System and Method for Generating High-Resolution Stereo Image and Depth Map

Abstract

A system and method for generating high-resolution stereo images and depth map in multi-camera systems having multiple cameras with different resolutions and view angles. One method is to improve the lower resolution image and combining it with the higher resolution image, then the resulting image is processed by extensive algorithms to ensure utmost high quality. The system can also handle non-planar image contents. The process is to generate a crude depth map first and then divide the map into multiple layers. Each layer will be separately registered. The results from the registered layers will be merged to improve the depth map generation. The improved depth map could be repeatedly fed back to the beginning of the process to further improve the registration performance. The system and method can generate stereo images using uncalibrated cameras with different views and resolutions.

Claims

1. A system for generating high-resolution stereo image and depth map comprising: a first camera having a first image; a second camera having a second image; an up-sampler connected to the first image; a two-step image alignment module is connected to a first output of the up-sampler and the second image; a pan-sharpening module is connected to a second output of the up-sampler; a pan band creator is connected to an output of the two-step image alignment module to provide an input to the pan-sharpening module for producing a first high resolution image; a stereo image generator for combining the first high resolution image and the second image to generate a stereo image; a disparity map generator receives the stereo image to generate a disparity map; and the generated disparity map is connected to a first input of a depth map generator.

2. A system for generating high-resolution stereo image and depth map as claim in claim 1, wherein, the first image is a low resolution image, and the second image is a high resolution image.

3. A system for generating high-resolution stereo image and depth map as claim in claim 2, further comprising: a divider connected between the depth map generator and the two-step image alignment module; and the depth map generator receives a third output from the up-sampler.

4. A system for generating high-resolution stereo image and depth map as claim in claim 3, wherein, the divider divides an output of the depth map generator into multiple levels to manage any non-coplanar image contents.

5. A system for generating high-resolution stereo image and depth map as claim in claim 1, wherein, the first and second cameras are uncalibrated with different views and resolutions.

6. A method for generating high-resolution stereo image and depth map comprising the steps of: upsampling a first image; aligning the upsampled first image with a second image using a two-step image alignment process; creating a Panchromatic (pan) band using multispectral bands of the second image; pansharpening the upsampled first image with the pan band created by the second image to generate a high resolution left image; forming a stereo image using the high resolution left image and the second image; generating a disparity map using the stereo image; and generating a first depth map using the disparity map.

7. A method for generating high-resolution stereo image and depth map as claim in claim 6, wherein the two-step image alignment step further comprising the steps of: combining the upsampled first image and the second image to generate a second depth map; dividing the second generated depth map into multiple levels; registering the multiple levels individually; merging all the registered levels to improve the second depth map; aligning the divided depth map with the upsampled left image.

8. A method for generating high-resolution stereo image and depth map as claim in claim 7, wherein, the second image is of high resolution; and the improved depth map is repeatedly fed back to the first depth map generating step to further improve the registering performance.

9. A method for generating high-resolution stereo image and depth map as claim in claim 7, wherein, the registering step is performed by Scale Invariant Features Transform (SIFT) or Speeded Up Robust Features (SURF) with Random Sample Consensus (RANSAC).

10. A method for generating high-resolution stereo image and depth map as claim in claim 6, wherein, the pansharpening step is performed by Gram-Schmidt Adaptive (GSA) algorithm.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0017] FIG. 1: Signal flow of a new stereo image formation and depth map generation system.

[0018] FIG. 2: Alignment scheme to deal with non-coplanar image contents.

[0019] FIG. 3. Block diagram of the two-step image alignment approach.

[0020] FIG. 4: Procedures in stereo image formation.

[0021] FIG. 5: Image rectification.

[0022] FIG. 6: Relationship between disparity and depth.

[0023] FIG. 7: Original images. Left: left Mastcam; right: right Mastcam.

[0024] FIG. 8: Cropping left image then upsampling.

[0025] FIG. 9: Matched feature points for aligning right image to left (smaller image is left image).

[0026] FIG. 10: Right image on top of left image after alignment.

[0027] FIG. 11: Left: image before pansharpening; right: image after pansharpening.

[0028] FIG. 12: Rectified stereo images.

[0029] FIG. 13: Depth maps without pansharpening (left) and with pansharpening (right).

[0030] FIG. 14: Left original low resolution (LR) and Right high resolution (HR) images.

[0031] FIG. 15: Depth map generated using LR left and HR right images.

[0032] FIG. 16: Comparison of original and pansharpened left images.

[0033] FIG. 17: Zoomed version of the previous image in FIG. 16.

[0034] FIG. 18: Stereo image formed by using pan-sharpened left and original right images.

[0035] FIG. 19: Depth map estimated using pan-sharpened left image and original right image.

DETAILED DESCRIPTION OF THE INVENTION

[0036] The present invention utilizes a new approach to high resolution stereo image formation and high resolution depth map generation. As shown in FIG. 1, the approach consists of the following steps: [0037] 1. First, an accurate two-step image registration approach is used to align the left and right images. The left image is used as the reference for registration. The coarse step is to use Scale Invariant Features Transform (SIFT) or Speeded Up Robust Features (SURF) features with Random Sample Consensus (RANSAC). The fine step is to apply a diffeomorphic algorithm to achieve subpixel accuracy. The common area between the left and right images is then extracted. The alignment can achieve sub-pixel accuracy. In the event that the image contents are not coplanar, we propose a scheme shown in FIG. 2 to use the upsampled left image and the right image to get a depth map. We then divide the depth map into many layers and perform registration for each layer separately. The alignment results from different layers are then merged. [0038] 2. Second, a Panchromatic (pan) band is created using the multispectral bands in the right image. One simple way to create the pan band is to take the average of all the available bands in the right camera. Any pansharpening algorithms, References [5]-[14], can be used for pansharpening. In the examples of the present invention, a pansharpening algorithm known as Gram-Schmidt Adaptive (GSA) algorithm was used to pansharpen the left image. [0039] 3. Third, the pansharpened left image and the original high resolution right image are then used to create a stereo image. In the stereo image creation process, some standard procedures, including feature points extraction, feature points matching between left and right images, estimation of fundamental matrix, outliers removal based on epipolar constraint, and image rectification, can be used. [0040] 4. Fourth, based on the stereo image, a disparity map and a depth map can then be generated.

[0041] FIG. 1 and FIG. 2 show the signal flow. In the subsequent sections, the key algorithms could be utilized in the present invention are summarized.

Two-Step Image Registration

[0042] The block diagram of the two-step image alignment approach is shown in Fig. The first step of the two-step image alignment approach is using RANSAC (Random Sample Consensus) technique, Reference [1], for an initial image alignment. In this first step, we use the two RGB images from the left and right imagers. The left image is first upsampled using bicubic interpolation to the same resolution of the right image. First, SURF features, Reference [2] and SIFT features, References [3][4], are extracted from the two images. These features are then matched within the image pair. This is followed by applying RANSAC to estimate the geometric transformation. Assuming the left camera image is the reference image, the right camera image content is then projected into a new image that is aligned with the reference image using the geometric transformation.

[0043] The second step of the two-step alignment approach uses this aligned right image with RANSAC and the left camera image as inputs and applies the Diffeomorphic Registration, Reference [3], technique. Diffeomorphic Registration is formulated as a constrained optimization problem, which is solved with a step-then-correct strategy, Reference [3]. This second step reduces the registration errors to subpixel levels and makes it possible to conduct accurate pansharpening.

Pansharpening Algorithm

[0044] The goal of pansharpening, References [5]-[14], is to fuse a low-spatial resolution left image with a high-spatial resolution panchromatic image (pan) from the right camera. In the present invention, after the two-step registration, the image from the left camera can be considered as a blurred version of the right one. The images from the left camera are sharpened by pansharpening using high spatial resolution images from the right camera as the panchromatic reference image.

[0045] Pansharpening techniques can be classified into two main categories:

[0046] (1) the Component Substitution (CS) approach; and

[0047] (2) the MultiResolution Analysis (MRA) approach.

[0048] The CS approach is based on the substitution of a component with the pan image and the MRA approach relies on the injection of spatial details that are obtained through a multiresolution decomposition of the pan image into the resampled MS bands. In the present invention, we focus on the CS-based approach. Under the assumption that the components containing the spatial structure of multispectral images at all spectral bands are highly correlated, the transformed low-resolution MS images can be enhanced by substituting the components containing the spatial structure of the pan image through a histogram matching. The output pansharpened data are finally achieved by applying the inverse transformation to project the data back to the original space.

[0049] The Gram-Schmidt Adaptive (GSA), Reference [5], algorithm is applied for its simplicity and performance in our experiments.

Stereo Image Formation

[0050] FIG. 4 shows the key steps in stereo image formation. Given a pansharpened left image and the original high resolution right image, we first perform a feature points extraction. This step is similar to the two-step image registration described earlier. Both SIFT and SURF features can be used. In the past, we found that SIFT features are more robust than SURF features. Second, the feature correspondence can be achieved with RANSAC, which matches feature points that belong to the same physical locations. Third, the fundamental matrix is estimated based on the corresponded feature points. Fourth, outliers are removed using the epipolar constraint. Finally, an image rectification step is performed.

[0051] FIG. 5 illustrates the stereo rectification process. Camera 1 and Camera 2 are arranged in different image planes as illustrated in FIG. 5(a); the purpose is to get a true stereo pair from these two images through image rectification. The rectification process can be broken down into two steps. First, by finding and applying a homography to each image, these two images are transformed into new ones, which are identical to the ones captured by two parallel cameras, as illustrated in FIG. 5(b). Second, the wide or narrow baseline of the two parallel cameras is adjusted to a proper value (e.g. standard base line) by translating the new images with a proper value. Thus, the desired stereo pair is constructed in FIG. 5(c).

Disparity Estimation

[0052] Disparity is the difference between two pixels that correspond to the same physical point in the stereo image pair. Once the stereo images are created, a feature correspondence process is needed to determine the pixels that belong to the same physical point. Based the feature correspondence results, the disparity map is computed for every pixel in the image.

Depth Estimation

[0053] FIG. 6 illustrates the relationship between depth and disparity. Once disparity map is found, we can use the following formula to compute the depth, L, for each pixel.

L=Bf/X

where B is the baseline between the two cameras, f is the focal length, and X is the disparity at a particular pixel.

Experimental Results

[0054] In the following figures, FIG. 7 to FIG. 13, we will illustrate how to apply the proposed stereo image formation and depth map generation to a real Mastcam image pair. Referring to FIG. 7, it shows the original left and right images. Left image has low resolution and normally has a wide field of view. In our 2-step image registration algorithm, the left image is required to be smaller than the right image. The cropped and upsampled left image is shown in FIG. 8. The SURF features are extracted by running the RANSAC algorithm, and then apply the fine registration using the diffeomorphic algorithm. The aligned feature points are shown in FIG. 9. FIG. 10 shows the aligned image with left image overlaid on top of the right. A pan band can be created by using the aligned right image. Pansharpening is performed to improve the resolution of the left image with help from the pan band of the right. The resulting pansharpened image using the GSA algorithm is shown in the right-hand side of FIG. 11.

[0055] The stereo images could then be created by using two methods:

[0056] 1. to use the low resolution left image and the downsampled right image; and

[0057] 2. to use the high resolution left and right images.

[0058] Both stereo images are shown in FIG. 12. Finally, the disparity maps from the two stereo images are shown in FIG. 13. It can be seen the disparity map from the high resolution stereo image pair looks smoother than that without pansharpening.

[0059] FIG. 14 to FIG. 19 show the application of the proposed system in the present invention to another pair of images that have non-planar image contents. Referring to FIG. 15, Depth map is generated using low resolution left image and high resolution right image. This map is divided into multiple layers and each layer will be separately registered. The registration results from different layers will be merged into a single one. The aligned right image is then used for creating the pan band for pansharpening the left image. All the subsequent steps (stereo and disparity generation) will be the same as mentioned before.

[0060] In FIG. 16, the original image and the pansharpened left image are compared. A specific area of FIG. 16 is zoomed out to further illustrate the pansharpened left image as shown in FIG. 17. The resultant stereo image shown in FIG. 18 is formed by using the pansharpened left image and the original right images. Finally, the Depth map is generated by using the pansharpened left image and the original right image as shown in FIG. 19.

[0061] It will be apparent to those skilled in the art that various modifications and variations can be made to the system and method of the present disclosure without departing from the scope or spirit of the disclosure. It should be perceived that the illustrated embodiments are only preferred examples of describing the invention and should not be taken as limiting the scope of the invention.

REFERENCES

[0062] [1] R. Hartley and A. Zisserman, Multiple View Geometry in Computer Vision, Cambridge University Press, 2003. [0063] [2] H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool. SURF: Speeded Up Robust Features, Computer Vision and Image Understanding (CVIU). Vol. 110, No. 3, pp. 346-359, 2008. [0064] [3] B. Ayhan, M. Dao, C. Kwan, H. Chen, J. F. Bell III, and R. Kidd, A Novel Utilization of Image Registration Techniques to Process Mastcam Images in Mars Rover with Applications to Image Fusion, Pixel Clustering, and Anomaly Detection, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol. 10, Issue: 10, Pages: 4553-4564, 2017. [0065] [4] D. G. Lowe, Object recognition from local scale-invariant features. IEEE International Conference on Computer Vision, vol. 2, pp. 1150-1157, 1999. [0066] [5] G. Vivone, L. Alparone, J. Chanussot, M. Dalla Mura, Garzelli, and G. Licciardi, A critical comparison of pansharpening algorithms, IEEE Int. Conf. Geoscience and Remote Sensing (IGARSS), pp. 191-194, July 2014. [0067] [6] J. Zhou, C. Kwan, and B. Budavari, Hyperspectral Image Super-Resolution: A Hybrid Color Mapping Approach, SPIE Journal of Applied Remote Sensing, Vol. 10, 035024, 2016. [0068] [7] C. Kwan, J. H. Choi, S. Chan, J. Zhou, and B. Budavari, Resolution Enhancement for Hyperspectral Images: A Super-Resolution and Fusion Approach, IEEE International Conference on Acoustics, Speech, and Signal Processing, New Orleans, March 2017. [0069] [8] M. Dao, C. Kwan, B. Ayhan, and J. Bell, Enhancing Mastcam Images for Mars Rover Mission, 14th International Symposium on Neural Networks, Hokkaido, Japan, June 2017. [0070] [9] C. Kwan, B. Budavari, M. Dao, B. Ayhan, and J. F. Bell, Pansharpening of Mastcam images, IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Fort Worth, July 2017. [0071] [10] C. Kwan, B. Ayhan, and B. Budavari, Fusion of THEMIS and TES for Accurate Mars Surface Characterization, IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Fort Worth, July 2017. [0072] [11] C. Kwan, B. Budavari, A. Bovik, and G. Marchisio, Blind Quality Assessment of Fused WorldView-3 Images by Using the Combinations of Pansharpening and Hypersharpening Paradigms, IEEE Geoscience and Remote Sensing Letters, Volume: 14, Issue: 10, pp. 1835-1839, 2017. [0073] [12] B. Ayhan, M. Dao, C. Kwan, H. Chen, J. F. Bell III, and R. Kidd, A Novel Utilization of Image Registration Techniques to Process Mastcam Images in Mars Rover with Applications to Image Fusion, Pixel Clustering, and Anomaly Detection, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol. 10, Issue: 10, Pages: 4553-4564, 2017. [0074] [13] C. Kwan, J. Zhou, and B. Budavari, A New Pansharpening Approach for Hyperspectral Images, Colorimetry and Image Processing, InTech, 2018. [0075] [14] Y. Qu, H. Qi, B. Ayhan, C. Kwan, and R. Kidd, Does Multispectral/Hyperspectral Pansharpening Improve the Performance of Anomaly Detection? IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Fort Worth, July 2017. [0076] [15] X. Li, C. Kwan, and B. Li, Stereo Imaging with Uncalibrated Camera, Advances in Visual Computing, Second International Symposium, ISVC 2006, Lake Tahoe, N.V., USA, Nov. 6-8, 2006. [0077] [16] Y. Qu, H. Qi, B. Ayhan, C. Kwan, and R. Kidd, Does Multispectral/Hyperspectral Pansharpening Improve the Performance of Anomaly Detection? IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Fort Worth, July 2017.