method according to another embodiment of the present invention.

[0025] The present invention will be described in more detail below with reference to the drawings.

[0026] An embodiment of the present invention presents an adaptive quantization method for point cloud data that considers a distance range. This is a point cloud INT8 quantization method applicable to embedded systems, which minimizes data loss and accuracy loss caused by quantization by adaptively adjusting quantization parameters according to distance while considering the characteristics of the point cloud data.

[0027] FIG. 1 is a diagram illustrating the configuration of an adaptive quantization device according to an embodiment of the present invention. As illustrated, the adaptive quantization device according to an embodiment of the present invention comprises a data acquisition unit (110), a data classification unit (120), a data quantization unit (130), and a data integration unit (140).

[0028] The data acquisition unit (110) acquires point cloud data from the lidar sensor. The point cloud data generated by the lidar sensor displays position information in meters along the x, y, and z axes from the origin to the target point, and while the range of values ​​varies depending on the sensor, it has a range of approximately -40 to 40 meters.

[0029] The data classification unit (120) calculates the distance from the origin for the point cloud data obtained by the data acquisition unit (110) and classifies the point cloud data based on the calculated distance.

[0030] The data quantization unit (130) performs quantization in different ways on the point cloud data classified by the data classification unit (120). The data integration unit (140) integrates the point cloud data quantized in different ways by the data quantization unit (130).

[0031] The adaptive quantization process by the presented device will be described in detail below with reference to FIG. 2. FIG. 2 is a diagram illustrating the flow of a distance-based adaptive quantization method according to another embodiment of the present invention.

[0032] As described above, first, the data acquisition unit (110) acquires point cloud data from the lidar sensor (S210). The acquired point cloud data is FP32 data.

[0033] Then, the data classification unit (120) calculates the distance from the origin for the point cloud data obtained in step S210 (S220). The distance is calculated using the Euclidean distance formula.

[0034] The next data classification unit (120) classifies point cloud data based on the distance calculated in step S220 (S230). Specifically, as shown in FIG. 3, 1) if the calculated distance (d) is less than the threshold value (D) (S230-Yes), it is classified as point cloud data close to the origin, and 2) if the calculated distance (d) is greater than or equal to the threshold value (D) (S230-No), it is classified as point cloud data far from the origin.

[0035] Then, the data quantization unit (130) performs quantization in different ways on the point cloud data classified through step S230 (S240, S250). Specifically, 1) for point cloud data close to the origin, quantization parameters for near-field data quantization (sclae s , zero_point s 1) Quantize by applying ), and 2) for point cloud data far from the origin, quantization parameters for far-distance data quantization (sclae l , zero_point l Quantization is performed by applying ). Through steps S240 and S250, the point cloud data becomes INT8 data.

[0036] Figure 4 shows quantization formulas for point cloud data near the origin and point cloud data far from the origin. According to Figure 4, for point cloud data far from the origin, the coordinate value x i , y i , z i For a specific value x d , y d , z d It can be seen that quantization is performed by excluding a specific value x. d , y d , z d The value of the point between the point and the origin that is at a distance of D can be calculated or a specific absolute value can be used. The reason for doing this is to reduce the rounding error during quantization and the clipping error that occurs when clipping values ​​outside the quantization range by scaling down the range of the wide-range point cloud data.

[0037] Afterwards, the data integration unit (140) integrates the quantized point cloud data from step S240 and the quantized point cloud data from step S250 to complete the quantized point cloud data (S260).

[0038] FIG. 5 is a diagram illustrating the configuration of an adaptive quantization device according to another embodiment of the present invention. The adaptive quantization device according to an embodiment of the present invention can be used to extract and quantize feature maps of point cloud data by integrating with the preprocessing process of a LiDAR-based 3D object recognition model.

[0039] An adaptive quantization device according to an embodiment of the present invention is configured to include, as illustrated, a data acquisition unit (310), a voxelization unit (320), a voxel classification unit (330), a voxel quantization unit (340), a convolution operation unit (350), and a voxel integration unit (360).

[0040] The data acquisition unit (310) acquires point cloud data from the lidar sensor. The voxelization unit (320) voxels the point cloud data acquired by the data acquisition unit (310).

[0041] The voxel classification unit (330) calculates the distance from the origin for the voxels generated by the voxelization unit (320) and classifies the voxels based on the calculated distance. The voxel quantization unit (340) performs quantization in different ways for the voxels classified by the voxel classification unit (330).

[0042] The convolution operation unit (350) passes a feature map composed of voxels quantized by the voxel quantization unit (340) through a 1D convolution layer. The voxel integration unit (360) integrates the feature maps that have passed through the 1D convolution layer.

[0043] FIG. 6 is a diagram illustrating the flow of a distance-based adaptive quantization method according to another embodiment of the present invention.

[0044] As described above, first, the data acquisition unit (310) acquires point cloud data from the lidar sensor (S410). The acquired point cloud data is FP32 data.

[0045] Then, the voxelization unit (320) voxels the point cloud data obtained in step S410 and converts it into a grid-shaped feature map (S420). The voxel classification unit (330) calculates the distance from the origin for the voxels generated in step S420 (S430).

[0046] The next voxel classification unit (330) classifies the voxels based on the distance calculated in step S430 (S440). Specifically, 1) if the calculated distance (d) is less than the threshold value (D) (S440-Yes), it is classified as a voxel close to the origin, and 2) if the calculated distance (d) is greater than or equal to the threshold value (D) (S440-No), it is classified as a voxel far from the origin.

[0047] Then, the voxel quantization unit (340) performs quantization in different ways for the voxels classified through step S440 (S450, S460). Specifically, 1) for voxels close to the origin, quantization parameters for near-distance voxel quantization (sclae s , zero_point s 1) Quantize by applying ), and 2) for voxels far from the origin, quantization parameters for far-distance voxel quantization (sclae l , zero_point l Quantization is performed by applying ). Meanwhile, for voxels far from the origin, quantization is performed by subtracting a specific value from the coordinate values. Through steps S450 and S460, the voxels become INT8 data.

[0048] The convolution operation unit (350) 1) passes a feature map composed of quantized near-distance voxels from step S450 through a 1D convolution layer (S470), and 2) passes a feature map composed of quantized far-distance voxels from step S460 through a 1D convolution layer (S480). This is to allow the convolution layer to learn relative distance information between near and far distances that is lost due to distance-based quantization for the INT8 type feature map.

[0049] Subsequently, the voxel integration unit (360) integrates (concatenates) the feature maps that have passed through the 1D convolutional layer in steps S470 and S480 to complete the quantized feature maps (voxels) (S490). The feature maps completed in step S490 are input into a 3D object recognition model and used for 3D object recognition.

[0050] Meanwhile, in the above embodiment, point cloud data were grouped into voxels, but it is also possible to group them into pillars. That is, point cloud data can be grouped into various grid shapes to enable distance-based adaptive quantization.

[0051] Up to now, an adaptive quantization method considering a distance range for LiDAR-based point cloud compression has been described in detail with reference to preferred embodiments.

[0052] In the above embodiment, through adaptive quantization considering the distance range for point cloud data, rounding error and clipping error during quantization are drastically reduced, enabling data compression and integer conversion while minimizing data loss and accuracy loss.

[0053] The adaptive quantization method and apparatus according to an embodiment of the present invention can be universally applied when INT8 quantization of point cloud data is required in an embedded system.

[0054] Furthermore, it is possible to implement an object recognition method or device by combining an object recognition model with the method and device according to the embodiment of the present invention, and it goes without saying that the technical concept of the present invention can also be applied in this case.

[0055] Meanwhile, it goes without saying that the technical concept of the present invention may also be applied to a computer-readable recording medium containing a computer program that enables the device and method according to the present embodiment to perform their functions. Furthermore, the technical concept according to various embodiments of the present invention may be implemented in the form of computer-readable code recorded on a computer-readable recording medium. A computer-readable recording medium may be any data storage device that can be read by a computer and store data. For example, a computer-readable recording medium may be a ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical disk, hard disk drive, etc. Additionally, computer-readable code or a program stored on a computer-readable recording medium may be transmitted through a network connected between computers.

[0056] Furthermore, although preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above. Various modifications are possible by those skilled in the art without departing from the essence of the invention as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present invention.

Claims

1. Step of acquiring point cloud data; For the acquired point cloud data, a step of calculating the distance from the origin; A step of classifying point cloud data based on calculated distances; A step of quantizing classified point cloud data in different ways; An adaptive quantization method characterized by including a step of integrating quantized point cloud data.

2. In Claim 1, The quantization step is, An adaptive quantization method characterized by quantizing classified point cloud data by applying different quantization parameters (scale, zero_point).

3. In Claim 2, The classification stage is, An adaptive quantization method characterized by classifying point cloud data based on whether the calculated distance is less than a threshold.

4. In Claim 3, The quantization step is, An adaptive quantization method characterized by quantizing point cloud data where the calculated distance is greater than or equal to a threshold value by subtracting specific values ​​from the coordinate values.

5. In Claim 1, The acquisition stage is. An adaptive quantization method characterized by acquiring point cloud data from a lidar sensor.

6. In Claim 1, The method further includes the step of grouping the acquired point cloud data into a grid shape, and The computation step, classification step, quantization step, and integration step are, An adaptive quantization method characterized by performing on grouped point cloud data.

7. In Claim 6, The grid pattern is, An adaptive quantization method characterized by being a voxel or a filler.

8. In Claim 6, The method further includes the step of passing a feature map, composed of point cloud data grouped into a quantized grid shape, through a convolutional layer; The integration phase is, An adaptive quantization method characterized by integrating feature maps that have passed through a convolutional layer.

9. In Claim 8, The integrated feature map is, An adaptive quantization method characterized by being used as input to a 3D object recognition model.

10. Acquisition unit for acquiring point cloud data; A classification unit that calculates the distance from the origin for the acquired point cloud data and classifies the point cloud data based on the calculated distance; A quantization unit that quantizes classified point cloud data in different ways; An adaptive quantization device characterized by including an integration unit that integrates quantized point cloud data.

11. A step of classifying point cloud data based on distance from the origin; A step of quantizing classified point cloud data in different ways; Step of integrating quantized point cloud data; An object recognition method characterized by including a step of recognizing an object based on integrated point cloud data.

12. A classification unit that classifies point cloud data based on distance from the origin; A quantization unit that quantizes classified point cloud data in different ways; Integration unit for integrating quantized point cloud data; An object recognition device characterized by including a recognition unit that recognizes an object based on integrated point cloud data.

Images