1 Introduction

Robots and unmanned systems are increasingly rising in the new era of artificial intelligence and fourth industrial revolution [2, 3, 21, 24, 26,27,28,29,30,31,32,33,34]. The advancements in digital image processing improved the development and the intelligence of robots and made them more popular [2, 22]. Computer vision First Person View (FPV) systems is one of the applications that enables remotely control robots. FPV vision systems are useful for indoor and outdoor missions and provide different type of useful feedback. Conventional visual FPV feedback systems utilizing conventional cameras usually cover less than 100 degrees field of view. Cameras could be equipped with fisheye lenses to widen the field of view. The use of fisheye lenses induces a type of optical distortion to the captured images. This type of optical distortion is called barrel distortion and occurs due to the optical design of the lens.

In this paper, the design and implementation of a wide-angle stereo vision system based on two fisheye cameras with low computational cost is presented. The system is suitable for many real-time applications. It is capable of providing 11 fps (frames per second). During the work, several commercial systems and existing algorithms were studied and considered. The system proposed in this paper compared with existing systems in the literature is different. The proposed work presents a stereo vision FPV system that covers 310 degrees field of view and outputs 11 fps using a laptop and 6 fps using a stand-alone embedded computing device. Up to the authors’ knowledge no reported similar systems in the literature that process the fisheye effect from two cameras and produces a stereo image achieves this rate. A comparison between various image processing and computer vision algorithms on a resource constraint embedded device was also conducted. Although some commercial similar systems may exist, these did not present scientific backgrounds or design details, hence, this work contributes to the body of knowledge through scientifically presenting and describing the proposed system.

The rest of the paper is organized as follows. In Section II, related literature is reviewed. Section III covers the design and implementation of the proposed system. Section IV presents experimental results and performance analysis. Finally, Section V concludes the paper and highlights future work.

2 Literature review

Utilizing two cameras to build binocular vision systems has several advantages. One main advantage is the wider field of view when compared to monocular systems. In this section similar existing systems and previous work done in this field are overviewed.

To obtain a wide field of view, some systems are implemented using a fixed wide-angle camera mounted on a moving object. The distorted images are then corrected using various kinds of algorithms to provide natural looking images. The obtained rectified images are passed to detection tools and processes to be applied for different kinds of purposes and applications such as in advanced driver assistance systems (ADAS) to detect the host vehicle or to track objects [1, 7, 14, 37, 42]. The wide field of view is achieved in other systems by using one conventional camera with a convex mirror. The output image is an annulus image. This image is then converted to a rectangular form to obtain a panorama image. The resultant image covers 360 degrees field of view [7].

Other systems are implemented using multiple conventional cameras mounted on a moving tripod. The purpose is to capture a series of successive non-instantaneous overlapping images. Then stitch them together using a state of the shelf stitchers to cover 360 degrees field of view [25, 36, 38].

The systems presented in [5, 11, 35, 39] are stereo vision systems implemented using conventional cameras. These systems are used for different kinds of applications such as 3D point reconstruction and object localization and mapping for autonomous mobile robots.

The work presented by Zhang et al. focuses on stereo matching of fisheye images without distortion correction. According to the authors, the matching algorithm the researchers proposed is expected to meet the need of stereo matching. In their work, were the optical axis of the two fisheye cameras are perpendicular, the researchers concentrate on finding the matching points between two fisheye spherical images for 3D applications. The researchers proposed an algorithm based on two algorithms called Maximally Stable Extremal Regions (MSER) and Affine Scale Invariant Feature Transform (ASIFT). In the reported results the researchers reported best time needed to find 37 matching points to be 3.566 s. This time is based on running the algorithm on a personal computer featuring a dual core processor running at 2.2 GHz and 5GB of memory [41]. The reported processing time did not include the cost of generating a stereo image out of the two captured fisheye images.

In [8, 19, 20] stereo vision systems are implemented to construct the position of vehicles. The systems used two cameras equipped with fisheye lenses. The distorted images are then calibrated and rectified. The stereo matching algorithm is used to compute the 3D points of the rectified images.

Wei et al. presented a robust scheme for fisheye video correction. The proposed method solved the time-varying problem which increases the distortion while the objects are moving. Six distinct but related correction criterions are used. The paper compared the corrected image with other methods. The proposed algorithm was tested on a PC with Intel 2.5 GHz Core 2 CPU and 2 GB memory. This approach supports interactive video processing with a runtime of around 0.42 s per frame (i.e. 2.3 fps) [40].

In [4] a real-time fisheye lens distortion correction system is designed and implemented using a FPGA-based camera. The work generated a complex image processing application and optimization on the source level on the original code to exploit the memory architecture in the FPGA. The performance comparison between the Core 2 Quad and the FPGA implementation was shown. In the Core 2 quad software the fps rate was 5.26 but in the FPGA it was 22 fps.

In [15] the system has been implemented using two back-to-back fisheye cameras to capture 360degrees FoV images. The cameras are exploited by a smartphone system. MediaTek’s smartphone with 5 megapixels and 182 degrees FoV is used to acquire the forms. The acquired images are wrapped and blended into a 4 k panorama image. The presented paper has designed and implemented a wrapping and projection techniques to eliminate the fisheye lens distortion. Moreover, a memory efficient blending technique has been introduced to blend the two fisheye images together. The system achieved 7.8 Mpixel/s, i.e. 1 K resolution with 15 fps by using 2.5 GHz Octa-core CPU, and power VR GPU.

The work presented by Ho et al. has introduced a novel method to align images generated by a dual-fisheye camera by employing interpolation grids based on rigid Moving Least Squares (MLS) to produce seamless stitching panorama images. Besides, they have reduced the jitter in videos that are generated by image-based stitching algorithms by incorporating a new temporal-coherent algorithm to maintain smooth-to-smooth frame transition. The system has been implemented using C++ and Matlab software, and the Rigid MLS has accelerated by GPU. The presented method achieved 360 × 180 degrees FoV seamless stitching panorama images [13].

In [12] proposed a stereo system that provides a panorama image with 360 × 65 degree FoV in horizontal and vertical directions. Also, the presented work addressed some constraints in stereo systems like camera modeling, fisheye intrinsic calibration, stereo self-calibration, and depth estimation. A bundle adjustment based approach and markerless stereo self-calibration method are used to optimize and to reduce the number of calibration parameters. The system introduced a Region of Interest (ROI) extraction modules to obtain distortion-free pinhole images from partial regions in fisheye images. The system was implemented using two mvBlueFox-Mlc202bG2 cameras with image resolutions 1280 × 960, 235degree FoV and 1280 × 1024, 245 degree FoV.

3 Proposed system

The proposed system takes into consideration noisy and low feature images. It is also a low computational complexity system. The proposed system is implementable on a laptop computer and on a resource constraint embedded computing device such as NI myRio-1900 [17]. The proposed system’s performance suits many applications with real-time requirements. The block diagram in Fig. 1 illustrates a high-level presentation of the wide-angle stereo vision system presented in this work.

Fig. 1
figure 1

The block diagram of the proposed system

Full size image

3.1 System hardware

The proposed system utilizes two wide-angle fisheye cameras. The cameras are low-cost car reverse backup rear view mini color Charge-Coupled Device (CCD) cameras. Each one covers 170 degrees field of view with resolution 756(H) × 720(V) pixels.

The dimensions of each camera is 18 mm × 18 mm × 22 mm and powered by a 12 V DC / 200 mA power source. The horizontal angle between the two cameras is 140 degrees as illustrated in Fig. 2. The cameras capture the overlapping images and transmit them via the communication links to the computing system. In this paper the computing system calibrates, corrects and stitches the images to produce a live panorama image suitable for many real-time applications.

Fig. 2
figure 2

The horizontal angle between the cameras

Full size image

The computer used for system development is an Intel® Core™ i7-2670QM CPU @ 2.20 GHz, with 8.00 GB installed memory laptop and Intel® HD Graphics 3000 graphics card. Moreover, the system is tested on a National Instruments (NI) myRIO-1900 embedded computing device. NI myRIO-1900 is a portable Reconfigurable I/O (RIO) device with dimensions of 136.6 mm × 86 mm and weighs 193 g, shown in Fig. 3.

Fig. 3
figure 3

NI myRIO-1900

Full size image

The NI myRIO-1900 Processor is Xilinx Z-7010 operating at 667 MHz. The processor features 2 cores, 512 MB of Nonvolatile memory, 256 MB DDR3 memory with 533 MHz clock frequency and a 16-bit data bus. The used NI myRIO-1900 features two USB ports that will be used to connect the two cameras. Figure 4 illustrates the hardware connection using the laptop and NI myRIO-1900.

Fig. 4
figure 4

System hardware connection

Full size image

3.2 System software

In the proposed system, the horizontal angle between the two fisheye cameras is experimentally determined by trying to reduce the overlapping area as much as possible yet try to cover the widest field of view. Before starting the distortion correction process, a calibration process is carried out to determine the intrinsic and extrinsic parameters of the cameras. After the calibration process, a barrel distortion correction algorithm is implemented. The barrel distortion correction algorithm shall consider the execution time constraints to enable the system to instantaneously handle the two images coming from the two fisheye cameras. The system undistorts every two instantaneous images and then combines them together to obtain a natural looking wide range view that covers 310 degrees field of view. Figure 5 illustrates the proposed system software design to generate a live panorama image from the images captured by the two fisheye cameras. The system starts by calibrating the fisheye lenses. Then the barrel distortion is removed from the acquired overlapped images. After that, the corrected overlapped images are correlated to measure the similarity between the images. Finally, the non- overlapped images are stitched to a composite image.

Fig. 5
figure 5

Software design of the proposed system

Full size image

3.2.1 Fisheye lens calibration

The calibration process is carried out to obtain the intrinsic and extrinsic parameters of the camera [6, 23]. These parameters are used for image correction, distance determination, stereo matching and accurate measurements. Intrinsic calibration maps the camera and the image coordinates to get the camera properties such as focal length, principal point, and lens distortion. Figure 6 illustrates the intrinsic parameters of the camera. Intrinsic parameters determine the projection of a 3D object into a 2D image, therefore determines the relative position of the camera relative to the object’s coordinates. Furthermore, extrinsic parameters define the location and the orientation of the camera recovered by rotation and translation matrices. In multi-camera systems, the extrinsic parameters describe the relative position and attitude of each camera to the other.

Fig. 6
figure 6

The intrinsic parameters of the camera

Full size image

The fisheye lens calibration in this work is performed by acquiring a predetermined circular grid via the wide-angle fisheye camera. The camera is equipped with a 170-degree field of view fisheye lens. The same grid is used later as a reference template for the system. The grid consists of 9 equal circles distributed regularly on an A4 sheet. The radius of each circle equals 2.5 cm. The horizontal and the vertical distances (dx, dy) between two adjacent centers equals 10 cm and 7 cm respectively as shown in Fig. 7.

Fig. 7
figure 7

Calibration grid parameters (a) Grid parameter description and the proposed calibration grid taken by 170 degrees cameras

Full size image

The grid is converted into an 8-bit greyscale image, then thresholding is conducted to produce a binary image. As shown in Fig. 8, the detected dark particles are the circles, boarders, and some other noise distributed in different positions in the image. To remove the unwanted boarders, a particles filter function is applied. This filter removes or keeps the particles depending on predetermined parameters such as particle area. As a result, the particle filter keeps the circles of the grid and removes the boarders and noise. The target point circles’ area ranges from 100 to 10,000 square pixels. The filter is set to remove any particles outside this range and produce a clean grid image. Figure 9a and b illustrate the calibration grid after applying the particles filter and computing the target points in pixels and real-world coordinates.

Fig. 8
figure 8

Illustration for the calibration grid before and after applying thresholding technique a A greyscale threshold calibration grid and b The calibration grid after applying thresholding technique

Full size image
Fig. 9
figure 9

The detected target points after applying the particles filter function and converting circular dots to target points a The target center of mass (x, y) of the circles in pixels and b The target center of mass (x, y) of the circles in real-world coordinates

Full size image

The previously computed reference points are then passed to a distortion learning calibration function. This function learns the distortion model of the camera and the lens setup. Generally, learning the image makes a map between the pixels’ coordinates and the real-world coordinates. After the learning process is complete, the function is applied on all acquired images by the cameras.

3.2.2 Barrel distortion correction

Barrel distortion follows the quadratic form of degree two. This effect appears as curved lines on the captured images. To fit this curve and the radius to a linear shape, different interpolation methods may be used. However, the interpolation uses known data to estimate the value of an unknown point. Hence, to correct the distorted image pixels, the pixel coordinates and their new intensity are required to be projected on the destination image. Pixel coordinates are obtained from the calibration process and the new value of these distorted pixels are obtained from the interpolation method. Some of the widely used interpolation methods that are used to correct barrel distortion are the nearest linear neighbor interpolation and bilinear interpolation [16]. Nearest neighbor interpolation method estimates the value of the interpolated point depending on the closest pixel value. In the bilinear interpolation method, the interpolated point is estimated by the average weight of 2 × 2 nearest neighborhood pixels. However, in the bilinear interpolation method the resultant image appears smoother than the image that is obtained from the nearest neighbor method. But the execution time of the bilinear method is longer than that of the nearest neighbor method. Figure 10a and b show a simple representation of the nearest neighbor interpolation and bilinear interpolation methods, respectively. Because the nearest neighbor interpolation method is faster than the other methods, it is adapted in this work. Figure 11a, b, c, and d show the overlapping images before and after barrel distortion correction, respectively, by using the nearest neighbor interpolation method.

Fig. 10
figure 10

Representation of interpolation methods a Nearest neighbor interpolation method and b Bilinear interpolation method

Full size image
Fig. 11
figure 11

a Left overlapping distorted image b Left overlapping corrected image c Right overlapping distorted image and d Right overlapping corrected image

Full size image

3.2.3 System calibration

In this step, the intersection and similarities between the two captured fisheye images are measured. This process consists of two main parts: (a) finding the overlapping particles in the two corrected images. (b) the found overlapping particles are then processed to figure out the overlapping area and the relative position of the first camera to the second camera.

Particle detection and analysis

The angle between the two cameras is experimentally determined to cover the largest possible field of view and at the same time to maintain the mid area from getting lost or seriously affected by losing details. Because the barrel distortion correction may produce loss in parts of the image near the borders of each image, in this work, the best optimal angle was experimentally found to be 140 degrees horizontally. After obtaining the undistorted images from the cameras, the resulting images are processed as follows:

  • Auto Threshold: Thresholding an image is usually implemented at the first step of the machine vision application depending on the image type. This function separates the image into a particles’ region and a background region. It converts the grey level image into a binary image depending on the objects that the function is looking for. It uses different kinds of techniques to separate the pixels’ intensities such as clustering, entropy, inter variance, metric, and moment techniques. In this work, the clustering threshold technique is used because it can threshold the image into multi-classes. It sorts the histogram of the image into a discrete number of classes to obtain the center of mass of each class. The histogram represents the number of pixels with the same color in a fixed list of color range. Figure 12, represents the cropped corrected overlapping images before and after applying the clustering technique.

Fig. 12
figure 12

The corrected images before and after applying thresholding technique a Half left corrected image b Half right corrected image c Half left corrected image after applying threshold and d Half right corrected image after applying thresholding technique

Full size image

  • Basic Morphology: This function isolates the objects in the image into non-overlapping regions based on the topography surface of the image. It depends on the relative ordering of pixel values not on their numerical values. Furthermore, the auto threshold technique separates the particles from the background. But basic morphology makes the particles in the image clearer and unattached. It separates the adjacent particles in the image depending on a small template that probes the image at all possible locations. This template is called the structuring element. It is a matrix of zeros and ones that fits or hits the input image. If all elements in the matrix are ones then it fits the corresponding pixels under the structuring element. If the matrix elements are zeros and ones, it hits the pixels under the structuring elements. Figure 13 describes the operation of the structuring elements on the image. Where the grey and white blocks represent ‘1’ and ‘0’ pixel values respectively. When probing structuring element #1 on locations A, B and C, it fits just location A and hits location A, B and C. And when probing structuring element #2 on the same locations, it fits and hits A and B. various techniques are used to separate the objects. The reason of using this function is to increase the gap between the particles and remove the thin edges. This makes the particle detection process easier and more accurate. The open objects technique with a 3 × 3 structuring element is used in this work. Figure 14 shows the image before and after applying the basic morphology on the half overlapping corrected images.

Fig. 13
figure 13

The operation of the structuring elements on an image

Full size image
Fig. 14
figure 14

The images before and after applying the basic morphology on the half overlapping corrected images a The half left processed image and b The half right processed image

Full size image

  • Particle removal: After isolating the objects in non-overlapping areas, some small particles persist in the image. These particles may cause inaccurate correlation measurements, and as a result may lead to inaccurate stereo calibration. Therefore, wrong overlapping areas are determined. The function is a filter that removes the large or the small particles in the image. The number of iterations determines the number of erosions to be applied in the image. In the here presented work, the small particles are removed, the number of erosions equal to 9 is experimentally determined. Figure 15 illustrates the corrected overlapping image before and after removing the small objects.

Fig. 15
figure 15

The corrected overlapping image before and after removing the small objects a The half left processed image and b The half right processed image

Full size image

  • Particle analysis: this is the last step before measuring the correlation between the two images. This step provides around eighty measurements for every particle found, such as the center of mass, the bounding rectangle, the object’s area, etc. To measure the correlation between two images the most suitable ten properties are taken, the center of mass pixels (x,y), first pixel (x,y), the bounding rectangle (Top, Left, Right, and Bottom), the area, and the area of the particle over the image area. Figure 16 represents the detected particles of the found objects bounded by rectangles for the left and right corrected overlapping images.

Fig. 16
figure 16

The detected particles of the found objects bounded by rectangles for the left and right corrected overlapping images a The half left image and b The half right image

Full size image

Correlation and offset measurements

The cross correlation measurement is a known concept used to find the similarity between two objects. It finds the average difference between two different objects. In this work, the regional correlation is done by taking the measurements found in the last step of (a), namely: particle analysis. The properties are the center of mass (x,y), first pixel (x,y), surrounding rectangle, the area of particles, and the area of the particle over the image area. These ten values (A) are taken for each particle found from the two images. The first particle found from the left image is correlated with all particles found from the right image, and the second particle found in the left image is correlated with all particles in the right image and so on. Assuming the number of particles found in the left image are five particles and in the right image are five as well, the correlation loop will repeat twenty five times to correlate each particle in the left image with all particles in the right image. Next, the correlation values finding process between the images is described.

According to the following pseudo code, MCVn is computed.

1 set n = 0

2 set m = 0

3 Compute Array Xdiff = |PnL – PmR|

4 Compute Xsum = Add All elements of Xdiff

5 Compute Xavg = Xsum /A

6 Increment m

7 Repeat 3–6 until m = max

8 MCVn = min(Xavgn)

9 Increment n

10 Repeat 2–8 until n = max

Where,

MCVn :

is the minimum correlation value number n.

n :

is the number of particles in the left image.

m :

is the number particles in the right image.

A :

is the number of used properties.

Xdiff :

is a one dimensional array of size (A) containing the resultant values of subtracting two one-dimensional arrays PnL and PmR.

PnL :

is a one dimensional array of size (A) containing the properties mentioned above for particle number n in the Left image.

PmR :

is a one dimensional array of size (A) containing the properties mentioned above for particle number m in the Right image.

Xsum :

is the summation of all elements of Xdiff.

Xavg :

is the average of Xsum over A the number of used properties.

The predetermined ten properties (A) of particle 1 in the left image are subtracted from each particle in the right image then the average summation of each one is computed. The properties values of the particles found in the left and right images are illustrated in Fig. 17. These values are used in the previous pseudo code. The correlation values of the particles in the left image with the others in the right image are 736.703 and 359.525 pixels respectively. The minimum correlation value indicates the maximum particles similarity. According to the values that are obtained from the pseudo code, the left particle (Red) with the right particle (Green) have the maximum similarity. The center of mass (x) of these particles are taken to find the average between them ((106.07+ 148.41)/2 = 127.24 pixels). This average value is used in the next step to measure the image locations.

Fig. 17
figure 17

The resultant properties values for the particles found in left and right image

Full size image

3.2.4 Panorama image stitching

The corrected overlapped images are cropped to focus on the overlapping area. The size of the images before and after cropping are 526 × 225 and 263 × 225 pixels respectively. The (x) average component is found in the system calibration process to be added to the remaining part of the left cropped image. The size of the remaining cropped left image is equal to 263 × 225 pixels. The left image location and the right image location are computed to be (0, 0) and (390, 0) respectively in the destination image. This offset is the right image’s location relative to the left image, or the amount in pixels to shift the right image from the left image as illustrated in Fig. 18. If there is no similarity between the images, then the system will stitch the whole images together without removing any part of them because no overlapping area is found. This correlation is done one time at the beginning of the process, for the first frame from each camera to determine the overlapping area between the cameras.

Fig. 18
figure 18

The resultant panorama image

Full size image

4 Experimental setup and results

In this work, National Instruments (NI) LabVIEW [18] software is used to implement the proposed algorithm. The proposed system and currently available algorithms in the literature are tested using videos captured by wide-angle fisheye cameras.

The system is implemented as shown in the Fig. 1. Each block in the system is designed and tested separately to make sure that each sub-system works properly and gives the expected results. Afterwards, the overall system is integrated and tested over different parameters and setups.

Different templates of grids are used to calibrate the cameras. The circles’ coordinates in the sheet are very important. Hence, the overall points should cover the whole sheet in different positions to make a proper calibration. Increasing the number of circles in the grid increases the accuracy of the calibration, because the calibration process detects these circles and determines the center of mass of each one. Increasing the number of circles, increases the detected coordinates in the grid. As a result, it increases the accuracy of the calibration but at the expense of the processing time, because detecting and calibrating each point in the grid takes a considerable amount of time. Therefore, reducing the processing speed of the overall system. The accuracy of calibration appears if the system detects the dots properly and determines the center of mass (x,y) of each one (pixels and their real-world coordinates). In this work, different image sizes are calibrated and the time of each one is computed. The most suitable calibration grid is taken to test the images. Figure 19 shows the calibration processing time for the different grids using the laptop and the NI myRio-1900. Figure 20 illustrates the calibration processing time using different image sizes on the laptop and on the NI myRio-1900.

Fig. 19
figure 19

The calibration processing time for the different grids using the laptop and the NI myRio-1900

Full size image
Fig. 20
figure 20

The calibration processing time using different image sizes on the laptop and the NI myRio-1900

Full size image

In this work it is required to generate a natural looking image with the shortest possible processing time. As mentioned earlier, the fastest interpolation method has been chosen (nearest interpolation). Another evaluation parameter that must be considered is the scaling factor of the corrected image. This evaluation parameter has two factors, the first one is “scale to preserve the image area” and the second one is “scale to fit”. Scale to preserve area displays the image such that the features of the input image have the same size of the corrected image, but the image size is smaller than the input image. Scale to fit displays the image such that the corrected image size has the same input image size. The processing speed of the scale to preserve image area factor is faster than scale to fit factor. Figure 21 illustrates the barrel distortion correction processing time for both scaling factors applied on the chosen calibration grid on the laptop and on the NI myRio-1900. Table 1 illustrates the processed frames per second using one and two cameras. The calibration grid with a 640 × 480 image size, scale to preserve image area factor, and using nearest interpolation method type is used.

Fig. 21
figure 21

The barrel distortion correction time with the interpolation method and the scaling factor

Full size image
Table 1 The processed frames per seconds for the proposed work using the nearest neighbor interpolation method and scale to preserve area factor
Full size table

Figure 22 shows a sequence of screenshots taken from a live video of the generated panorama image using the proposed system. The sequence is from left to right, up to down. A red rectangle is drawn around an object, an apple iPhone in this case, moving back and forth between the images of the two fisheye cameras in the constructed panorama image after going through all the steps described above. The rectangle is drawn in the first row only to enable the reader to better see the object in the other four rows. Tables 2 and 3 illustrate a comparison of the proposed work with other research.

Fig. 22
figure 22

A sequence of screenshots from a live video with an object moving back and forth between the right and left fisheye images after removing the barrel distortion and generating the panorama image

Full size image
Table 2 Comparison table between the proposed work and other works – Part I
Full size table
Table 3 Comparison table between the proposed work and other works – Part II
Full size table

5 Conclusion and future work

In this work, a wide-angle stereo vision system based on two fisheye cameras is designed and implemented. The two barrel distorted images are calibrated and corrected instantaneously. The resultant images are correlated to measure the similarity between them. Finally, the system stitches the images into a composite one to generate a panoramic view. The proposed system processes the input fisheye images instantaneously and produces the resultant panoramic images as a live video that suites many applications with strict time constraints. The number of panoramic frames produced by the proposed system using the laptop was 11 fps. The frames per second processing speed of the stereo system using the NI myRio-1900 embedded computing system was 6 fps. The system is able to generate panorama images that cover 310 degrees field of view horizontally. The proposed system achieves a frame rate that fits many real-time applications.

For future work, the proposed system could be enhanced by implementing it using Field Programmable Gate Arrays (FPGA). Moreover, using the stereoscopic system gives the ability to produce 3D views of the scenes providing depth perception. Furthermore, extra work could be done on the quality of the produced panorama image and use of evaluation measures to help improve the output stitched image [9, 10]. Another future research direction is to increase the number of cameras to produce a 360 degree field of view in all directions and in real-time.