A three-dimensional target detection method, medium and system

Abstract

The present invention discloses a three-dimensional target detection method, medium and system, the method comprises the steps of: S1, obtaining the original point cloud, and downsampling the original point cloud layer by layer, extracting point-by-point features of different scales; S2, upsampling the point-by-point features of different scales, giving each feature point high-level geometric structure information, and obtaining updated point cloud data; S3, encoding the updated point cloud data with features, and obtaining preliminary features; S4, generating a bird's-eye view of the preliminary features to obtain a preliminary feature map, and then using a regional Transformer block and a moving region strategy to extract features on the preliminary feature map, constructing a multi-scale feature pyramid, and fusing information from different scales through top-down and bottom-up feature fusion methods to obtain a final multi-scale feature map; S5, performing target detection based on the multi-scale feature map. The present invention has the advantages of accurate and reliable detection.

Description

Technical Field

[0001] The present invention mainly relates to the field of autonomous driving technology, and specifically to a three-dimensional target detection method, medium and system. Background Art

[0002] In recent years, with the rapid development of deep learning, laser point cloud target detection networks based on deep learning have gradually become a key technology for realizing three-dimensional detection of vehicles, pedestrians, etc. In traditional three-dimensional target detection methods, the processing of point cloud data has always been a challenging problem. To address this problem, new methods and technologies are constantly proposed in this field to improve the accuracy and efficiency of detection.

[0003] As a cutting-edge field in the modern automobile industry, autonomous driving technology is developing rapidly and gradually changing the transportation system. Its core lies in simulating the operating behavior of human drivers through advanced electronic technology and artificial intelligence algorithms to improve driving safety and efficiency. In this process, environmental perception technology plays a vital role, which is the premise for the decision-making of the autonomous driving system.

[0004] Environmental perception relies on a variety of on-board sensors, including passive sensors such as cameras, and active sensors such as laser radar (LiDAR). LiDAR is particularly important in autonomous driving due to its high accuracy and strong environmental adaptability. It generates three-dimensional point cloud data of the surrounding environment by emitting lasers and receiving reflected signals, providing accurate spatial information for the vehicle.

[0005] However, how to effectively process these point cloud data, extract useful information, and perform target detection and recognition based on this is a key technical challenge facing the field of autonomous driving. Traditional target detection methods, such as those based on segmentation and clustering, have limited generalization capabilities due to their reliance on manually designed features. In contrast, deep learning-based methods can automatically learn the deep features of data and show better performance and robustness. To address this issue, new methods and technologies are constantly being proposed in this field to improve the accuracy and efficiency of detection.

[0006] The application of deep learning in 3D object detection can be divided into several categories: point set-based methods directly process the original point cloud and can retain complete geometric information; voxel-based methods convert point clouds into regular grids for easy processing by convolutional neural networks; in addition, Transformer-based models use self-attention mechanisms to process point cloud data, demonstrating the ability to handle complex relationships. Although these methods have their own advantages, they still face challenges in practical applications, especially when dealing with complex scenes such as occlusion and truncation. Summary of the invention

[0007] In view of the technical problems existing in the prior art, the present invention provides a method, medium and system for accurately and reliably detecting three-dimensional targets.

[0008] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0009] A three-dimensional target detection method comprises the steps of:

[0010] S1, obtain the original point cloud data, and downsample the original point cloud data layer by layer to extract point-by-point features of different scales;

[0011] S2, upsampling the point-by-point features of different scales, giving each feature point high-level geometric structure information, and obtaining updated point cloud data;

[0012] S3, feature encoding the updated point cloud data to obtain preliminary features;

[0013] S4, generate a bird's-eye view of the preliminary features, and generate a preliminary feature map based on the bird's-eye view, then use the regional Transformer block and the moving region strategy to extract features on the preliminary feature map, build a multi-scale feature pyramid, and fuse information from different scales through top-down and bottom-up feature fusion methods to obtain the final multi-scale feature map;

[0014] S5. Target detection based on multi-scale feature maps.

[0015] Preferably, the specific process of step S4 is:

[0016] S41, generating a bird's-eye view of the preliminary features, and then generating an original feature map;

[0017] S42, normalizing the dimension of the original feature map to the number of channels C;

[0018] S43, down-sampling the standardized feature map to generate a preliminary feature map;

[0019] S44, after generating the preliminary feature map, the regional Transformer block is used to perform self-attention calculation on the feature map to extract features;

[0020] S45. After feature extraction is completed, a strategy of fusing feature pyramid network and path enhancement network is adopted to construct a multi-scale feature pyramid; through top-down and bottom-up feature fusion methods, at each layer of the feature pyramid, a multi-layer perceptron is used to unify the number of channels C to a preset dimension, and then splicing and upsampling are performed to obtain fused multi-scale features;

[0021] S46. Input the fused multi-scale features into the multi-head region proposal network to perform preliminary classification and regression tasks.

[0022] Preferably, in step S44, a moving region strategy is introduced, and after the initial region self-attention calculation is completed, these regions are moved and the self-attention is recalculated to capture more surrounding context information.

[0023] Preferably, the specific process of step S1 is: sending the original point cloud data into the point-based three-dimensional feature encoding network to extract the features of the original information of the point cloud, sampling and combining the input original point cloud (N, D) at the farthest point, and downsampling layer by layer using the PointNet network to generate point-by-point features of different scales; wherein N represents the number of point clouds, and D represents the dimension of the point cloud, including the coordinates of the x-axis, y-axis, and z-axis, and the reflection intensity.

[0024] Preferably, in step S1, a multi-scale grouping strategy is adopted to take concentric spheres of different sizes, obtain multiple local neighborhood features with the same center but different scales, and then splice them to capture spatial information from local to global.

[0025] Preferably, the specific process of step S3 is: sending the updated point cloud data into a voxel-based three-dimensional feature encoding network, and using three-dimensional sparse convolution for feature encoding. The specific operation is: setting the size of each input voxel, and alternately using submanifold convolution and three-dimensional sparse convolution for feature encoding.

[0026] Preferably, in step S5, a random point cloud elimination operation is performed on the selected point cloud subset within the target true value bounding box.

[0027] Preferably, the specific process of the random point cloud elimination operation is:

[0028] First, the true bounding box of the target is divided into a blocks, and all point clouds of one of the blocks are randomly selected and deleted to simulate the situation where the object is partially occluded; where a is a natural number, a≥4;

[0029] Secondly, the true bounding box is randomly divided into b blocks above and below or left and right, and all point clouds of a block in the divided area are randomly deleted to simulate the situation where the object is truncate; where b is a natural number, b≥2.

[0030] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described above are executed.

[0031] The present invention further discloses a three-dimensional target detection system, comprising a memory and a processor connected to each other, wherein a computer program is stored in the memory, and when the computer program is run by the processor, the steps of the method described above are executed.

[0032] Compared with the prior art, the advantages of the present invention are:

[0033] The present invention is based on the point and voxel three-dimensional feature coding layer. It combines the point-based three-dimensional feature extraction method with the voxel-based three-dimensional sparse convolutional network for the first time, deeply integrates the original features of the point cloud and the characteristics of voxel fast feature coding, summarizes multi-scale local point cloud information, makes full use of the fine-grained features and information of the original point cloud, and focuses on local point cloud features. This network structure can deeply mine the original features of point cloud data, especially for small-sized objects, and significantly improves the accuracy and reliability of detection.

[0034] The two-dimensional feature encoding network constructed by the present invention integrates the Transformer of the moving area and the FPN-PAN feature pyramid. By performing self-attention calculations in different downsampled feature map areas and motion areas, the associated communication with surrounding information is established through sliding, and rich scene context information is encoded at multiple scales; then, the FPN-PAN layer is used to splice information of different scales and send it to the multi-head RPN to better extract targets of different scales, improve the three-dimensional target detection performance, and be more suitable for downstream tasks such as target detection. Through the self-attention mechanism and multi-scale feature fusion, the encoding of local features and global context information is strengthened, and the detection performance of the model for targets of different scales is improved, especially the ability to recognize targets in complex scenes.

[0035] The present invention is based on the moving region Transformer and feature pyramid, and aims to achieve detection accuracy comparable to that of current advanced models by deeply fusing the original features of point clouds with the advantages of fast feature encoding of voxels, while making full use of bird's-eye view features. Through a series of innovative designs, while improving detection accuracy, the model's ability to adapt to complex environments is also enhanced, contributing new strength to the development of autonomous driving technology.

[0036] The data enhancement method based on random point cloud elimination of the present invention simulates incomplete point cloud data, randomly removes point clouds of different sizes and positions from some truth boxes, and trains the model to learn to infer complete object shape information from partial information, thereby improving the model's adaptability to occlusion and truncation problems in the real world, thereby significantly improving the generalization ability and robustness of the model, making it more stable in the face of complex environments and improving the three-dimensional detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The flowchart of the three-dimensional object detection method in an embodiment of the present invention.

[0038] Figure 2Schematic diagram of data enhancement based on random point cloud elimination in the present invention; (a), (c), and (e) are schematic diagrams of simulated objects being partially occluded; (b), (d), and (f) are schematic diagrams of simulated objects being truncated.

[0039] Figure 3 Schematic diagram of a non-overlapping region with a size of m×n in the present invention.

[0040] Figure 4 Schematic diagram of the structure of the multi-scale feature pyramid in the present invention. DETAILED DESCRIPTION

[0041] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0042] like Figure 1 As shown, the three-dimensional target detection method provided by the embodiment of the present invention has the following specific steps:

[0043] S1, data input and feature extraction;

[0044] In order to effectively utilize the original fine-grained information of the point cloud, the original point cloud is sent to a point-based three-dimensional feature encoding network (such as PointNet) to extract the features of the original point cloud information. The specific operation is to sample and combine the input point cloud (N, D) at the farthest point, and use the PointNet network to downsample layer by layer to generate point-by-point features of different scales; N represents the number of point clouds, and D represents the dimension of the point cloud, including the coordinates of the x-axis, y-axis, and z-axis as well as the reflection intensity;

[0045] Furthermore, in order to avoid the disadvantage of different sparsity of point clouds at near and far distances, a multi-scale grouping strategy is adopted to take concentric spheres of different sizes, obtain multiple local neighborhood features with the same center but different scales, and then splice them to capture spatial information from local to global. In this process, four iterations are performed, and the number of sampling points is 4096, 1024, 256, and 64 respectively; and the corresponding different radius ranges are [0.1, 0.5], [0.5, 1.0], [1.0, 2.0], [2.0, 4.0]; the number of points in the grouping sphere is [16, 32], [16, 32], [16, 32], [16, 32]; the final feature vectors are [32, 64], [128, 128], [256, 256], [512, 512] respectively.

[0046] S2, feature upsampling and fusion;

[0047] S21. Upsample the point-by-point features to give each feature point high-level geometric structure information, and perform jump connections through the concatenated features obtained through reverse interpolation and downsampling operations to enhance the spatial continuity and hierarchical fusion of the features, and obtain the fused point cloud data (N, 128), which provides sufficient and rich feature representation for subsequent target detection tasks.

[0048] S22. Subsequently, high-level geometric structure information of the point cloud in the voxels is updated in batches according to the index of the point cloud.

[0049] In addition, the original point cloud position information is spliced ​​into the corresponding point cloud high-level geometric structure information to enrich the feature information, and the point cloud high-level geometric structure information and original position information are processed using maximum pooling and average pooling respectively.

[0050] S3, voxel feature encoding;

[0051] The processed voxels are sent to the voxel-based three-dimensional feature encoding network. The fine-grained features of the point cloud and the global structural features of the voxels are combined, and a series of three-dimensional sparse convolutions are used for feature encoding. The specific operation is as follows: the size of each input voxel is set to (0.05m, 0.05m, 0.1m), and submanifold convolution and three-dimensional sparse convolution are used alternately for feature encoding to obtain preliminary features.

[0052] This method effectively solves the problem of computing resources and memory consumption of traditional 3D convolutional networks when processing large-scale point cloud data by combining the fine-grained features of point clouds and the global structure features of voxels.

[0053] S4, 2D feature encoding and fusion;

[0054] S41, first, by multiplying the height dimension in the 3D feature map with the feature channel, performing voxel operation, obtaining a 2D bird's-eye view image, inputting it into the 2D backbone network for preliminary feature extraction, and then obtaining the original feature map;

[0055] S42, adjust the input original feature map by using a multi-layer perceptron (MLP), a batch normalization layer (BN), and a ReLU function to normalize the dimension of the feature map to the required number of channels C;

[0056] S43. Next, two layers of two-dimensional convolutional layers with a convolution kernel size of 3×3 and a stride of 2×2 are used to downsample the standardized feature maps to generate two sets of preliminary feature maps of H / 2×W / 2×2C and H / 4×W / 4×4C; these feature maps provide the basis for subsequent region division and self-attention calculation;

[0057] S44, after generating the preliminary feature map, further process the preliminary feature map using a regional Transformer block to obtain an intermediate feature map. Specifically, Figure 3 As shown in the figure, the preliminary feature map is divided into several non-overlapping regions of size m×n, and self-attention calculation is performed independently in each region to perform preliminary feature extraction. The above local self-attention mechanism overcomes the inherent limitations of traditional convolution operations and can more effectively capture local information while significantly reducing computational complexity. By introducing the self-attention mechanism, the accuracy of the network for point cloud data is improved. The network can capture key features more accurately, which helps to accurately locate and identify targets, thereby improving the accuracy of target detection tasks.

[0058] In order to further enhance the expressiveness of features, a moving region strategy is introduced. After the initial region self-attention calculation is completed, more surrounding context information can be captured by moving these region Transformers and recalculating self-attention. This moving region strategy realizes the effective integration of local and global information, allowing the model to capture richer feature representations at different scales.

[0059] Although traditional convolution operations are good at extracting local features, they are insufficient in capturing global information. Through the above-mentioned region division and self-attention calculation, the model of the present invention can independently perform self-attention calculations in multiple non-overlapping regions. By moving these regions, the model can not only capture local features, but also integrate global information, thereby better understanding targets in complex scenes. This optimized self-attention mechanism makes the model more efficient and accurate when processing irregular and sparse point cloud data, improving the quality of feature extraction and the robustness of the model.

[0060] S45. After completing the preliminary feature extraction, a strategy of integrating feature pyramid network (FPN) and path enhancement network (PAN) is adopted to construct a multi-scale feature pyramid, such as Figure 4 As shown in the figure, the expression ability of multi-scale features is enhanced through top-down and bottom-up feature fusion. At each layer of the feature pyramid, a multi-layer perceptron is used to unify the number of channels to 128 dimensions, and then splicing and upsampling are performed to obtain fused multi-scale features. This multi-scale feature fusion strategy can effectively integrate information from different levels and significantly improve the detection ability of objects of various scales.

[0061] The present invention realizes the effective fusion of multi-scale features by fusing the feature pyramid network (FPN) and the path enhancement network (PAN), thereby enhancing the perception range of the model. The feature pyramid network enhances the expression ability of multi-scale features through top-down and bottom-up feature fusion, enabling the model to capture information of large-scale and small-scale targets at the same time, thereby improving the overall detection performance.

[0062] S46. Finally, the fused multi-scale features are input into the multi-head region proposal network (RPN) to perform preliminary classification and regression tasks. The design of the multi-head RPN can accurately detect objects of different scales, and through multi-head parallel processing, the detection accuracy and robustness of objects of different scales can be further improved.

[0063] S5. Data enhancement and target detection based on random point cloud elimination;

[0064] Based on the characteristics of point cloud data captured by LiDAR technology, that is, point clouds are mainly distributed on the surface of objects, a carefully designed random elimination operation is performed on the selected point cloud subset within the target ground truth bounding box. Based on the missing point cloud phenomenon on the surface of the object, that is, occlusion and truncation, the partial invisible situation of the object is simulated, thereby enhancing the adaptability and robustness of the deep learning model to such scenes. Based on this principle, the following operations are performed on the selected ground truth target bounding box part with random probability:

[0065] S51. First, divide the target's true bounding box into a blocks, randomly select and delete all the point clouds in one of the blocks (e.g. Figure 2 The dark areas shown in (a), (c), and (e) are used to simulate the situation where the object is partially blocked, where a is a natural number, a≥4, and 8 is taken in this embodiment.

[0066] S52, then, randomly divide the true value bounding box into upper and lower or left and right b blocks, and randomly delete all point clouds of one of these divided blocks (such as Figure 2 The dark areas shown in (b), (d), and (f) are used to simulate the situation where the object is truncated; wherein b is a natural number, b≥2, and is 3 in this embodiment.

[0067] Furthermore, considering the inherent sparsity of point cloud data, this operation is only applied to the first half of the objects within the detection range to ensure the effectiveness and rationality of the enhancement operation.

[0068] Since three-dimensional point cloud data is usually sparse and irregular, this poses a challenge to traditional convolutional neural networks. The present invention uses a data enhancement method based on random point cloud elimination to allow the model to be exposed to more incomplete and irregular point cloud data during training. By randomly deleting part of the point cloud data, occlusion and truncation in the real environment are simulated, and the complete form of the object is inferred from the partially missing information, thereby improving the adaptability to irregular data.

[0069] The present invention is based on the point and voxel three-dimensional feature coding layer. It combines the point-based three-dimensional feature extraction method with the voxel-based three-dimensional sparse convolutional network for the first time, deeply integrates the original features of the point cloud and the characteristics of voxel fast feature coding, summarizes multi-scale local point cloud information, makes full use of the fine-grained features and information of the original point cloud, and focuses on local point cloud features. This network structure can deeply mine the original features of point cloud data, especially for small-sized objects, and significantly improves the accuracy and reliability of detection.

[0070] The two-dimensional feature encoding network constructed by the present invention integrates the Transformer of the moving area and the FPN-PAN feature pyramid. By performing self-attention calculations in different downsampled feature map areas and motion areas, the associated communication with surrounding information is established through sliding, and rich scene context information is encoded at multiple scales; then, the FPN-PAN layer is used to splice information of different scales and send it to the multi-head RPN to better extract targets of different scales, improve the three-dimensional target detection performance, and be more suitable for downstream tasks such as target detection. Through the self-attention mechanism and multi-scale feature fusion, the encoding of local features and global context information is strengthened, and the detection performance of the model for targets of different scales is improved, especially the ability to recognize targets in complex scenes.

[0071] The present invention is based on the moving region Transformer and feature pyramid, and aims to achieve detection accuracy comparable to that of current advanced models by deeply fusing the original features of point clouds with the advantages of fast feature encoding of voxels, while making full use of bird's-eye view features. Through a series of innovative designs, while improving detection accuracy, the model's ability to adapt to complex environments is also enhanced, contributing new strength to the development of autonomous driving technology.

[0072] The data enhancement method based on random point cloud elimination of the present invention simulates incomplete point cloud data, randomly removes point clouds of different sizes and positions from some truth boxes, and trains the model to learn to infer complete object shape information from partial information, thereby improving the model's adaptability to occlusion and truncation problems in the real world, thereby significantly improving the generalization ability and robustness of the model, making it more stable in the face of complex environments and improving the three-dimensional detection performance.

[0073] The embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, and the computer program executes the steps of the above method when it is executed by a processor. The embodiment of the present invention also provides a three-dimensional object detection system, including a memory and a processor connected to each other, and the memory stores a computer program, and the computer program executes the steps of the above method when it is executed by the processor. The medium and system of the present invention correspond to the above method and also have the advantages described in the above method.

[0074] The present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned method embodiment can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable storage media include: any entity or device that can carry computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing computer programs and / or modules stored in the memory, and calling data stored in the memory. The memory may include a high-speed random access memory and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0075] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A three-dimensional target detection method, characterized in that: Includes steps: S1, obtain the original point cloud data, and downsample the original point cloud data layer by layer to extract point-by-point features of different scales; S2, upsampling the point-by-point features of different scales, giving each feature point high-level geometric structure information, and obtaining updated point cloud data; S3, feature encoding the updated point cloud data to obtain preliminary features; S4, generate a bird's-eye view of the preliminary features, and generate a preliminary feature map based on the bird's-eye view, then use the regional Transformer block and the moving region strategy to extract features on the preliminary feature map, build a multi-scale feature pyramid, and fuse information from different scales through top-down and bottom-up feature fusion methods to obtain the final multi-scale feature map; S5, target detection based on multi-scale feature maps; In step S5, a random point cloud elimination operation is performed on the selected point cloud subset within the target true value bounding box; The specific process of random point cloud elimination operation is: First, the true bounding box of the target is divided into a blocks, and all point clouds of one of the blocks are randomly selected and deleted to simulate the situation where the object is partially occluded; where a is a natural number, a≥4; Secondly, the true bounding box is randomly divided into b blocks above and below or left and right, and all point clouds of a block in the divided area are randomly deleted to simulate the situation where the object is truncate; where b is a natural number, b≥2.

2. The three-dimensional target detection method according to claim 1, characterized in that: The specific process of step S4 is: S41, generating a bird's-eye view of the preliminary features, and then generating an original feature map; S42, normalizing the dimension of the original feature map to the number of channels C; S43, down-sampling the standardized feature map to generate a preliminary feature map; S44, after generating the preliminary feature map, the regional Transformer block is used to perform self-attention calculation on the feature map to extract features; S45. After the feature extraction is completed, a strategy of fusing feature pyramid network and path enhancement network is adopted to construct a multi-scale feature pyramid; through top-down and bottom-up feature fusion methods, at each layer of the feature pyramid, a multi-layer perceptron is used to unify the number of channels C to a preset dimension, and then splicing and upsampling are performed to obtain the fused multi-scale features; S46. Input the fused multi-scale features into the multi-head region proposal network to perform preliminary classification and regression tasks.

3. The three-dimensional target detection method according to claim 2, characterized in that: In step S44, a moving region strategy is introduced. After the initial region self-attention calculation is completed, these regions are moved and self-attention is recalculated to capture more surrounding context information.

4. The three-dimensional target detection method according to claim 1, 2 or 3, characterized in that: The specific process of step S1 is: the original point cloud data is sent to the point-based three-dimensional feature encoding network to extract the features of the original point cloud information, and the input original point cloud (N, D) is sampled and combined at the farthest point, and downsampled layer by layer using the PointNet network to generate point-by-point features of different scales; Where N represents the number of point clouds, and D represents the dimension of the point cloud, including the coordinates of the x-axis, y-axis, and z-axis, as well as the reflection intensity.

5. The three-dimensional target detection method according to claim 4, characterized in that: In step S1, a multi-scale grouping strategy is adopted to take concentric spheres of different sizes, obtain multiple local neighborhood features with the same center but different scales, and then splice them to capture spatial information from local to global.

6. The three-dimensional target detection method according to claim 1, 2 or 3, characterized in that: The specific process of step S3 is: sending the updated point cloud data into the voxel-based three-dimensional feature encoding network, and using three-dimensional sparse convolution for feature encoding. The specific operation is: setting the size of each voxel of the input, and alternately using submanifold convolution and three-dimensional sparse convolution for feature encoding.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 6.

8. A three-dimensional object detection system, comprising a memory and a processor connected to each other, wherein a computer program is stored in the memory, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 6.

Images