Classifications

• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
  • H04N23/60—Control of cameras or camera modules
  • H04N23/68—Control of cameras or camera modules for stable pick-up of the scene, e.g. compensating for camera body vibrations
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
  • H04N23/60—Control of cameras or camera modules
  • H04N23/68—Control of cameras or camera modules for stable pick-up of the scene, e.g. compensating for camera body vibrations
  • H04N23/681—Motion detection
  • H04N23/6812—Motion detection based on additional sensors, e.g. acceleration sensors
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
  • H04N23/60—Control of cameras or camera modules
  • H04N23/68—Control of cameras or camera modules for stable pick-up of the scene, e.g. compensating for camera body vibrations
  • H04N23/682—Vibration or motion blur correction

Definitions

• the invention relates to the stabilization of a sequence of images acquired by an observation device of an imaging system such as a video camera, subjected to involuntary movements or vibrations.
• the observation device typically operates at wavelengths of the visible or infrared range.
• a first solution relates to video cameras on a terrestrial, naval or airborne platform that are generally mounted on a steerable mount, for example of theodolite type or in a mirror with two axes orientable mirror.
• Servos ensure the stabilization of the line of sight by exploiting the information of a gyroscope or gyrometers integral with the camera.
• Another solution relates to portable light cameras of the camcorder or binocular type (intensified or infrared) which are manually oriented and stabilized by the observer but remain subject to the residual erratic movements of the latter.
• Solutions for stabilizing the line of sight of the camera are implemented by combining two gyrometric sensors integral with the camera housing to an opto-mechanical component deviating variable-angle prism type or eccentric lens by actuators included in the lens from the camera. In these two cases, we speak of stabilization "a priori”.
• Other solutions use image processing to recalibrate the successive images of a video sequence. In this case, we speak of "posteriori" electronic stabilization.
• the object of the invention is to be able to use a low-cost image stabilization process operating whatever the content of the image (contrasted or not, structured or not, ).
• the principle of the image stabilization method according to the invention is based on a combination of prior stabilization and post-stabilization.
• This first stabilization is obtained by determining the attitude of the camera.
• the subject of the invention is a method of stabilizing the images of a scene, acquired by means of an observation device of an imaging system, comprising a step of digital processing of a stream of successive images. It is mainly characterized in that it comprises a step of acquisition of gyrometric measurements by means of at least one gyrometric sensor integral with the observation device, of using these gyrometric measurements to determine so-called approximate shifts experienced between images successive, and in that the image processing step comprises a substep of using the approximate offsets and the stream of acquired images to determine said shifts fine between successive images.
• the advantage of the hybrid motion estimation is thus to allow the stabilization of the image sequence (of the image flow) whatever the scene.
• the image processing step comprises a sub-step of evaluating the fine offsets to choose between the approximate offsets and the fine offsets, those to be applied to the image stream to stabilize it.
• the image processing step comprises a substep of temporal filtering offsets.
• the offset calculations are obtained for example by correlation.
• the subject of the invention is also a device for stabilizing the images of a scene comprising an observation device for an imaging system and a digital image processing element, characterized in that it also comprises at least one gyrometric sensor integral with the observation device and the image processing element.
• the image processing element comprises a microprocessor capable of implementing shift calculations between successive images.
• this logical network programmable device being able to acquire gyrometric measurements of the sensors and an image stream of the observation device and preferably another programmable logic array disposed at the output of the microprocessor and connected to the observation device and able to apply the offsets provided by the microprocessor to the flow of images from the first programmable logic network.
• these programmable logic networks are grouped into a single programmable logic array.
• FIG. 1 schematically represents a stabilization device according to FIG.
• the invention in FIG. 2 schematically represents the various steps of the hybrid estimation of the method according to the invention.
• FIG. 3 schematically represents the different steps of the second level of registration of the hybrid estimation. From one figure to another, the same elements are identified by the same references.
• FIG. 1 An example of a stabilization device according to the invention is shown in FIG. 1. It comprises a video camera 1 and two gyrometric sensors 2 integral with the camera.
• the image of the camera comprising pixels distributed in lines (x axis) and columns (y axis); there is preferably one sensor for the pixel lines and another for the columns.
• These sensors are typically MEMS (acronym for the English expression "Micro Electro Mechanical System”) made on silicon. They are connected to an FPGA 3 (acronym for the English expression Field Programmable Gâte Array).
• an FPGA is a gate array network programmable logic by the user, used for a particular function.
• the FPGA delivers a stream of images and a stream of gyrometric measurements; these flows are synchronized so that the offset measured by the gyrometers is related to the corresponding image.
• This processing of synchronized gyro-images and image flow consists mainly of a so-called hybrid motion estimation, the purpose of which is to provide an estimate of the parasitic movements between successive images. More specifically, the microprocessor 4 calculates the offsets to compensate for geometrically recalibrating the images relative to each other.
• the purpose of stabilization is to correct the involuntary movements of the camera in 3D space. Perfect correction is achieved by combining rotation and translations following the three axes (x, y, z). In the context of the correction of vibrational movements of small amplitudes and high frequencies, a correction in translation along the axes of the image (axes x and y) is sufficient.
• the microprocessor 4 supplies to another FPGA 5 the calculated offsets.
• This FPGA whose function is the geometric transformation of the acquired image stream, then produces a stream of stabilized images.
• a single FPGA is used to perform the functions of the two FPGAs.
• the flow of measurements from the gyrometric sensors has a frequency higher than the flow of the video camera.
• This gyrometric flow is filtered temporally in order to reduce the noise of the low-cost sensors, synchronized with respect to the acquisition of the images (ie an offset along the x-axis and an offset along the y-axis are associated with each image) and the angular values given by the gyrometers are converted into number of pixels according to the field of the camera.
• An approximate offset is obtained, that is to say with a precision of the order of 4 to 6 pixels for a field of about 4 °, or 2 to 3 pixels for a field of about 8 °; it is intended to be exploited by the second level of registration (registration by image processing).
• the second level of registration the fine registration (of a precision less than one pixel for example), is carried out by image processing, for example by correlation of successive images two to one. two.
• This second level uses the first level of registration to initialize its estimate, this initialization thus making it possible to save processing time and make the measure reliable by imposing the consistency of the two estimates.
• This second level of registration is detailed below.
• This estimate of offsets can be seen as a hierarchical process producing a first approximate estimate by gyro measurements which is then refined by image processing possibly several times, according to increasingly finer resolutions.
• This principle allows in particular to preserve a low calculation complexity for the process whereas a simple correlation would prove very expensive.
• the third step consists in merging the two previous offset levels so as to combine the accuracy of the image processing estimation with the reliability of the gyro estimation.
• the aim here is to provide the best estimation of the movements according to the context: fast or slow movements, of high or low amplitude, content of the contrasted or unstressed image, structured or not, etc.
• the fine registration process includes a self-evaluation method detailed below and aimed at estimating the reliability of the processing, for example by controlling the level or the local curvature of a peak of correlation. Depending on the result of this self-evaluation, one or the other level of registration is preferred:
• the reliability is for example defined according to a minimum threshold on the correlation result and a minimum threshold on the correlation gradient.
• the advantage of the hybrid motion estimation is thus to allow the stabilization of the image sequence (of the image flow) whatever the scene.
• the last step is the temporal filtering of the different offsets measured over time, the purpose of which is to separate the different components of the offset to eliminate the offsets that are caused by involuntary movements and keep only those resulting from voluntary movements of the line of sight as a movement of panoramic movement for example.
• the first correspond to displacements of small amplitudes and high temporal frequencies whereas the second correspond to displacements of large amplitudes and low temporal frequencies.
• the step of the second level of registration fine registration
• This figure illustrates an exemplary embodiment of the fine estimation of the image processing offsets based on the principle multi-resolution correlation:
• the current image ⁇ t and the previous image I M are first zoomed out, that is to say undergo a resolution reduction step of a factor N.
• This step includes for each image I t and I M two stages: a low-pass filtering of the image, a subsampling by which only one pixel is conserved
• I 0 the image zoomed out, (note I Dt the image l t zoomed and drank the image read out) i, j row and column indices of pixels in the image.
• the zoomed images lo t and I DM are then correlated to provide a so-called sub-resolved offset estimate.
• any function is used to quantify the similarity between a reference image bt and a secondary image l D n.
• a distance function ie dissimilarity function
• Minkowski metric type ie dissimilarity function
• 7 C 7 D, ⁇ Dt-I) ⁇ ⁇ 7 Dt (UJ) - 7 Dt-I (U j) noting: p the order of the distance, lot and I Dt- i the two images to be compared, i and j the row and column indices of the pixels in the image, N the number of lines of the image and M the number of columns.
• the web or webs of correlation distance noted Nd is constructed by moving the lo reference image correlation carrier t
• H x and H y the correlation horizons in x and y.
• the indices u, v corresponding to the minimum on the distance layer give the estimated offset between the reference image lo t and the secondary image I Dt- 1 -
• the correlation horizons Hx and Hy determine the maximum amplitude of the displacements estimable by image processing.
• the hierarchical process is as follows:
• the gyrometers make it possible to obtain an approximate offset with a precision of the order of 4 to 6 pixels in a small camera field, obtained at the end of a very low computation complexity and much less important than that of an exhaustive correlation.
• the low-resolution correlation that is to say performed on the zoomed images (see Fig. 3), initialized by the estimation of the gyrometers and using horizons of the order of the accuracy of the gyrometers, allows to obtain a precision of the order of 2 pixels.
• the high-resolution correlation carried out on the images l t and
• a- zooming the successive images k times to obtain k reduced resolution image streams, k being an integer greater than or equal to 1, b- calculating offsets from the kth reduced resolution image flow and approximate offsets, where k is strictly greater than 1, iterate the following process: compute offsets from the (k-1) th reduced resolution image flow and offsets previous ones, k k-1 d- calculate the fine offsets from the previous offsets and the original image flow. "Unzoom" means to reduce the resolution.
• the hierarchical process thus works by successive refinements of the motion estimation, allowing gain in time of computation (the low-resolution correlation being of reduced complexity compared to the high-resolution correlation) and gain in reliability.
• a final step of self-assessment of the correlation quality makes it possible to determine the overall reliability of this process, reliability used during the merger to make the decision on whether or not to use this motion estimation by image processing.
• the reliability of motion estimation by image processing is estimated through the qualification of the correlation web. This is to validate that the peak correlation (maximum correlation or minimum distance Lp) is sufficiently marked.
• information of: height of the peak is used simultaneously, for example by analyzing that the difference, in absolute value, between the height of the peak of correlation and the average on the correlation layer is greater than a threshold, curvature on the peak, for example by analyzing for each of the directions x and y, that the difference in absolute value between the height of the correlation peak and the average of its two immediate neighbors is greater than a threshold.

Landscapes

• Engineering & Computer Science (AREA)
• Multimedia (AREA)
• Signal Processing (AREA)
• Studio Devices (AREA)
• Image Analysis (AREA)
• Image Processing (AREA)
• Testing, Inspecting, Measuring Of Stereoscopic Televisions And Televisions (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=37667204

Family Applications (1)

Country Status (4)

Families Citing this family (53)

Family Cites Families (10)

• 2006
  • 2006-06-29 FR FR0605885A patent/FR2903200B1/fr active Active
• 2007
  • 2007-06-25 EP EP07765590A patent/EP2033162A2/de not_active Withdrawn
  • 2007-06-25 US US12/306,862 patent/US8111294B2/en active Active
  • 2007-06-25 WO PCT/EP2007/056294 patent/WO2008000711A2/fr not_active Ceased

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events