Large indoor space moving object positioning method fusing depth information

Abstract

The invention relates to the technical field of target detection, and discloses a depth information fused large indoor space moving object positioning method, which comprises the following steps of performing dead-corner-free deployment of a monitoring camera according to characteristics of a large indoor space; carrying out joint calibration on the plurality of monitoring cameras; performing initial extraction of video foreground information by adopting a Vibe algorithm; the detection precision is improved by adopting a method of fusing image pixel features and image depth information; and carrying out joint detection positioning on the moving target by fusing multivariate information, namely foreground information, depth information and pixel features. According to the method, the adaptability to illumination change is improved through local adaptive pixel feature adjustment, a Vibe algorithm is adopted to fuse regional image depth-pixel value correlation analysis, a regional image depth-pixel value correlation equation is established, a space attention mechanism is introduced, interference of background noise is reduced through dynamic adjustment of regional attention weight, and the accuracy of regional image depth-pixel value correlation analysis is improved. And high-precision positioning of a large indoor space moving object is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a method for positioning moving objects in a large indoor space by integrating depth information. Background Art

[0002] Video-based moving target extraction, that is, extracting moving targets through image segmentation methods for further processing. The existing technologies in the field of moving target detection have made significant progress, but there are still many problems. At present, the video-based moving target detection methods mainly include background subtraction, frame difference and optical flow methods. These methods perform well in fixed camera scenes, but in complex and changing dynamic environments, such as lighting changes and dynamic backgrounds, the detection effect is often unsatisfactory. Although the background subtraction method is classic and widely used, it is easily disturbed by external factors such as lighting changes and dynamic backgrounds, and requires frequent background updates. The frame difference method is simple to calculate and can quickly respond to changes in the scene, but this feature also makes it susceptible to noise, and the detection effect is not good for targets with slow movement or subtle changes. The optical flow method does not rely on a fixed background model and can more flexibly process motion information in dynamic scenes, but the amount of calculation is large and the real-time performance is limited. The above technologies have their own advantages and disadvantages, but it is difficult to meet the requirements of accuracy, real-time and ease of use at the same time. Summary of the invention

[0003] In view of this, an object of the present invention is to provide a method for positioning moving objects in a large indoor space by integrating depth information, which can realize stereoscopic detection of moving targets in a large indoor space.

[0004] In order to achieve the above-mentioned invention object, the technical solution adopted is as follows:

[0005] The method for positioning a large indoor moving object by integrating depth information comprises the following steps:

[0006] S1. According to the structural characteristics of large indoor spaces, determine the number, specifications, deployment locations and angles of cameras to achieve blind spot monitoring of large indoor spaces and establish an initial background library;

[0007] S2. Jointly calibrate multiple surveillance cameras according to the camera deployment position and angle, and realize the time synchronization of each camera, determine the time synchronization of the video at each angle, and provide a basis for the subsequent surveillance image processing process;

[0008] S3, preprocessing the surveillance video image, including: image enhancement, image mean filtering, and local adaptive adjustment of image brightness and contrast, and at the same time, judging the sudden change of illumination. If a sudden change of illumination occurs, reinitialize the background;

[0009] S4, using the Vibe algorithm to establish, match and update the background library model of the video, thereby realizing the extraction of the initial foreground information of the video;

[0010] S5. Feature extraction and feature matching of the same-frame video image to obtain pixel depth information, perform depth-pixel value correlation analysis, establish depth-pixel value correlation equation according to the depth information, set the threshold range, and introduce a dynamic spatial attention mechanism to detect depth change areas. Perform a logical OR operation on the depth change area and the background information to obtain the final moving target detection result.

[0011] As a further improvement of the present invention, in step S3, image enhancement, image mean filtering, and local adaptive adjustment of image brightness and contrast are performed, and the specific steps are as follows:

[0012] S3-1, image enhancement, improving the resolution of the original video image;

[0013] S3-2, image filtering: using a mean filtering method to denoise the video image and filter out the noise in the video image;

[0014] S3-3, local adaptive contrast adjustment, divide the image into multiple areas according to its characteristics, use logarithmic compression method, adjust the base of logarithmic compression method based on the characteristics of local area pixels, and reduce the impact of light changes on detection effect.

[0015] As a further improvement of the present invention, in step S4, the Vibe algorithm is used to establish, match and update the background library model of the video, and the specific steps are as follows:

[0016] S4-1. The first frame image is judged for sudden illumination changes. If there is no sudden illumination change, for the current pixel, a suitable neighborhood size range is selected in combination with the similar spatiotemporal characteristics of the neighborhood to ensure that the background model includes a sufficient number of samples. Assuming that a certain pixel in the first frame is x and its pixel value is v(x), then its model sample set M(x) is:

[0017] M(x)={V 1 ,V 2 ,...V n}

[0018] Where N is the number of samples;

[0019] S4-2, perform background detection: calculate the Euclidean distance d between the pixel value of the current frame image and the background model, set a distance threshold R, compare, and then obtain pixel similarity;

[0020] S4-3. A time sampling update strategy is adopted, that is, when a pixel is determined to be background, it is updated with a probability of 1 / ω.

[0021] As a further improvement of the present invention, in step S5, feature extraction and feature matching of the same frame video image are performed to obtain pixel depth information, perform depth-pixel value correlation analysis, establish a depth-pixel value correlation equation according to the depth information, set a threshold range, and introduce a dynamic spatial attention mechanism to detect depth change areas, perform a logical OR operation on the depth change area and the background information, and obtain the final moving target detection result, which specifically includes the following steps:

[0022] S5-1, initialize the static background library, extract and match the features of the static background images of adjacent surveillance cameras, based on the parallax principle:

[0023]

[0024] Then, the depth information of each pixel value of the static background image is obtained;

[0025] D (x，y) ={D (x1,y1) , D (x2,y2) , D (x3，y3) ......D (xn，yn)}

[0026] Where f is the focal length of the surveillance camera, a is the distance between two adjacent cameras, and d is the parallax;

[0027] S5-2, perform depth-pixel value correlation analysis on the static background library, calculate the gradient change of image depth information, determine the region division boundary according to the characteristics of the static background image and the gradient change value of the depth information, and initially divide the image into i regions; based on the characteristics of the pixel values, perform secondary division on the i regions after division, divide the entire image into i×j regions, and use polynomial regression to establish the static background library depth-pixel value correlation equation by region;

[0028] I n (x, y) = f ij (D n (x, y)) = β 0 +β 1 D n (x,y)+β 2 D n (x, y) 2 +…+β n D n (x, y) n

[0029] Among them, I ij (x, y) is a set of static background pixel values; D n (x, y) is the static background depth information set; β 0 , β1 , ...β n is the regression coefficient;

[0030] S5-3. Integrate the spatial attention mechanism and assign an attention weight to each i×j region according to the depth change value and pixel change characteristics. The pixel depth calculation point sampling for the next frame, the initial value of the attention weight is 0.5, and the dynamic weight update rule is used to update the weights of each area of ​​the image. The specific rules are as follows:

[0031]

[0032] In the formula is the attention weight of a region at time t+1 and α is the step length of weight increase, β is the step length of weight decrease; ΔD i is the depth change value of the area, specifically ΔD i =|D i (t)-D i (t-1)|;ΔD i The pixel change value of this area is ΔP i =|P i (t)-P i (t-1)|;Δ D , Δ P They are the depth threshold and pixel threshold for judging whether the area has significant changes;

[0033] S5-4. After the depth-pixel value correlation equation of the static background library is established, the depth region change judgment is first performed on each subsequent frame. The depth judgment of the next frame image is performed by regional sampling. The number of sampled pixels is the product of the attention weight and the number of pixels in the region. If the number of sampling points where both the pixel value and the depth have changed in a certain region reaches 60% of the total number of sampling points in the region, it is determined to be a candidate depth change region. At this time, the region will be marked, and the lower limit of the attention weight of the marked region will be updated to 0.2;

[0034] Traverse the candidate depth change area pixels, pre-judge their pixel values, set the pixel number threshold N1, if there are more than N pixels in the neighborhood of a certain pixel, the pixel change value of N1 pixels is greater than the threshold K;

[0035]

[0036] Among them, ΔP i′,j′ =P i′,j′ (t 1 )-P i′,j′ (t 0 ) represents the pixel (i′, j′) at time t 1 and t0 The pixel value changes;

[0037] If the above conditions are met, the pixel is judged to be a change boundary, and its boundary is expanded outward to 1.2 times the number of pixels, that is, 1.2N 1 pixels, that is

[0038] Expanded Boundary = Neighborhood (Changed Boundary, 1.2N 1 )

[0039] Determine the pixel change boundary and expand it evenly to 1.2N 1 pixels, continue to judge the depth change value of the area, then set the depth change threshold T, and calculate the depth change information according to the regional depth-pixel value correlation equation:

[0040] ΔD=f(D(t 1 ),P(t 1 ))-f(D(t 0 ),P(t 0 ))

[0041] If ΔD>T, the area is judged as a depth change area;

[0042] S5-5, performing a logical OR operation on the obtained depth change area and the foreground information to obtain a final moving target detection result.

[0043] The beneficial effects of the present invention are as follows: the present invention can improve the adaptability of the traditional Vibe algorithm to light changes in dynamic target extraction, integrate depth information for judgment and extraction, effectively improve the accuracy of dynamic target extraction, and add a dynamic spatial attention mechanism to reduce the calculation and judgment of unnecessary areas, thereby reducing the calculation overhead and improving the recognition ability and efficiency of moving targets in large indoor spaces. The stereoscopic detection of moving targets in the monitoring video by multiple cameras in conjunction with an efficient moving target detection algorithm reduces the workload of relevant management personnel and effectively improves the efficiency and accuracy of moving target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0045] Figure 1 It is the overall flow chart of the present invention;

[0046] Figure 2 This is a schematic diagram of initializing the background sample library;

[0047] Figure 3It is a flow chart of regional depth value-pixel correlation analysis;

[0048] Figure 4 It is a schematic diagram of image area division;

[0049] Figure 5 It is a flowchart of dynamic update of regional attention weights. DETAILED DESCRIPTION

[0050] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0051] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0052] like Figure 1-5 As shown, the method for positioning a large indoor space moving object by integrating depth information includes the following steps:

[0053] S1. According to the structural characteristics of large indoor spaces, determine the number, specifications, deployment locations and angles of cameras to achieve blind spot monitoring of large indoor spaces and establish an initial background library;

[0054] S2. Jointly calibrate multiple surveillance cameras according to the camera deployment position and angle, and realize the time synchronization of each camera, determine the time synchronization of the video at each angle, and provide a basis for the subsequent surveillance image processing process;

[0055] S3, preprocessing the surveillance video image, including: image enhancement, image mean filtering, and local adaptive adjustment of image brightness and contrast, and at the same time, judging the sudden change of illumination. If a sudden change of illumination occurs, reinitialize the background;

[0056] S4, using the Vibe algorithm to establish, match and update the background library model of the video, thereby realizing the extraction of the initial foreground information of the video;

[0057] S5. Feature extraction and feature matching of the same-frame video image to obtain pixel depth information, perform depth-pixel value correlation analysis, establish depth-pixel value correlation equation according to the depth information, set the threshold range, and introduce a dynamic spatial attention mechanism to detect depth change areas. Perform a logical OR operation on the depth change area and the background information to obtain the final moving target detection result.

[0058] In step S3, image enhancement, image mean filtering, and local adaptive adjustment of image brightness and contrast are performed. The specific steps are as follows:

[0059] S3-1, image enhancement, improving the resolution of the original video image;

[0060] S3-2, image filtering: using a mean filtering method to denoise the video image and filter out the noise in the video image;

[0061] S3-3, local adaptive contrast adjustment, divide the image into multiple areas according to its characteristics, use logarithmic compression method, adjust the base of logarithmic compression method based on the characteristics of local area pixels, and reduce the impact of light changes on detection effect.

[0062] In step S4, the Vibe algorithm is used to establish, match and update the background library model of the video. The specific steps are as follows:

[0063] S4-1. The first frame image is judged for sudden changes in illumination. If there is no sudden change in illumination, for the current pixel, a suitable neighborhood size range is selected in combination with similar spatiotemporal characteristics of the neighborhood to ensure that the background model includes a sufficient number of samples, such as Figure 2 As shown in the figure, assuming that a pixel point in the first frame is x and its pixel value is v(x), then its model sample set M(x) is:

[0064] M(x)={V 1 , V 2 , ..., V n}

[0065] Where N is the number of samples;

[0066] S4-2, perform background detection: calculate the Euclidean distance d between the pixel value of the current frame image and the background model, set a distance threshold R, compare, and then obtain pixel similarity;

[0067] S4-3. A time sampling update strategy is adopted, that is, when a pixel is determined to be background, it is updated with a probability of 1 / ω.

[0068] In step S5, feature extraction and feature matching of the same frame video image are performed to obtain pixel depth information, and depth-pixel value correlation analysis is performed. A depth-pixel value correlation equation is established according to the depth information, a threshold range is set, and a dynamic spatial attention mechanism is introduced to detect the depth change area. The depth change area is logically or operated with the background information to obtain the final moving target detection result, which specifically includes the following steps:

[0069] S5-1, initialize the static background library, extract and match the features of the static background images of adjacent surveillance cameras, based on the parallax principle:

[0070]

[0071] Then, the depth information of each pixel value of the static background image is obtained;

[0072] D (x，y) ={D (x1,y1) , D (x2,y2) , D (x3，y3) ......D (xn，yn)}

[0073] Where f is the focal length of the surveillance camera, a is the distance between two adjacent cameras, and d is the parallax;

[0074] S5-2. Perform depth-pixel value correlation analysis on the static background library. The analysis process is as follows: Figure 3 As shown, the gradient change of image depth information is calculated, the region division boundary is determined according to the characteristics of the static background image and the gradient change value of the depth information, and the image is initially divided into i regions; based on the characteristics of the pixel value, the i regions after division are divided twice, and the entire image is divided into i×j regions. The static background library depth-pixel value correlation equation is established by region using polynomial regression;

[0075] I n (x,y)=f ij (D n (x, y)) = β 0 +β 1 D n (x,y)+β 2 D n (x,y) 2 +…+β n D n (x,y) n

[0076] Among them, I ij (x, y) is a set of static background pixel values; D n (x, y) is the static background depth information set; β 0 , β 1 , ...β n is the regression coefficient;

[0077] S5-3. Integrate the spatial attention mechanism and assign an attention weight to each i×j region according to the depth change value and pixel change characteristics. The pixel depth calculation point sampling for the next frame, the initial value of the attention weight is 0.5, and the dynamic weight update rule is used to update the weights of each area of ​​the image. The specific rules are as follows:

[0078]

[0079] In the formula is the attention weight of a region at time t+1 and α is the step length of weight increase, β is the step length of weight decrease; ΔD i is the depth change value of the area, specifically Δ D =|D i (t)-D i (t-1)|;ΔD i The pixel change value of this area is ΔP i =|P i (t)-P i (t-1)|;Δ D , Δ P They are the depth threshold and pixel threshold for judging whether the area has significant changes;

[0080] S5-4. After the depth-pixel value correlation equation of the static background library is established, the depth area change judgment is first performed on each subsequent frame. The depth judgment of the next frame image is performed by regional sampling. The number of sampled pixels is the product of the attention weight and the number of pixels in the area. If the number of sampling points where both the pixel value and the depth have changed in a certain area reaches 60% of the total number of sampling points in the area, it is determined to be a candidate depth change area. At this time, the area will be marked, and the lower limit of the attention weight of the marked area is updated to 0.2;

[0081] Traverse the candidate depth change area pixels, pre-judge their pixel values, set the pixel number threshold N1, if there are more than N pixels in the neighborhood of a certain pixel, the pixel change value of N1 pixels is greater than the threshold K;

[0082]

[0083] Where ΔP i′,j′ =P i′,j′ (t 1 )-P i′,j′ (t 0 ) represents the pixel (i′, j′) at time t 1 and t 0 The pixel value changes;

[0084] If the above conditions are met, the pixel is judged to be a change boundary, and its boundary is expanded outward to 1.2 times the number of pixels, that is, 1.2N 1 pixels, that is

[0085] Expanded Boundary = Neighborhood (Changed Boundary, 1.2N 1 )

[0086] Determine the pixel change boundary and expand it evenly to 1.2N 1 pixels, continue to judge the depth change value of the area, then set the depth change threshold T, and calculate the depth change information according to the regional depth-pixel value correlation equation:

[0087] ΔD=f(D(t 1 ),P(t 1 ))-f(D(t 0 ),P(t 0 ))

[0088] If ΔD>T, the area is judged as a depth change area;

[0089] S5-5, performing a logical OR operation on the obtained depth change area and the foreground information to obtain a final moving target detection result.

[0090] The present invention provides a method for positioning moving objects in a large indoor space by integrating depth information. The method has high detection performance. The method improves adaptability to illumination changes by adaptively adjusting local pixel features of an input image. The static background is subjected to depth-pixel value correlation analysis, and the region division boundary is determined according to the characteristics of the static background image and the gradient change value of the depth information. A spatial attention mechanism is introduced, and a regional weight dynamic adjustment rule is adopted to flexibly adjust the weight of each region, thereby increasing attention to high-weight regions and reducing unnecessary calculation overhead. A regional depth-pixel value correlation equation is established, and the depth information of a video image is integrated to perform moving target judgment, thereby improving the moving target detection accuracy and detection efficiency, and having certain robustness.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, component splitting or combination, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for positioning moving objects in a large indoor space by integrating depth information, characterized in that: The following steps are involved: S1. According to the structural characteristics of large indoor spaces, determine the number, specifications, deployment locations and angles of cameras to achieve blind spot monitoring of large indoor spaces and establish an initial background library; S2. Jointly calibrate multiple surveillance cameras according to the camera deployment position and angle, and realize the time synchronization of each camera, determine the time synchronization of the video at each angle, and provide a basis for the subsequent surveillance image processing process; S3, preprocessing the surveillance video image, including: image enhancement, image mean filtering, and local adaptive adjustment of image brightness and contrast, and at the same time, judging the sudden change of illumination. If a sudden change of illumination occurs, reinitialize the background; S4, using the Vibe algorithm to establish, match and update the background library model of the video, thereby realizing the extraction of the initial foreground information of the video; S5. Feature extraction and feature matching of the same-frame video image to obtain pixel depth information, perform depth-pixel value correlation analysis, establish depth-pixel value correlation equation according to the depth information, set the threshold range, and introduce a dynamic spatial attention mechanism to detect depth change areas. Perform a logical OR operation on the depth change area and the background information to obtain the final moving target detection result.

2. The method for positioning moving objects in large indoor spaces by integrating depth information according to claim 1, characterized in that: In step S3, image enhancement, image mean filtering, and local adaptive adjustment of image brightness and contrast are performed. The specific steps are as follows: S3-1, image enhancement, improving the resolution of the original video image; S3-2, image filtering: using a mean filtering method to denoise the video image and filter out the noise in the video image; S3-3, local adaptive contrast adjustment, divide the image into multiple areas according to its characteristics, use logarithmic compression method, adjust the base of logarithmic compression method based on the characteristics of local area pixels, and reduce the impact of light changes on detection effect.

3. The method for positioning moving objects in large indoor spaces by integrating depth information according to claim 1, characterized in that: In step S4, the Vibe algorithm is used to establish, match and update the background library model of the video. The specific steps are as follows: S4-1. The first frame image is judged for sudden illumination changes. If there is no sudden illumination change, for the current pixel, a suitable neighborhood size range is selected in combination with the similar spatiotemporal characteristics of the neighborhood to ensure that the background model includes a sufficient number of samples. Assuming that a certain pixel in the first frame is x and its pixel value is v(x), then its model sample set M(x) is: M(x)={V1,V2,...V n } Where N is the number of samples; S4-2, perform background detection: calculate the Euclidean distance d between the pixel value of the current frame image and the background model, set a distance threshold R, compare, and then obtain pixel similarity; S4-3. A time sampling update strategy is adopted, that is, when a pixel is determined to be background, it is updated with a probability of 1 / ω.

4. The method for positioning moving objects in large indoor spaces by integrating depth information according to claim 1, characterized in that: In step S5, feature extraction and feature matching of the same frame video image are performed to obtain pixel depth information, and depth-pixel value correlation analysis is performed. A depth-pixel value correlation equation is established according to the depth information, a threshold range is set, and a dynamic spatial attention mechanism is introduced to detect the depth change area. The depth change area is logically or operated with the background information to obtain the final moving target detection result, which specifically includes the following steps: S5-1, initialize the static background library, extract and match the features of the static background images of adjacent surveillance cameras, based on the parallax principle: Then, the depth information of each pixel value of the static background image is obtained; D (x,y) ={D (x1,y1) ,D (x2,y2) ,D (x3,y3 )......D (xn,yn) } Where f is the focal length of the surveillance camera, a is the distance between two adjacent cameras, and d is the parallax; S5-2, perform depth-pixel value correlation analysis on the static background library, calculate the gradient change of image depth information, determine the region division boundary according to the characteristics of the static background image and the gradient change value of the depth information, and initially divide the image into i regions; based on the characteristics of the pixel values, perform secondary division on the i regions after division, divide the entire image into i×j regions, and use polynomial regression to establish the static background library depth-pixel value correlation equation by region; I n (x,y)=f ij (D n (x,y))=β0+β1D n (x,y)+β2D n (x,y) 2 +…+β n D n (x,y) n Among them, I ij (x, y) is a set of static background pixel values; D n (x, y) is the static background depth information set; β0, β1, ...β n is the regression coefficient; S5-3. Integrate the spatial attention mechanism and assign an attention weight to each i×j region according to the depth change value and pixel change characteristics. The pixel depth calculation point sampling for the next frame, the initial value of the attention weight is 0.5, and the dynamic weight update rule is used to update the weights of each area of ​​the image. The specific rules are as follows: In the formula is the attention weight of a region at time t+1 and α is the step length of weight increase, β is the step length of weight decrease; ΔD i is the depth change value of the area, specifically ΔD i =|D i (t)-D i (t-1)|;ΔD i The pixel change value of this area is ΔP i =|P i (t)-P i (t-1)|;Δ D , Δ P They are the depth threshold and pixel threshold for judging whether the area has significant changes; S5-4. After the depth-pixel value correlation equation of the static background library is established, the depth area change judgment is first performed on each subsequent frame. The depth judgment of the next frame image is performed by regional sampling. The number of sampled pixels is the product of the attention weight and the number of pixels in the area. If the number of sampling points where both the pixel value and the depth have changed in a certain area reaches 60% of the total number of sampling points in the area, it is determined to be a candidate depth change area. At this time, the area will be marked, and the lower limit of the attention weight of the marked area is updated to 0.2; Traverse the candidate depth change area pixels, pre-judge their pixel values, set the pixel number threshold N1, if there are more than N pixels in the neighborhood of a certain pixel, the pixel change value of N1 pixels is greater than the threshold K; Among them, ΔP i′,j′ =P i′,j′ (t1)-P i′,j′ (t0) represents the change in pixel value of pixel point (i′, j′) at time t1 and t0; If the above conditions are met, the pixel is judged to be a change boundary, and its boundary is expanded outward to 1.2 times the number of pixels, that is, 1.2N1 pixels, that is ExpandedBoundary = Neighborhood(ChangeBoundary,1.2N1) Determine the pixel change boundary and expand it evenly to 1.2N1 pixels. Continue to determine the depth change value of the area, set the depth change threshold T, and calculate the depth change information according to the regional depth-pixel value correlation equation: ΔD=f(D(t1),P(t1))-f(D(T0),P(t0)) If ΔD>T, the area is judged as a depth change area; S5-5, performing a logical OR operation on the obtained depth change area and the foreground information to obtain a final moving target detection result.

Images