Video image rapid cutting and dimension reduction method and system based on hierarchical bounding box

Abstract

The invention provides a video image rapid cutting and dimension reduction method and system based on a hierarchical bounding box, and is applied to the technical field of image processing, and the method comprises the steps: collecting multi-view dynamic scene point cloud data, and dividing a space through the hierarchical bounding box, so as to build a bounding box structure with a density feature; then pre-filtering with an optical diffractive array and separating the point clouds based on wavelength selectivity and density distribution to generate a first subset and a background subset; then, nodes in the first subset are modulated through adaptive pulse coding to obtain high-precision data, and nodes in the second subset are compressed into low-bit-rate data through run length coding; and inputting the two types of data into the improved k-d tree based on a bounding box spatial relationship to generate a two-dimensional point set representing key features, and constructing a spatial index to realize real-time scene retrieval. According to the method, the key geometric features of the image can be reserved while the image dimension reduction is realized, so that the efficient real-time space retrieval in a dynamic scene is improved.

Description

Technical Field

[0001] This application relates to the field of image dimensionality reduction technology, and in particular to a method and system for fast cropping and dimensionality reduction of video images based on hierarchical bounding boxes. Background Technology

[0002] In dynamic 3D scenarios such as real-time perception in autonomous driving and immersive augmented reality interaction, there is an urgent need to solve the problem of image dimensionality reduction. To address this problem, gigabyte-level 3D point clouds need to be efficiently compressed into megabyte-level 2D feature images that can be processed in real time, while preserving key information such as object edge sharpness and surface texture. At the same time, it is necessary to complete the entire chain response from data acquisition to spatial retrieval in a short time to support high-precision real-time decision-making on resource-constrained mobile platforms.

[0003] The current mainstream approach adopts a hierarchical feature distillation architecture. This architecture extracts spatiotemporal features of point clouds through a 3D convolutional neural network, uses a channel attention mechanism to filter high-value 3D points, and then projects the filtered point cloud onto a 2D feature plane through graph pooling operations. Finally, it constructs a spatial index tree to achieve fast retrieval of dimensionality-reduced images.

[0004] However, the above architecture is difficult to adapt to dynamic scene deformation due to the fixed convolution kernel, which causes the edges of moving objects to break and become distorted during the dimensionality reduction process. Furthermore, the staged feature distillation and 2D projection produce cumulative errors, causing the final 2D image to lose curvature details. In addition, graph pooling operation destroys the original spatial topology, causing structural misalignment between the dimensionality reduction image and the 3D scene. Summary of the Invention

[0005] This application provides a method and system for fast video image cropping and dimensionality reduction based on hierarchical bounding boxes, which solves the problem of poor image dimensionality reduction effect in the prior art.

[0006] Firstly, this application provides a fast video image cropping and dimensionality reduction method based on hierarchical bounding boxes, including:

[0007] Point cloud data of the image to be processed is acquired in a multi-view dynamic scene. The point cloud data is spatially divided using hierarchical bounding boxes to obtain a bounding box hierarchy structure, which consists of multiple bounding box nodes.

[0008] By pre-filtering the multi-view dynamic scene using a diffraction element array, the bounding box nodes in the bounding box hierarchy are divided into a first subset and a second subset based on wavelength selectivity and point cloud density distribution characteristics.

[0009] Adaptive pulse code modulation is used to process the bounding box nodes in the first subset to dynamically allocate the number of bits for encoding according to the curvature change rate of the point cloud, thereby obtaining encoded data. Run-length encoding is used to compress the bounding box nodes in the second subset to obtain compressed data.

[0010] Based on the spatial topological relationship of the bounding box hierarchy, the encoded data and compressed data are input into the improved kd tree for non-uniform clustering to obtain a two-dimensional feature point set.

[0011] A spatial index structure is constructed based on the spatial distribution of the two-dimensional feature point set to achieve real-time spatial retrieval of dynamic scenes from multiple perspectives.

[0012] Optionally, the bounding box nodes in the first subset are processed using adaptive pulse code modulation to dynamically allocate the number of encoding bits according to the point cloud curvature change rate, thereby obtaining encoded data. The bounding box nodes in the second subset are then compressed using run-length encoding to obtain compressed data, including:

[0013] Traverse each point within the bounding box nodes of the first subset, calculate the curvature change rate of points in the local neighborhood, and classify the points in the local neighborhood into one of a number of predefined change rate levels based on the curvature change rate.

[0014] Obtain the three-dimensional coordinate values ​​of the bounding box nodes in the first subset, scale the three-dimensional coordinate values ​​according to the allocated encoding bit length, and convert them into binary encoding of a specified bit length to form encoded data;

[0015] Obtain the 3D coordinate values ​​of the bounding box nodes in the second subset, and scan them according to the preset spatial scanning order. Identify points with the same 3D coordinate values ​​that appear consecutively as a continuous sequence, and represent the combination of continuous sequences as compressed data.

[0016] Secondly, this application provides a fast video image cropping and dimensionality reduction system based on hierarchical bounding boxes, including:

[0017] The acquisition module is used to acquire point cloud data of the image to be processed in a multi-view dynamic scene, and to spatially divide the point cloud data using hierarchical bounding boxes to obtain a bounding box hierarchy structure, which is composed of multiple bounding box nodes.

[0018] The separation module is used to pre-filter the multi-view dynamic scene through the diffraction element array, so as to divide the bounding box nodes in the bounding box hierarchy into a first subset and a second subset based on the wavelength selectivity and the distribution characteristics of point cloud density.

[0019] The first processing module is used to process the bounding box nodes in the first subset using adaptive pulse code modulation to dynamically allocate the number of encoding bits according to the point cloud curvature change rate to obtain encoded data, and to compress the bounding box nodes in the second subset using run-length encoding to obtain compressed data.

[0020] The second processing module is used to input the encoded data and compressed data into the improved kd tree for non-uniform clustering based on the spatial topological relationship of the bounding box hierarchy to obtain a two-dimensional feature point set.

[0021] The construction module is used to construct a spatial index structure based on the spatial distribution of the two-dimensional feature point set, so as to realize real-time spatial retrieval of dynamic scenes from multiple perspectives.

[0022] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a method for fast video image cropping and dimensionality reduction based on hierarchical bounding boxes as described in the first aspect above.

[0023] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a method for fast video image cropping and dimensionality reduction based on hierarchical bounding boxes as described in the first aspect.

[0024] The fast video image cropping and dimensionality reduction method based on hierarchical bounding boxes provided in this application has the following beneficial effects:

[0025] This application achieves efficient spatial organization of point cloud data by collecting multi-view dynamic scene point cloud data and dividing it into hierarchical structures using bounding boxes. This accurately preserves the geometric features and point cloud density distribution of different regions, providing a structured foundation for subsequent data separation and processing. At the same time, by using a diffraction element array for pre-filtering to separate the first subset and the second subset, it can accurately distinguish key geometric information from redundant second subsets based on wavelength selectivity and point cloud density characteristics, thereby reducing the interference of invalid data on subsequent processing and further improving the targeting of data processing.

[0026] Then, by applying adaptive pulse code modulation to the first subset and run-length encoding to the second subset, encoding resources can be dynamically allocated according to the curvature of the point cloud, and efficient compression of the second subset can be achieved while ensuring high precision of geometric details, thereby balancing data precision and storage / transmission efficiency.

[0027] Then, based on the bounding box hierarchical structure topology, the encoded data is input into an improved kd-tree for non-uniform clustering to obtain a two-dimensional feature point set. This can accurately preserve key geometric features during dimensionality reduction, thereby achieving an effective transformation from three-dimensional to two-dimensional, and further providing concise and critical feature data for subsequent retrieval. Finally, a spatial index structure is constructed based on the two-dimensional feature point set, which can realize fast spatial retrieval of dynamic scenes from multiple perspectives, thereby greatly improving retrieval efficiency and meeting the response requirements of real-time interactive scenarios.

[0028] Furthermore, this application ensures high-precision encoding of the first subset by dynamically allocating the number of bits based on the rate of curvature change, thereby accurately preserving key features such as object edges; then, run-length encoding is used to compress the second subset, which can effectively reduce the bit rate of the second subset, thereby achieving efficient data compression. Overall, while ensuring the accuracy of key geometric features, the amount of data is greatly optimized, and the processing and transmission efficiency is improved.

[0029] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 The flowchart of a fast video image cropping and dimensionality reduction method based on hierarchical bounding boxes provided in this application is shown.

[0032] Figure 2 The flowchart illustrating the generation of bounding box nodes in the first subset and the second subset provided in this application is shown.

[0033] Figure 3 This paper presents a schematic diagram of a video image fast cropping and dimensionality reduction system based on hierarchical bounding boxes provided in this application.

[0034] Figure 4 A schematic diagram of the structure of a computing device provided in this application is shown. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0036] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0037] Research indicates that image dimensionality reduction of multi-view point cloud data faces severe challenges in dynamic 3D scenarios such as real-time perception in autonomous driving and immersive augmented reality interaction. While the current mainstream hierarchical feature distillation architecture can achieve dimensionality reduction, it has significant shortcomings. For example, fixed convolutional kernels are difficult to adapt to the deformation of objects in dynamic scenes, resulting in broken and distorted edges of moving objects after dimensionality reduction. The phased feature distillation and 2D projection processes generate cumulative errors, causing the final 2D image to lose the curvature details of objects. Furthermore, graph pooling operations disrupt the original spatial topology, causing structural misalignment between the dimensionality-reduced image and the 3D scene. These problems seriously affect the accuracy and real-time performance of subsequent decisions.

[0038] To address the aforementioned issues, this application proposes a fast video image cropping and dimensionality reduction method based on hierarchical bounding boxes. The core idea of ​​this method is as follows: First, point cloud data of a multi-view dynamic scene is acquired, and spatial partitioning is performed using hierarchical bounding boxes to obtain a hierarchical structure. Then, a first subset and a second subset are separated through pre-filtering using a diffraction element array. Next, adaptive pulse code modulation is applied to the first subset to retain high-precision information, and run-length encoding is used to achieve efficient compression of the second subset. Subsequently, based on the bounding box topology, a two-dimensional feature point set is obtained through improved kd-tree clustering, and a spatial index is constructed. This application also solves the problem of fixed convolutional kernels not adapting to deformation through dynamically adaptable hierarchical partitioning, reduces accumulated errors through targeted subset separation and encoding processing to retain curvature details, and avoids structural misalignment through clustering processing based on the original topology, effectively meeting the high-precision and real-time requirements of image dimensionality reduction in dynamic scenes.

[0039] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] Figure 1 This application provides a flowchart of a method for fast video image cropping and dimensionality reduction based on hierarchical bounding boxes, as shown in the following embodiments. Figure 1 As shown, the method includes:

[0041] 101. Collect point cloud data of the image to be processed in a multi-view dynamic scene, and use hierarchical bounding boxes to spatially divide the point cloud data to obtain a bounding box hierarchy structure, which is composed of multiple bounding box nodes.

[0042] Point cloud data refers to a set of three-dimensional spatial points in a dynamic scene of an image to be processed, acquired by a multi-view sensor. Each point in this set contains corresponding coordinate information. The bounding box hierarchy refers to a tree-shaped spatial index composed of multi-level cubic containers. Each container node, also called a bounding box node or bounding box, is used to record the distribution characteristics of the internal points.

[0043] In one specific implementation, step 101 includes:

[0044] 1011. Take the point cloud data as a whole as the first-level initial bounding box node, set the minimum number of points threshold and the minimum spatial volume threshold within the initial bounding box node, calculate the spatial volume of the initial bounding box node, and count the total number of points contained within the initial bounding box node.

[0045] The initial bounding box node, also known as the root bounding box node or root bounding box, has a minimum number of points threshold, which is the minimum number of points required for a single bounding box node to stop splitting; the minimum spatial volume threshold is the minimum physical size at which the node stops splitting.

[0046] In step 1011, the point cloud data of the image to be processed acquired in the dynamic scene is used as the root node bounding box, and spatial partitioning termination conditions are set, such as: minimum point count threshold of 120 points and minimum spatial volume threshold of 0.8. Then, the spatial volume of the initial bounding box node is calculated simultaneously, such as 30m long × 25m wide × 20m high = 15000. And accurately count the total number of points in the internal point cloud, such as 128,000 points.

[0047] 1012. When the total number of points is greater than the minimum number of points threshold and the spatial volume is greater than the minimum spatial volume threshold, the points are cut along the three coordinate axes respectively. Two symmetrical spatial interface positions are generated in each coordinate axis direction. The volume ratio of the central region between the two symmetrical spatial interfaces under the same coordinate axis is calculated, and the coordinate axis direction with the smallest volume ratio of the central region is selected as the dividing direction.

[0048] Among them, the spatial interface refers to the virtual plane that cuts space along the coordinate axis direction, and the volume ratio of the central region refers to the volume ratio of the unallocated central region after symmetrical cutting.

[0049] In step 1012, when the total number of points is greater than the minimum point threshold of 120 points and the spatial volume is greater than the minimum spatial volume threshold of 0.8... At that time, the fixed cutting gap is calculated based on the bounding box level depth using the following formula. ,like: ,in, This means taking the value down to 0.01m, 0.05% means the ratio of the cutting gap to the maximum axis length, and Lmax means the maximum length of the bounding box in the X, Y, and Z coordinate axes. Therefore, Lmax = max(Lx, Ly, Lz), where Lx represents the length in the X-axis direction, Ly represents the length in the Y-axis direction, and Lz represents the length in the Z-axis direction.

[0050] and, Where Xmax represents the maximum coordinate value of the bounding box in the X-axis direction, and Xmin represents the minimum coordinate value of the bounding box in the X-axis direction; similarly... Where Ymax represents the maximum Y-axis coordinate value and Ymin represents the minimum Y-axis coordinate value; Where Zmax represents the maximum coordinate value of the Z-axis, and Zmin represents the minimum coordinate value of the Z-axis.

[0051] For example, when Xmin, Ymin, and Zmin are 0, the bounding box has the following dimensions: Lx = 30.0 − 0.0 = 30.0m, Ly = 25.0 − 0.0 = 25.0m, and Lz = 20.0 − 0.0 = 20.0m. Therefore, Lmax = max(30.0, 25.0, 20.0) = 30.0m. m.

[0052] Furthermore, the calculation generates two tangential surfaces with a cutting distance of δ at the geometric center positions of each axis, that is, two symmetrical spatial interface positions are generated in the X, Y, and Z coordinate axes, respectively:

[0053] The two tangents along the Z-axis are: and That is, the two tangents of the Z-axis are 10.0m and 10.01m;

[0054] The two tangents along the Y-axis are: and That is, the two tangents of the Y-axis are 12.5m and 12.51m;

[0055] The two tangents along the X-axis are: and That is, the two tangents of the X-axis are 15m and 15.01m;

[0056] Subsequently, the volume ratio Rn of the central region between the two symmetrical spatial interfaces on the same coordinate axis is calculated as follows: (× the product of the lengths of the other two axes) ÷ the total volume of the bounding box, where n is the subscript of R, and includes the three subscripts x, y, and z. That is:

[0057] Rx=(0.01×25×20)÷15000=5÷15000≈0.00067≈0.03%;

[0058] Ry=(0.01×30×20)÷15000=6÷15000=0.0004=0.04%;

[0059] Rz=(0.01×30×25)÷15000=7.5÷15000=0.0005=0.05%;

[0060] By comparing the percentage values, Rmin = min(0.03%, 0.04%, 0.05%) = 0.03% = Rx, the X-axis with the smallest percentage of the unallocated central area volume is finally selected as the cutting direction.

[0061] 1013. Based on the division direction, divide the region of the initial bounding box node into two sub-regions, create lower-level bounding box nodes for the sub-regions, establish a hierarchical relationship link between the lower-level bounding box nodes and the initial bounding box node, and add them to the queue to be processed.

[0062] In step 1013, based on the selected X-axis cutting direction, the X-axis range of the left child node is [Xmin, 15.0), i.e., [0.0, 15.0) m, with a length of 15.0 m; the X-axis range of the right child node is (15.01, Xmax], i.e., (15.01, 30.0], with a length of 14.99 m. Therefore, the volume of the left child node = 15.0 × 25.0 × 20.0 = 7500. The volume of the right child node is 14.99 × 25.0 × 20.0 = 7495. .

[0063] Furthermore, the points contained in the left child node must satisfy... The points contained in the right child node must satisfy the following condition: Then, a traversal check was performed on the 128,000 point cloud data of the parent node, resulting in 63,200 points in the left range and 64,800 points in the right range, totaling 128,000, which matches the number of points in the parent node. Next, bidirectional pointers were created from the parent node pointing to the left and right child nodes, since the left child node has a volume of 7500. >0.8 And the number of nodes 63200 > 120, and the volume of the right child node is 7495. >0.8 Since the number of points 64800 > 120, the recursive partitioning condition is met, and the two child nodes are finally added to the queue to be processed.

[0064] 1014. When there are unprocessed bounding box nodes in the queue to be processed, repeat the space partitioning process until there are no unprocessed bounding box nodes in the queue to be processed, and output the bounding box hierarchy structure composed of bounding box nodes of all levels.

[0065] In step 1014, based on the spatial volume and total number of points of the unprocessed bounding box nodes, for example, the left child node of the first item in the queue (63200 points, 7500... Since both the number of points and the volume exceed the corresponding threshold, the cyclic processing of steps 1012-1013 needs to be performed. This process is similar to the steps in step 1013, and will not be described again here.

[0066] For example, a recursive 7-layer mechanism is used, starting from the root node with a depth of 0, to the first split with a depth of 1, the second split with a depth of 2, and so on until the 7th split terminates, ultimately forming a bounding box tree structure.

[0067] This application can completely eliminate the problem of empty nodes generated by the traditional equal division method in non-uniform scenes by dynamically selecting the cutting direction through the strategy of minimizing the proportion of the center volume; combined with the dual threshold termination mechanism of spatial volume and point cloud number, it can accurately match the feature distribution density of the actual scene; the generated bounding box hierarchical structure improves the spatial retrieval efficiency and establishes a high-precision spatial indexing framework for subsequent density feature analysis.

[0068] 102. Pre-filtering of multi-view dynamic scenes using a diffraction element array, based on wavelength selectivity and point cloud density distribution characteristics, divides the bounding box nodes in the bounding box hierarchy into a first subset and a second subset.

[0069] Among them, the diffraction element array refers to a physical device composed of multiple independent optical processing units arranged according to a preset angle difference. It separates incident light by wavelength selective separation. Each unit corresponds to the spectral resolution of a specific viewing angle range. The first subset is a set of bounding box nodes carrying geometric contour details, and the second subset is a set of bounding box nodes carrying continuous background features.

[0070] In one specific implementation, such as Figure 2 As shown, step 102 includes:

[0071] 1021. Receive the reflected light signal from the multi-view dynamic scene and input the reflected light signal into a diffraction element array. The diffraction element array is composed of multiple diffraction units arranged with a preset angle difference. Each diffraction unit selectively decomposes the reflected light signal into a first optical path and a second optical path for independent transmission according to wavelength.

[0072] Among them, the reflected light signal refers to the collection of light wave energy reflected from the surface of an object to the acquisition device, including physical attribute data such as intensity and wavelength, which is used to characterize the original optical features of the scene; the diffraction unit refers to the smallest functional module in the array that performs spectral separation, which decomposes the mixed wavelength incident light into independent transmission channels according to preset bands; the first optical path refers to the high-frequency reflected light carrying the edge and texture of the object; the second optical path refers to the low-frequency diffuse reflected light carrying the smooth area.

[0073] In step 1021, multi-view optical signals are input, and reflected optical signals from N cameras are received, where each signal contains wavelength λ and incident angle θ. Then, the optical signal matrix L... M×N The input is a diffraction element array, where M is the number of spectral channels; then each diffraction unit modulates the phase of the incident light according to a preset tilt angle and outputs a separated optical path that satisfies the Bragg condition, such as: λ1∈[λmin, λc], where λ1 represents the wavelength of the first optical path, λmin represents the minimum wavelength of the first optical path, λc is the cutoff wavelength, λ2∈(λc, λmax], where λ2 represents the wavelength of the second optical path, λmax represents the maximum wavelength of the second optical path, and the cutoff wavelength λc can be determined by the period of the diffraction unit. In this embodiment, the size of this cutoff wavelength is not specifically limited.

[0074] 1022. Independent first processing end and second processing end are respectively set for the first optical path and the second optical path, so as to receive the light intensity distribution of the first optical path through the first processing end and generate the first image data; and to receive the phase distribution of the second optical path through the second processing end and generate the second image data.

[0075] Among them, light intensity distribution refers to the light energy intensity data per unit area, reflecting the abrupt changes in the geometric contour of the object surface; the first image data carries geometric contour details, the second image data carries continuous background features, and phase distribution refers to the spatial offset data of the light wave waveform, reflecting the uniformity and continuity of the surface material.

[0076] In step 1022, the light intensity distribution I(x, y) of the first optical path T1 is measured using a photoelectric sensor array at the first processing end, and the first image data of 1024×1024 pixels is output, which carries geometric contour details; then, the phase distribution ϕ(x, y) of the second optical path T2 is analyzed by a Michelson interferometer at the second processing end, and the second image data of size 1024×1024 is generated based on the phase distribution, which carries continuous background features.

[0077] 1023. Extract the distribution features of point cloud density inside each bounding box node. The distribution features of point cloud density include the density change range. Bounding box nodes whose density change range exceeds a preset density threshold are marked as first associated nodes. Bounding box nodes whose density change range does not exceed the preset density threshold are marked as second associated nodes.

[0078] Among them, the distribution characteristics of point cloud density refer to the statistical quantity of the number of spatial sampling units within the bounding box node.

[0079] In step 1023, the parent bounding box node with size Lx×Ly×Lz is divided into... The number of child nodes is calculated using the following formula: Where r is the subdivision granularity, representing the size reference of the child node in each direction; then the density variation amplitude of each child node is calculated, that is, the standard deviation of the point cloud density within each child node: ,in, This represents the number of point clouds within the k-th child node. This represents the local density of the j-th point within the k-th child node. This represents the average local density of all points within this child node, where the standard deviation of the point cloud density is... >τ d The time marker is designated as the first associated node. ≤τ d The time marker is the second associated node, where τ d This is a preset density change threshold.

[0080] 1024. Map the pixel clusters in the first image data whose brightness gradient exceeds a preset gradient threshold to the corresponding first associated node, map the pixel regions in the second image data with uniform phase distribution to the corresponding second associated node, and determine the set to which each bounding box node belongs as the first subset or the second subset according to the mapping relationship.

[0081] Among them, brightness gradient refers to the brightness difference rate between adjacent pixels, which is used to identify edge abrupt changes; pixel cluster mapping refers to the operation of establishing coordinate association between optical feature data and spatial bounding box nodes; and the embodiments of this application do not specifically limit the size of the preset gradient threshold.

[0082] In step 1024, the gradient map corresponding to the first image data is determined based on the convolution kernel of the Sobel operator in the x and y directions, and then the gradient map ∇G is binarized based on a preset gradient threshold; subsequently, based on the camera's intrinsic parameter matrix K, the binary image coordinates are mapped to three-dimensional space, and node binding is performed to obtain the bounding box nodes in the first subset and the bounding box nodes in the second subset.

[0083] This application can accurately identify key geometric features and background areas through a dual verification mechanism of optical hardware-level spectral separation and spatial density statistics; then, it dynamically correlates physical optical properties with point cloud spatial structure, which can eliminate the high-frequency distortion and misjudgment risks caused by traditional digital filtering, thereby establishing a highly reliable classification foundation for subsequent differentiated processing.

[0084] 103. Adaptive pulse code modulation is used to process the bounding box nodes in the first subset to dynamically allocate the number of bits for encoding according to the curvature change rate of the point cloud, thereby obtaining encoded data. Run-length encoding is used to compress the bounding box nodes in the second subset to obtain compressed data.

[0085] Among them, the rate of change of curvature refers to the rate of change of the curvature of the point cloud surface in the local neighborhood, which can be calculated from the coordinates of adjacent points and is used to characterize the geometric feature complexity.

[0086] Furthermore, the embodiments of this application do not specifically limit the specific implementation process of adaptive pulse code modulation and run-length encoding. In one specific implementation, step 103 includes:

[0087] 1031. Traverse each point within the bounding box node of the first subset, calculate the curvature change rate of the points in the local neighborhood, and classify the points in the local neighborhood into one of a number of predefined change rate levels according to the curvature change rate.

[0088] Among them, multiple rate of change levels correspond to different coding bit lengths according to the rate of curvature change from high to low, and the higher the rate of curvature change level, the more coding bit lengths are allocated.

[0089] Among them, the rate of change level refers to the predefined division of the curvature change range, and different levels correspond to different coding precision; the number of bits is the length of the binary data representing the three-dimensional coordinate value, and the higher the number of bits, the higher the coordinate precision.

[0090] In step 1031, during the traversal of each point cloud data in the first subset, a spherical detection region is first constructed with the current point as the center and a fixed radius of 0.1 meters. Then, the three-dimensional coordinates of all adjacent points within this spherical detection region are counted, and the standard deviation of these three-dimensional coordinate values ​​from the coordinates of the center point is calculated as the rate of curvature change. : ,in, Represents the coordinates of adjacent points. Let T be the average coordinates of the center point, and T be the number of adjacent points.

[0091] Subsequently, a first threshold of 0.02 meters was set based on the maximum allowable deformation tolerance of steel welding quality standards, and a second threshold of 0.01 meters was set with reference to the assembly connection accuracy specifications. The curvature change rate was then compared with three preset change rate level threshold ranges. Among them, the high change rate level σ>0.02 meters corresponds to complex surfaces such as welds, the medium change rate level 0.01 meters≤σ≤0.02 meters corresponds to medium rough surfaces such as bolts, and the low change rate level σ<0.01 meters corresponds to smooth surfaces such as flat steel plates. Then, the number of bits for encoding was allocated according to the change rate level. The high change rate point was allocated 8 bits, which can represent 256 states, the medium change rate was allocated 4 bits, which can represent 16 states, and the low change rate was allocated 2 bits, which can represent 4 states. This achieves a dynamic compression strategy where the greater the curvature change, the higher the encoding accuracy.

[0092] 1032. Obtain the three-dimensional coordinate values ​​of the bounding box nodes in the first subset, scale the three-dimensional coordinate values ​​according to the allocated encoding bit length, and convert them into binary encoding of a specified bit length to form encoded data.

[0093] Step 1032 may specifically include the following processes:

[0094] Step a1: Obtain the three-dimensional coordinate values ​​of the bounding box nodes in the first subset, wherein the three-dimensional coordinate values ​​include three component values, calculate the maximum and minimum values ​​of the three component values, and determine the scaling factor according to the correspondence between the maximum and minimum values ​​and the preset integer range.

[0095] In step a1, all points within the current bounding box node are traversed to dynamically calculate the numerical boundaries of each axis of the 3D coordinate system, and the minimum coordinate value Xmin and maximum coordinate value Xmax of the X-axis are obtained. Similarly, the minimum coordinate value Ymin and maximum coordinate value Ymax of the Y-axis and the minimum coordinate value Zmin and maximum coordinate value Zmax of the Z-axis are calculated, and the coordinate span of each axis is calculated. Similarly, ΔY and ΔZ are not elaborated here.

[0096] Step a2: Based on the scaling factor and the allocated number of bits, determine the specified number of bits corresponding to the binary code, convert the three component values ​​into integer values ​​within a preset integer range, convert the integer values ​​into binary codes of the specified number of bits, and aggregate the three binary codes to form a binary code.

[0097] The minimum value is zero, and the maximum value is the result of a power operation with base two and the number of bits allocated as the exponent minus one.

[0098] In step a2, the curvature change rate level marker for each point is assigned a number of bits for encoding. For high change rate points (σ > 0.02m), 16 bits (integer range 0-65) can be assigned; for medium change rate points (0.01m ≤ σ ≤ 0.02m), 12 bits (integer range 0-4) can be assigned; and for low change rate points (σ < 0.01m), 4 bits (integer range 0-15) can be assigned. Then, a coordinate-to-integer conversion is performed, applying a linear mapping from floating-point coordinates to integer values ​​for each point, using the scaling formula: Where a∈{x, y, z}, This represents the coordinate value of component a of the current point. To represent the minimum coordinate value of component a, represents the maximum coordinate value of component a, and b represents the number of bits used to encode the curvature level corresponding to the current point. Represents the integer value of the converted component 'a', with the sign... This indicates rounding down to the nearest integer.

[0099] Step a3: Arrange the binary codes of all bounding box nodes in the first subset in order to form encoded data.

[0100] In step a3, when generating the binary code, the integer value is... The binary sequence is converted to a specified number of bits, and zeros are padded to a fixed length in the high bits. Then, the three-axis binary codes are concatenated in the order X→Y→Z to generate a complete point code. Finally, all point codes are concatenated in the order of the spatial index of the bounding box nodes to generate coded data.

[0101] 1033. Obtain the three-dimensional coordinate values ​​of the bounding box nodes in the second subset, and scan them according to the preset spatial scanning order to identify points with the same three-dimensional coordinate values ​​that appear consecutively as a continuous sequence, and represent the combination of continuous sequences as compressed data.

[0102] Among them, spatial scanning order refers to the rule of traversing the point cloud according to a fixed path; continuous sequence refers to the set of continuous points with identical coordinate values ​​on the scanning path. The combination representation of continuous sequence includes: three-dimensional coordinate points and their corresponding occurrence counts.

[0103] In step 1033, all points within a node are arranged according to a preset spatial scanning order, namely, ascending Z-axis, ascending Y-axis, and ascending X-axis; then, the formula is used... Real-time comparison of coordinate differences between adjacent points, where, Represents the three-dimensional coordinates of the i-th point. Represents the three-dimensional coordinates of the (i-1)th point. Represents the spatial distance between two adjacent points, when Points less than a preset distance threshold are identified as identical points. Then, consecutive identical points are merged into a single compression unit, and the storage structure includes the reference coordinates and the repetition count. Finally, all compression units are concatenated in the scanning order to output compressed data.

[0104] This application enhances accuracy in key geometric feature regions by dynamically allocating the number of bits for encoding based on the rate of curvature change; subsequently, it employs spatial continuity detection compression to significantly reduce redundant storage. This dual-mode processing mechanism ensures the integrity of core features while achieving efficient reduction of the overall data volume, thus laying the foundation for real-time transmission and retrieval.

[0105] 104. Based on the spatial topological relationship of the bounding box hierarchy, the encoded data and compressed data are input into the improved kd-tree for non-uniform clustering processing to obtain a two-dimensional feature point set.

[0106] Among them, the "improved kd tree" in this application is an adaptive spatial partitioning structure specifically designed to achieve non-uniform clustering and dimensionality reduction of point clouds. Its improvement is reflected in the partitioning strategy and termination condition: First, it inherits the basic form of recursive spatial partitioning of traditional kd trees, but improves the partitioning basis, that is: for each region to be partitioned, calculate the coordinate variance of its internal point cloud in the X, Y, and Z coordinate axes, and select the direction with the largest variance as the partitioning axis for this time, so as to ensure that the partitioning is carried out along the direction with the most significant features; Secondly, multiple equally spaced candidate division positions are generated on the selected division axis, and the difference in the number of points after each candidate position divides the current region into two sub-regions is calculated. Finally, the position that minimizes the difference in the number of points is selected as the division plane, thereby achieving a balance in the number of points in the sub-regions and adapting to the non-uniform distribution of the point cloud. Finally, the division termination condition is not the traditional single point count threshold, but a combination of spatial volume and point cloud quantity judgment. That is, when the spatial volume of the sub-region is less than the preset minimum volume threshold or the number of points contained therein is less than the preset minimum point count threshold, the division stops and it is marked as the final spatial sub-unit. Then, the geometric centers of all spatial sub-units are projected to form a two-dimensional feature point set.

[0107] In one specific implementation, step 104 includes:

[0108] 1041. Based on the spatial location and spatial topology of the bounding box nodes recorded in the bounding box hierarchy, decode each encoded data into the original three-dimensional coordinate value, restore each compressed data into a continuous sequence, map the original three-dimensional coordinate value and the continuous sequence to the corresponding three-dimensional spatial coordinate position, and perform spatial topology splicing and recombination.

[0109] Spatial topology stitching refers to recombining scattered data into a complete 3D scene by utilizing the positional relationships of bounding box nodes.

[0110] In step 1041, the encoded data can be restored to the original three-dimensional coordinate values ​​using a preset coordinate restoration formula. At the same time, the compressed data is decompressed to generate a set of repeated points. Then, according to the spatial positional relationship recorded in the bounding box hierarchy, all points are stitched together to form a complete three-dimensional scene point cloud, providing basic data for subsequent processing.

[0111] 1042. After reorganization, select the coordinate axis direction with the largest variance of point cloud coordinates in the divided area, divide it into multiple candidate positions along the selected coordinate axis direction, calculate the difference in the number of points on both sides of each candidate position, select the candidate position with the smallest difference in the number of points as the final division position, and when the spatial volume of the final division position is less than a preset minimum volume threshold or the number of points in the space is less than a preset minimum number of points threshold, generate a spatial sub-unit and terminate the division.

[0112] Among them, the division region is a clustering unit whose volume is adjusted in real time according to the point density; the dispersion value refers to the degree to which the coordinate value deviates from the average value.

[0113] In step 1042, the density is calculated. Where S is the number of points and V is the partition volume, if the density d exceeds a preset gradient threshold, the partition volume is reduced; if it does not exceed the threshold, the partition volume is increased. This dynamic adjustment ensures that high-detail regions are processed finely, while background regions are efficiently covered.

[0114] Specifically, step 1042 includes:

[0115] Step b1: Calculate the arithmetic mean of the coordinate values ​​of all three-dimensional coordinate points in the X, Y, and Z directions within the divided region. Based on the arithmetic mean and the total number of coordinate values, obtain the dispersion value, and select the direction with the largest dispersion value as the selected coordinate axis direction.

[0116] In step b1, the arithmetic mean in direction a is calculated: ,in, Let represent the arithmetic mean along direction 'a', and let S represent the total number of 3D coordinate points within the currently defined region. Let represent the coordinate value of the i-th point in direction 'a'. Then, calculate the sum of the squares of the differences between each coordinate value and the arithmetic mean, divided by the total number of coordinate values, to obtain the dispersion value. The formula for calculating the dispersion value is: And select the dispersion value. The largest direction is used as the dividing direction.

[0117] Step b2: Traverse the coordinate values ​​of all three-dimensional coordinate points in the current division area along the selected coordinate axis, take the coordinate value with the smallest value as the minimum boundary value and the coordinate value with the largest value as the maximum boundary value; based on the minimum boundary value and the maximum boundary value, divide the area equally with a preset fixed spacing value to generate multiple candidate positions.

[0118] Step b3: For each candidate position in the candidate partition set, the current region is divided into a first sub-unit and a second sub-unit with the candidate position as the boundary. The number of three-dimensional coordinate points in the first sub-unit and the second sub-unit are counted respectively, and the absolute difference between the number of three-dimensional coordinate points in the two sub-units is calculated. The candidate partition plane position with the smallest absolute difference is selected as the final partition position.

[0119] In step b3, for each candidate position in the candidate partition set... ,by Divide the current region into boundaries. The first subunit and The second sub-unit, and the number of three-dimensional coordinate points in the two sub-units respectively. and And calculate the absolute difference between the two. ,choose The smallest candidate partition plane position is taken as the final partition position.

[0120] Step b4: Pre-set the minimum volume threshold and the minimum number threshold for spatial units, and define them as the minimum volume threshold and the minimum number threshold, respectively. When the volume of the final division position is less than the minimum volume threshold or the total number of three-dimensional coordinate points in the unit is less than the minimum number threshold, generate a spatial sub-unit marker and stop the division. Otherwise, divide the current region into two new regional units based on the final division position, and recursively execute the above division process for each new regional unit until all generated new regional units meet the stop division condition.

[0121] In step b4, a minimum volume threshold for the spatial unit is preset. and preset minimum quantity threshold When the final division of the spatial volume Or the number of three-dimensional coordinate points within a unit If the division fails, a spatial sub-unit marker is generated and the division stops; otherwise, the current region is divided into two new regional units based on the final division position, and the above division process is recursively executed for each new regional unit until all new regional units meet the division stopping condition.

[0122] 1043. The arithmetic mean position of all points in the spatial sub-unit is used as the geometric center, projected onto a preset two-dimensional plane to generate two-dimensional coordinate points, and the two-dimensional coordinate points of all spatial sub-units are combined to form a two-dimensional feature point set.

[0123] In step 1043, the geometric center coordinates of all points in the sub-unit where the division terminates are calculated, using the following formula: Where C represents the geometric center of the spatial sub-unit. This indicates the number of point clouds within a spatial sub-unit. , , Let C represent the three-dimensional coordinate components of the i-th point within the sub-unit. Then, project the center point C onto the horizontal plane to generate a two-dimensional coordinate point. ,in, and The x and y component coordinates of the geometric center C are given respectively, and the two-dimensional coordinates of all spatial sub-units are output. Then, they are arranged in order according to the spatial topological relationship to form a two-dimensional feature point set.

[0124] This application preserves the complete details of key geometric features through high-precision encoding, and significantly reduces the data volume by using "coordinates × number" compression storage for the background region. Subsequently, the spatial partitioning granularity is dynamically adjusted based on the point cloud density, which can achieve a balance between fine processing of high-detail areas and efficient coverage of smooth areas. Then, the three-dimensional point cloud is condensed into a two-dimensional key feature point set, which can accelerate the feature extraction process by combining non-uniform clustering mechanism while ensuring the integrity of the core structure.

[0125] 105. Construct a spatial index structure based on the spatial distribution of the two-dimensional feature point set to realize real-time spatial retrieval of dynamic scenes from multiple perspectives.

[0126] In one specific implementation, step 105 includes:

[0127] Step 1051: Using the planar region covered by the two-dimensional feature point set as the root node, divide the planar region into multiple sub-regions of equal area.

[0128] Step 1052: Count the number of two-dimensional feature points in each sub-region. If the number of two-dimensional feature points is less than or equal to a preset minimum point threshold, assign a region identifier to the corresponding sub-region, and associate the region identifier with the spatial location and spatial topology relationship of the bounding box node to construct a spatial index structure.

[0129] Among them, the root node refers to the two-dimensional boundary range of the entire scene, the preset minimum number of points threshold refers to the preset sub-region termination condition, the spatial identifier refers to the unique code that identifies the sub-region, and the bidirectional mapping relationship refers to the mutual association between the sub-region code and the spatial position and spatial topology of the 3D bounding box.

[0130] This application enables rapid target region localization driven by encoding by constructing a hierarchical spatial index structure. Subsequently, it establishes regional spatial topological relationships and optimizes resource allocation through a balancing mechanism of coarse indexes in sparse regions and fine indexes in dense regions, thereby improving the retrieval efficiency and real-time response capability of multi-view dynamic scenes.

[0131] Figure 3 This application provides a schematic diagram of the structure of a video image fast cropping and dimensionality reduction system based on hierarchical bounding boxes, as shown in the embodiment of the present application. Figure 3 As shown, the system includes:

[0132] The acquisition module 31 is used to acquire point cloud data of the image to be processed in a multi-view dynamic scene, and to spatially divide the point cloud data using hierarchical bounding boxes to obtain a bounding box hierarchical structure, wherein the bounding box hierarchical structure is composed of multiple bounding box nodes.

[0133] The separation module 32 is used to pre-filter the multi-view dynamic scene through the diffraction element array, so as to divide the bounding box nodes in the bounding box hierarchy into a first subset and a second subset based on the wavelength selectivity and the distribution characteristics of point cloud density.

[0134] The first processing module 33 is used to process the bounding box nodes in the first subset using adaptive pulse code modulation to dynamically allocate the number of encoding bits according to the curvature change rate of the point cloud to obtain encoded data, and to compress the bounding box nodes in the second subset using run-length encoding to obtain compressed data.

[0135] The second processing module 34 is used to input the encoded data and compressed data into the improved kd tree for non-uniform clustering based on the spatial topological relationship of the bounding box hierarchy to obtain a two-dimensional feature point set.

[0136] The construction module 35 is used to construct a spatial index structure based on the spatial distribution of the two-dimensional feature point set, so as to realize real-time spatial retrieval of dynamic scenes from multiple perspectives.

[0137] Figure 3 The aforementioned video image fast cropping and dimensionality reduction system based on hierarchical bounding boxes can perform... Figure 1The implementation principle and technical effects of the fast video image cropping and dimensionality reduction method based on hierarchical bounding boxes described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the fast video image cropping and dimensionality reduction system based on hierarchical bounding boxes in the above embodiments perform operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0138] In one possible design, Figure 3 The illustrated embodiment of a video image fast cropping and dimensionality reduction system based on hierarchical bounding boxes can be implemented as a computing device, such as... Figure 4 As shown, the computing device may include a storage component 41 and a processing component 42;

[0139] The storage component 41 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 42.

[0140] The processing component 42 is used for the above Figure 1 The embodiment describes a method for fast video image cropping and dimensionality reduction based on hierarchical bounding boxes.

[0141] The processing component 42 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method.

[0142] Storage component 41 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, programmable read-only memory, read-only memory, magnetic storage, flash memory, magnetic disk, or optical disk.

[0143] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0144] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0145] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0146] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0147] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The illustrated embodiment presents a method for fast video image cropping and dimensionality reduction based on hierarchical bounding boxes.

[0148] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0149] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0150] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the prior art, can be embodied in the form of software products. These computer software products can be stored in computer-readable storage media to enable a computer device (which may be a personal computer, server, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for fast video image cropping and dimensionality reduction based on hierarchical bounding boxes, characterized in that, include: Point cloud data of the image to be processed is acquired in a multi-view dynamic scene. The point cloud data is spatially divided using hierarchical bounding boxes to obtain a bounding box hierarchy structure, which consists of multiple bounding box nodes. By pre-filtering the multi-view dynamic scene using a diffraction element array, the bounding box nodes in the bounding box hierarchy are divided into a first subset and a second subset based on wavelength selectivity and point cloud density distribution characteristics. Adaptive pulse code modulation is used to process the bounding box nodes in the first subset to dynamically allocate the number of bits for encoding according to the curvature change rate of the point cloud, thereby obtaining encoded data. Run-length encoding is used to compress the bounding box nodes in the second subset to obtain compressed data. Based on the spatial topological relationship of the bounding box hierarchy, the encoded data and compressed data are input into the improved kd tree for non-uniform clustering to obtain a two-dimensional feature point set. A spatial index structure is constructed based on the spatial distribution of the two-dimensional feature point set to achieve real-time spatial retrieval of dynamic scenes from multiple perspectives.

2. The method according to claim 1, characterized in that, The pre-filtering of the multi-view dynamic scene using a diffraction element array, based on wavelength selectivity and point cloud density distribution characteristics, divides the bounding box nodes in the bounding box hierarchy into a first subset and a second subset, including: The system receives reflected light signals from the multi-view dynamic scene and inputs the reflected light signals into a diffraction element array. The diffraction element array is composed of multiple diffraction units arranged with a preset angle difference. Each diffraction unit selectively decomposes the reflected light signal into a first optical path and a second optical path for independent transmission according to wavelength. Independent first processing end and second processing end are respectively set for the first optical path and the second optical path, so that the light intensity distribution of the first optical path is received through the first processing end to generate first image data; and the phase distribution of the second optical path is received through the second processing end to generate second image data. Extract the distribution features of point cloud density inside each bounding box node. The distribution features of point cloud density include the density change amplitude. Bounding box nodes whose density change amplitude exceeds a preset density threshold are marked as first associated nodes. Bounding box nodes whose density change amplitude does not exceed the preset density threshold are marked as second associated nodes. In the first image data, pixel clusters with brightness gradients exceeding a preset gradient threshold are mapped to the corresponding first associated nodes. In the second image data, pixel regions with uniform phase distribution are mapped to the corresponding second associated nodes. Based on the mapping relationship, the set to which each bounding box node belongs is determined to be either the first subset or the second subset.

3. The method according to claim 1, characterized in that, The bounding box nodes in the first subset are processed using adaptive pulse code modulation to dynamically allocate the number of encoding bits according to the point cloud curvature change rate, thereby obtaining encoded data. The bounding box nodes in the second subset are then compressed using run-length encoding to obtain compressed data, including: Traverse each point within the bounding box nodes of the first subset, calculate the curvature change rate of points in the local neighborhood, and classify the points in the local neighborhood into one of a number of predefined change rate levels based on the curvature change rate. Obtain the three-dimensional coordinate values ​​of the bounding box nodes in the first subset, scale the three-dimensional coordinate values ​​according to the allocated encoding bit length, and convert them into binary encoding of a specified bit length to form encoded data; Obtain the 3D coordinate values ​​of the bounding box nodes in the second subset, and scan them according to the preset spatial scanning order. Identify points with the same 3D coordinate values ​​that appear consecutively as a continuous sequence, and represent the combination of continuous sequences as compressed data.

4. The method according to claim 1, characterized in that, Based on the spatial topological relationship of the bounding box hierarchy, the encoded data and compressed data are input into an improved kd-tree for non-uniform clustering processing to obtain a two-dimensional feature point set, including: Based on the spatial location and spatial topology of the bounding box nodes recorded in the bounding box hierarchy, each encoded data is decoded into the original three-dimensional coordinate value, each compressed data is restored into a continuous sequence, the original three-dimensional coordinate values ​​and the continuous sequence are mapped to the corresponding three-dimensional spatial coordinate positions, and spatial topology splicing and recombination are performed. After reorganization, the coordinate axis direction with the largest variance of point cloud coordinates within the division area is selected, and the area is divided into multiple candidate positions along the selected coordinate axis direction. The difference in the number of points on both sides of each candidate position is calculated, and the candidate position with the smallest difference in the number of points is selected as the final division position. When the spatial volume of the final division position is less than a preset minimum volume threshold or the number of points in the space is less than a preset minimum number of points threshold, a spatial sub-unit is generated and the division is terminated. The arithmetic mean of all points within the spatial sub-unit is used as the geometric center, projected onto a preset two-dimensional plane to generate two-dimensional coordinate points, and the two-dimensional coordinate points of all spatial sub-units are combined to form a two-dimensional feature point set.

5. The method according to claim 1, characterized in that, Constructing a spatial index structure based on the spatial distribution of the two-dimensional feature point set includes: Using the planar region covered by the two-dimensional feature point set as the root node, the planar region is divided into multiple sub-regions of equal area; The number of two-dimensional feature points in each sub-region is counted. If the number of two-dimensional feature points is less than or equal to a preset minimum point threshold, a region identifier is assigned to the corresponding sub-region. The spatial location and spatial topology relationship between the region identifier and the bounding box node are associated to construct a spatial index structure.

6. The method according to claim 1, characterized in that, Point cloud data is spatially partitioned using hierarchical bounding boxes to obtain a bounding box hierarchy. This hierarchy consists of multiple bounding box nodes, including: The point cloud data as a whole is used as the initial bounding box node of the first level. The minimum number of points and the minimum spatial volume threshold within the initial bounding box node are set. The spatial volume of the initial bounding box node is calculated, and the total number of points contained within the initial bounding box node is counted. When the total number of points is greater than the minimum number of points and the spatial volume is greater than the minimum spatial volume threshold, the points are cut along the three coordinate axes respectively. Two symmetrical spatial interface positions are generated in each coordinate axis direction. The volume ratio of the central region between the two symmetrical spatial interfaces under the same coordinate axis is calculated, and the coordinate axis direction with the smallest volume ratio of the central region is selected as the dividing direction. Based on the division direction, the region of the initial bounding box node is divided into two sub-regions. Sub-level bounding box nodes are created for the sub-regions. A hierarchical relationship link is established between the sub-level bounding box nodes and the initial bounding box node, and the nodes are added to the queue to be processed. If there are unprocessed bounding box nodes in the queue, repeat the space partitioning process until there are no unprocessed bounding box nodes in the queue, and output the bounding box hierarchy structure composed of bounding box nodes of all levels.

7. The method according to claim 3, characterized in that, Obtain the 3D coordinate values ​​of the bounding box nodes in the first subset, scale and convert the 3D coordinate values ​​into binary code of a specified number of bits according to the allocated encoding bit width, forming encoded data, including: Obtain the three-dimensional coordinate values ​​of the bounding box nodes in the first subset, wherein the three-dimensional coordinate values ​​include three component values, calculate the maximum and minimum values ​​of the three component values, and determine the scaling factor based on the correspondence between the maximum and minimum values ​​and a preset integer range; Based on the scaling factor and the allocated number of bits, determine the specified number of bits corresponding to the binary code; The three component values ​​are converted into integer values ​​within a preset integer range, and the integer values ​​are converted into binary codes of a specified number of bits. The three binary codes are then combined to form a binary code. Arrange the binary codes of all bounding box nodes in the first subset in order to form encoded data.

8. The method according to claim 4, characterized in that, Select the coordinate axis direction with the largest variance of the point cloud coordinates within the partitioned area, and divide the area into multiple candidate locations along the selected coordinate axis direction, including: Calculate the arithmetic mean of the coordinate values ​​of all three-dimensional coordinate points in the X, Y, and Z directions within the divided region. Based on the arithmetic mean and the total number of coordinate values, obtain the dispersion value, and select the direction with the largest dispersion value as the selected coordinate axis direction. Traverse the coordinate values ​​of all three-dimensional coordinate points in the selected coordinate axis direction within the current partitioned region, and take the coordinate value with the smallest value as the minimum boundary value and the coordinate value with the largest value as the maximum boundary value; Based on the minimum and maximum boundary values, multiple candidate positions are generated by dividing the area equally with a preset fixed spacing value.

9. A video image fast cropping and dimensionality reduction system based on hierarchical bounding boxes, characterized in that, include: The acquisition module is used to acquire point cloud data of the image to be processed in a multi-view dynamic scene, and to spatially divide the point cloud data using hierarchical bounding boxes to obtain a bounding box hierarchy structure, which is composed of multiple bounding box nodes. The separation module is used to pre-filter the multi-view dynamic scene through the diffraction element array, so as to divide the bounding box nodes in the bounding box hierarchy into a first subset and a second subset based on the wavelength selectivity and the distribution characteristics of point cloud density. The first processing module is used to process the bounding box nodes in the first subset using adaptive pulse code modulation to dynamically allocate the number of encoding bits according to the point cloud curvature change rate to obtain encoded data, and to compress the bounding box nodes in the second subset using run-length encoding to obtain compressed data. The second processing module is used to input the encoded data and compressed data into the improved kd tree for non-uniform clustering based on the spatial topological relationship of the bounding box hierarchy to obtain a two-dimensional feature point set. The construction module is used to construct a spatial index structure based on the spatial distribution of the two-dimensional feature point set, so as to realize real-time spatial retrieval of dynamic scenes from multiple perspectives.

10. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a method for fast video image cropping and dimensionality reduction based on hierarchical bounding boxes as described in any one of claims 1 to 8.

Images