 Research
 Open Access
Realtime video stabilization without phantom movements for micro aerial vehicles
 Wilbert G Aguilar^{1}Email author and
 Cecilio Angulo^{1}
https://doi.org/10.1186/16875281201446
© Aguilar and Angulo; licensee Springer. 2014
 Received: 17 December 2013
 Accepted: 5 September 2014
 Published: 22 September 2014
Abstract
In recent times, micro aerial vehicles (MAVs) are becoming popular for several applications as rescue, surveillance, mapping, etc. Undesired motion between consecutive frames is a problem in a video recorded by MAVs. There are different approaches, applied in video postprocessing, to solve this issue. However, there are only few algorithms able to be applied in real time. An additional and critical problem is the presence of false movements in the stabilized video. In this paper, we present a new approach of video stabilization which can be used in real time without generating false movements. Our proposal uses a combination of a lowpass filter and control action information to estimate the motion intention.
Keywords
 Video stabilization
 Micro aerial vehicles
 Real time
 Filter
 Motion intention
Introduction
The growing interest in developing unmanned aircraft vehicles (UAVs) is due to their versatility in several applications such as rescue, transport, or surveillance. A particular type of UAV that becomes popular nowadays are micro aerial vehicles (MAVs) by their advantage to fly in closed and reduced spaces.
Robust guidance, navigation, and control systems for MAVs[1] depend on the input information obtained from onboard sensors as cameras. Undesired movements are usually generated during the fly as a result of complex aerodynamic characteristics of the UAV. Unnecessary image rotations and translations appear in the video sequence, increasing the difficulty to control the vehicle.
There are multiple techniques in the literature[2–5] designed to compensate the effects of undesired movements of the camera. Recently, the video stabilization algorithm ‘L1 Optimal’ provided by the YouTube editor was introduced in[6]. Another interesting proposal is the Parrot’s Director Mode, implemented as an iOS application (iPhone operative system) for postprocessing of videos captured with Parrot’s AR.Drones.
Usually, offline video stabilization techniques are divided in three stages:

Local motion estimation

Motion intention estimation

Motion compensation
Local motion estimation
In this phase, the parameters that relate the uncompensated image and the image defined as reference are determined frame by frame. Optical flow[7, 8] and geometric transformation models[9–11] are two common approaches for local motion estimation. Our algorithm uses the latter one.
Geometric transformation models are based on the estimation of the motion parameters. For this estimation, interest points should be detected and described. A list of techniques performing this task can be found in the literature[12–14], but Binary Robust Invariant Scalable Keypoints (BRISK)[15], Fast Retina Keypoint (FREAK)[16], Oriented FAST and Rotated BRIEF (ORB)[17], Scale Invariant Feature Transform[18] (SIFT), and Speeded Up Robust Feature (SURF)[19] are common in solving computer vision problems[20]. We are using SURF in this phase as a stateoftheart algorithm because our contribution is not focused on reducing delays due to the calculation of interest points, not being significant. The delay due to smoothing techniques is higher.
The second part of the motion estimation process is interest points matching across consecutive frames. This is a critical part because the estimated motion parameters are directly dependent on the reliability of matched points. False correspondences will be removed using an iterative technique called Random Sample Consensus (RANSAC), a widely used technique based on a model from a set of points[21–24]. In our work, RANSAC uses a simple cost function based on gray level difference, minimizing the delay.
Motion intention estimation
In a second phase, for ensuring coherence in the complete motion sequence, the parameters estimated previously are validated in the global and not just in the relative motion between consecutive frames. The main objective of motion intention estimation is to obtain the desired motion in the video sequence suppressing highfrequency jitters from the accumulative global motion estimation.
Several motion smoothing methods are available for motion intention estimation such as particle filter[10], Kalman filter[11], Gaussian filter[25, 26], adaptive filter[26, 27], spline smoothing[28, 29], or point feature trajectory smoothing[30, 31]. In our approach, the control signal sent to the MAV is dealt as a known information; hence, a new and different methodology to those in the literature is considered. A combination of a secondorder lowpass filter, using as few frames as possible, and action control input is employed to estimate a reliable motion intention. We achieve to reduce the number of frames (time window) required for the smoothing signal using an optimization process.
Most of the techniques cited perform well in video stabilization applications, but there is an additional challenge, not studied in the literature and introduced in this paper, which we have called ‘phantom movement’.
Phantom movement
A phantom movement is mainly a false displacement generated in the scale and/or translation parameters due to the compensation of the highfrequency movements in the motion smoothing. Sometimes the motion smoothing process removes real movements and/or introduces a delay in them. Both cases are defined as phantom movements. This phenomenon represents a problem when teleoperating the MAV, and its effects in other stateoftheart algorithms will be shown in the ‘Results and discussion’ section.
Additionally, our proposal to solve this problem will be explained in the ‘Realtime video stabilization’ section.
Motion compensation
Finally, the current frame is warped using parameters obtained from the previous estimation phase to generate a stable video sequence.
This paper is organized as follows: the estimation of local motion parameters in our method is explained in the next section. In addition, we describe the combination of RANSAC and gray level differencebased cost function for robust local motion estimation. In the ‘Motion intention estimation’ section, we present a motion smoothing based on a lowpass filter. Then, the ‘Realtime video stabilization’ section focuses on the optimization of the algorithm with minimum number of frames to estimate the motion intention. Furthermore, we propose a novel approach to solve the problem of phantom movement. Experimental results and conclusions are presented in the last section.
Robust local motion estimation
To obtain a stabilized video sequence, we estimate the interframe geometric transformation, i.e., the local motion parameters. In this phase, we determine the relationship between the current and the reference frame as a mathematical model. This process can be structurally divided into two parts: (a) interest point detection, description, and matching and (b) interframe geometric transformation estimation using matched points. Additionally, an extra process to ensure robustness is (c) robust cumulative motion parameters.
Interest point detection, description, and matching
As mentioned in the latter section, there are several techniques for detecting and matching interest points. According to the results presented in[20], the computational cost of SURF is considerably lower than that of SIFT, with equivalent performance. Using the Hessian matrix and a spacescale function[32], SURF locates waypoints and describes their features using a 64dimensional vector. Once the vector descriptors are obtained, the interest point matching process is based on minimum Euclidian distance in the 64dimensional feature space.
Interframe geometric transform estimation using matched points
where I_{sp} = [x_{sp},y_{sp},1]^{ T } and I_{ t } = [x_{ t },y_{ t },1]^{ T } are the coordinates of the interest points at the reference image and the uncompensated image, respectively, and H_{ t } is the 3 × 3 geometric transformation matrix. How to chose the reference image will be analyzed in the next subsection.
The geometric transformation represented by H_{ t } can be characterized by different models as appropriate. The most common models are translation, affine, and projective models.
In previous works[36, 37], we have found experimentally that in both cases, handheld devices and onboard cameras of flying robots, most of the undesired movements and parasitic vibrations in the image are considered significant only on the plane perpendicular to the roll axis (scale, rotation, and translations). This type of distortion can be modeled by a projective transformation, and we use the affine model in our algorithm as a particular case of the projective one[38]. The benefit is twofold: a lower computation time than for the projective model and its ability for direct extraction of relevant motion parameters (scale, rotation roll, and translations in the xyplane).
Robust cumulative motion parameters
Affine transform can be estimated from three pairs of noncollinear points, but SURF and other techniques obtain hundreds of these pairs of points. RANSAC is performed iteratively N times using three pairs for each H_{ j }, obtaining N different affine transforms. The value of N is based on the required speed of the algorithm. An alternative is to use an accuracy threshold, but this procedure can be slower.
The affine model has been selected as the geometric transform between two frames, the one to be compensated and another used as reference, and there are several alternatives for selecting the reference frame. An experimental comparative study was carried out in[37] on three candidates to be the reference frame: the initial frame (Frame_{sp} = Frame_{0}), the previous frame (Frame_{sp} = Frame_{i1}), and the compensated previous frame (Frame_{sp} = Frame i1′). The analysis for the three proposed approaches was performed by using data obtained from an onboard camera of a real micro aerial vehicle. The obtained results show that the approach based on the previous frame is the best candidate to reference.
Finally, the transformation matrix H_{opt} is calculated and applied on the current frame for obtaining a warped frame similar to the reference frame, i.e., a stable video sequence.
Motion intention estimation
The RANSAC algorithm, based on the minimization of the gray level difference, is enough for obtaining a high accuracy in the image compensation of static scenes (scenes without moving objects)[36, 37]. However, our goal is to achieve a robust stabilization of video sequences obtained with onboard cameras in microaerial vehicles. Most of the unstable videos captured with either flying robots or handheld devices contain dynamic scenes (scenes with moving objects) mainly due to the camera motion. In this way, some movements of the capture device should not be eliminated, but softly compensated, generating a stable motion video instead of a static scene.
The process of approximating the capture device’s movements is known as motion intention. Several video stabilization algorithms use smoothing methods for the motion intention estimation such as Kalman filter, Gaussian filter, and particle filter. Our approach is based on a secondorder Butterworth filter, a lowpass filter used for smoothing of signals[39].
Our experimentation platform is a lowcost MAVs, which shows a complex dynamic behavior during indoor flight. Consequently, the videos captured with onboard cameras usually contain significant displacements and high speed movements on the plane perpendicular to the roll axis. Effects of wireless communication problems such as framebyframe display, lowfrequency videos, and video freezing should be also considered.
Using the lowpass filter as a motion intention estimator, most of the problems associated to indoor flight are avoided. However, freezing effects can eventually still appear by a low communication quality. Motion parameters computed from frozen frames must be discarded before to continue with the estimation process.
Once the affine transformation parameters (scale, rotation, and translations x and y) are extracted, as well as the values of parameters from frozen screens are removed, the lowpass filter computes the motion intention as an output without highfrequency signals. Low frequencies are associated to the intentional motion, and high frequencies are referred to undesired movements, thus the cutting frequency depending on the application and system characteristics. For cutting frequency, a higher value means an output video similar to the original movements, including the undesired movements, while a lower value means that the output video eliminates intentional movements. In our case, we use a secondorder filter with the same cutting frequency 66.67 Hz to smooth the signals of the four motion parameters. An alternative option is to use a different filter for each motion parameter.
Realtime video stabilization
A robust postprocessing algorithm for video stabilization has been detailed; however, the goal is a realtime version. In this context, it is worth noting that there are very few techniques for realtime video stabilization, and the first challenge to be solved being computational cost. Hence, calculation time is minimized in[40] by using efficient algorithms of interest point detection and description. This method reduces time in motion intention estimation by means of a Gaussian filter without accumulative global motion, using the stabilized frames in addition to the original frames. Our proposal uses an offline optimization process for obtaining the minimum number of frames that can be applied in real time to the system at hands without decrease in initial offline video stabilization performance. Furthermore, this filter will be combined with the known control action signal in order to eliminate the socalled ‘phantom’ movements (in fact, a sort of ‘freezing’) in the compensated video.
Optimized motion intention estimation
being the mean square error between monochromatic images with size M · N and Ip_{MAX} the maximum pixel intensity in the frame.
Based on the optimization of the objective evaluation metric ITF, two solutions have been obtained with a performance as high as using the complete video sequence to estimate the motion intention: a) using four (4) previous and four (4) posterior frames and b) using only six (6) previous frames.
Our work is focused on the realtime application, so it is important to analyze how this issue is affected for both options: a) For the first case, to use four previous and four posteriors frames means that the algorithm will be launched four frames after the video sequence initialization, and the stabilized sequence will be ready after four frames. b) In the second case, the algorithm starts six frames after the video sequence initialization, two frame later than in the first case, but for the rest of the sequence, it can be applied without added delay. Considering that the sample frequency was 10 Hz for the system at hands, a 0.4s delay would be introduced in the total computational time when using the first option. Our algorithm uses the second option, which relies only on precedent information.
Phantom movements
Previous video stabilization approaches have obtained good results eliminating undesired movements in images captured with handheld devices and complex systems, but all of them were evaluated using the ITF as cost function. Although the final video has achieved a good ITF performance, i.e., a stable video, the motion smoothing process has generated phantom movements.
For video postprocessing applications, the main objective is to stabilize the video; hence, phantom movements do not represent a problem. Notwithstanding, for realtime applications, the objective is to obtain a stable video, but it should be as real as possible. In this sense, it is important to decrease the difference between the real and estimated motion intention, preserving the ITF performance.
The root mean square error (RMSE)[43] is adopted in order to evaluate the reliability of the estimated motion with respect to the observed motion. A low RMSE means that the estimated motion intention is similar to the real motion intention.
where E_{x,j} and E_{y,i} are the estimated global motion of the j th frame in the xaxis and yaxis, respectively, T_{x,j} and T_{y,i} are the observed motions of the j th frame in the xaxis and yaxis, respectively, and F denotes the number of frames in the sequence.
Two alternatives exist to merge information when computing real motion intention (T_{x,j}, T_{y,i}): information obtained from the onboard inertial measurement unit (IMU) or the control action data. Choice depends on the accuracy of the model. In our algorithm, the control action is employed since IMU information is not very reliable in most of the micro aerial vehicles. In this way, the observed motion is defined as a combination between the control action and the smoothed motion signal.
Results and discussion
This section has been divided into three parts: experimental design, video stabilization performance, and comparison with another algorithm.
The experimental design
The AR.Drone 1.0, a lowcost quadrotor built by the French company Parrot (Paris, France), has been used as experimental platform for several reasons: low cost, energy conservation, safe flight, and vehicle size. The proposed methodology has been implemented in a laptop with the following characteristics: Intel Core i72670QM processor, 2.20 GHz with Turbo Boost up to 3.1 GHz and RAM 16.0 Gb. Real images of four different scenarios are obtained with the onboard camera (sample frequency = 10 Hz) and processed. Furthermore, a video has been recorded with a zenith camera to capture the real motion of the flying robot in the xyplane. RMSE is selected as objective measure of motion reliability, comparing the estimated motion with the observed motion. In order to obtain a position measure, we use a tracker based on optical flow[44] and camera calibration method for radial distortion[45]. Next, the RMSE is computed by comparing the estimate with the observed motion (from the zenith camera).
Video stabilization performance
Presence of nearby objects
Nearby objects in the scene represents one of the main problems of video stabilization, because most of interest points are generated in the objects’ region. The image compensation is computed using the objects’ motion instead of the scene motion. However, our process of matching interest points is based on the RANSAC algorithm and the gray level difference between consecutive frames as cost function. Consequently, the process of motion estimation is not performed on the objects’ interest points but on the whole scene.
Scenes with moving objects
Moving objects are another common problem. Some objects with many points cause, during the motion estimation, undesirable tracking of these objects. Once more, the RANSACbased process of matching interest point is not only referenced to moving objects but to the whole image.
Problems from onboard cameras for quadrotors
Scenes frame by frame, significant displacements, lowfrequency videos, freezing, and highspeed displacements are frequent problems in images captured with an onboard camera due to the complex dynamic of the quadrotors during indoor flight. In all of them, the change between two consecutive frames could be considerable, producing a critical problem in video stabilization. In our approach, motion intention estimation solves these problems, and previous rejection of data higher than a threshold provides additional robustness.
Phantom movements
They corresponds to a phenomenon present in previous video stabilization techniques, but not still reported. Independently of which approach is used, video stabilization process depends on a phase of motion intention estimation. The phantom movements are generated during the elimination of the highfrequency movements due to the motion intention estimation which reduces the frequency of the movements and the previous motion intention estimators which are not able to detect or correct these troubles. Our proposal eliminates this phantom movements using a combination of a lowpass filter, as a motion intention estimator, with the control action. However, this method slightly decreases the ITF value.
Comparison
Evaluation metrics
Video name  Evaluation metric  Original  L1Optimal  Our approach 

Video 1  ITF (dB)  14.09  19.62  19.48 
RMSE  0.046  0.028  
Video 2  ITF (dB)  13.43  19.57  19.52 
RMSE  0.051  0.023  
Video 3  ITF (dB)  14.65  20.16  19.89 
RMSE  0.047  0.029  
Video 4  ITF (dB)  16.96  21.24  21.12 
RMSE  0.036  0.017 
The obtained results show that our algorithm is comparable with the L1Optimal method. The performance of our approach with respect to the ITF measure is slightly lower than L1Optimal, which means that the output video is a little less stable. However, the RMSE value, which represents the similarity between the stabilized video motion and the intentional motion, is higher in our approach.
Conclusions
After conducting an initial study of motion smoothing methods, it has been experimentally checked that the lowpass filter has a high performance as algorithm for motion intention estimation, eliminating undesired movements.
However, this method can be optimized using a lower number of frames without decreasing the ITF measure, as we have presented in this paper. The cutting frequency depends on the model characteristics; hence, the information about the capture system implicates a considerable contribution in a calibration phase.
The phantom movements are a phenomenon that had not yet been studied in the video stabilization literature, but it is a key point in the control of complex dynamic systems as micro aerial vehicles, where the realism in the movements could mean the difference that prevents an accident.
The reliability factor is adapted to the purpose of the application. This application can be a postproduction with a high ITF value and a lower realism or the opposite situation, for realtime video stabilization used in teleoperation systems.
As a future work, we will extend our video stabilization method, using the quadrotor model estimated in[46], for aggressive environments with turbulence and communication problems. We will apply it for increasing the performance of detection and tracking algorithms. In[47] we presented a first application for face detection.
Endnote
^{a}Video results are provided:http://www.youtube.com/user/VideoStabilizerMAV/videos.
Declarations
Acknowledgements
This work has been partially supported by the Spanish Ministry of Economy and Competitiveness, through the PATRICIA project (TIN 201238416C0301). The research fellow Wilbert G. Aguilar thanks the funding through a grant from the program ‘Convocatoria Abierta 2011’ issued by the Secretary of Education, Science, Technology and Innovation SENESCYT of the Republic of Ecuador.
Authors’ Affiliations
References
 Kendoul F: Survey of advances in guidance, navigation, and control of unmanned rotorcraft systems. J. Field Robot 2012, 29(2):315378. 10.1002/rob.20414View ArticleGoogle Scholar
 Duric Z, Rosenfeld A: Shooting a smooth video with a shaky camera. Mach. Vis. Appl 2003, 13(5–6):303313.View ArticleGoogle Scholar
 Battiato S, Gallo G, Puglisi G, Scellato S: SIFT features tracking for video stabilization. In 14th International Conference on Image Analysis and Processing, 2007. ICIAP 2007. Modena, 10–14 Sept; 2007:825830.Google Scholar
 Hsu YF, Chou CC, Shih MY: Moving camera video stabilization using homography consistency. In 2012 19th IEEE International Conference on Image Processing (ICIP). Lake Buena Vista,30 Sept–3 Oct; 2012:27612764.View ArticleGoogle Scholar
 Song C, Zhao H, Jing W, Zhu H: Robust video stabilization based on particle filtering with weighted feature points. Consum. Electron. IEEE Trans 2012, 58(2):570577.View ArticleGoogle Scholar
 Grundmann M, Kwatra V, Essa I: Autodirected video stabilization with robust l1 optimal camera paths. In 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Colorado Springs, 20–25 June; 2011:225232.Google Scholar
 Chang HC, Lai SH, Lu KR: A robust and efficient video stabilization algorithm. In 2004 IEEE International Conference on Multimedia and Expo, 2004. ICME ’04. Taipei,27–30 June; 2004:29321.Google Scholar
 Strzodka R, Garbe C: Realtime motion estimation and visualization on graphics cards. IEEE Vis 2004, 2004: 545552.Google Scholar
 Lee KY, Chuang YY, Chen BY, Ouhyoung M: Video stabilization using robust feature trajectories. In 2009 IEEE 12th International Conference on Computer Vision. Kyoto, 27 Sept–4; Oct 2009:13971404.View ArticleGoogle Scholar
 Yang J, Schonfeld D, Mohamed M: Robust video stabilization based on particle filter tracking of projected camera motion. Circuits Syst. Video Technol. IEEE Trans 2009, 19(7):945954.View ArticleGoogle Scholar
 Wang C, Kim JH, Byun KY, Ni J, Ko SJ: Robust digital image stabilization using the Kalman filter. Consum. Electron. IEEE Trans 2009, 55(1):614.View ArticleGoogle Scholar
 Canny J: A computational approach to edge detection. Pattern Anal. Mach. Intell. IEEE Trans 1986, PAMI8(6):679698.View ArticleGoogle Scholar
 Harris C, Stephens M: A combined corner and edge detector. Proceedings of The Fourth Alvey Vision Conference 31 Aug–2 Sept 1988, 147151.Google Scholar
 Miksik O, Mikolajczyk K: Evaluation of local detectors and descriptors for fast feature matching. In 2012 21st International Conference on Pattern Recognition (ICPR). Tsukuba; 11–15 Nov 2012:26812684.Google Scholar
 Leutenegger S, Chli M, Siegwart RY: BRISK: Binary Robust Invariant Scalable Keypoints. In 2011 IEEE International Conference on Computer Vision (ICCV). Barcelona; 6–13 Nov 2011:25482555.View ArticleGoogle Scholar
 Alahi A, Ortiz R, Vandergheynst P: Freak: Fast retina keypoint. In 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Providence; 16–21 June 2012:510517.View ArticleGoogle Scholar
 Rublee E, Rabaud V, Konolige K, Bradski G: ORB: an efficient alternative to SIFT or SURF. In 2011 IEEE International Conference on Computer Vision (ICCV). Barcelona; 6–13 Nov 2011:25642571.View ArticleGoogle Scholar
 Lowe DG: Object recognition from local scaleinvariant features. In The Proceedings of the Seventh IEEE International Conference on Computer Vision, 1999. Kerkyra; 20–25 Sept 1999:115011572.View ArticleGoogle Scholar
 Bay H, Tuytelaars T, Gool L: SURF: Speeded Up Robust Features. In Computer Vision – ECCV 2006. Edited by: A Leonardis, H Bischof, A Pinz, and (eds.). Springer, Berlin; 2006:404417.View ArticleGoogle Scholar
 Luo J, Oubong G: A Comparison of SIFT, PCASIFT and SURF. Int. J. Image Process. (IJIP) 2009, 3(4):143152.Google Scholar
 Fischler MA, Bolles RC: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 1981, 24(6):381395. 10.1145/358669.358692MathSciNetView ArticleGoogle Scholar
 Tordoff B, Murray D: Guided sampling and consensus for motion estimation. In Computer Vision – ECCV 2002. Edited by: A Heyden, G Sparr, M Nielsen, P Johansen, and (eds.). Springer, Berlin; 2002:8296.View ArticleGoogle Scholar
 Choi S, Kim T, Yu W: Performance evaluation of RANSAC family. In Proceedings of the British Machine Vision Conference 2009. London; 7–10 Sept 2009:8118112.Google Scholar
 Derpanis KG: Overview of the RANSAC algorithm. Technical report, Computer Science, York University; 2010.Google Scholar
 Matsushita Y, Ofek E, Ge W, Tang X, Shum HY: Fullframe video stabilization with motion inpainting. Pattern Anal. Mach. Intell. IEEE Trans 2006, 28(7):11501163.View ArticleGoogle Scholar
 Rawat P, Singhai J: Adaptive motion smoothening for video stabilization. Int. J. Comput 2013, 72(20):1420.Google Scholar
 Wu S, Zhang DC, Zhang Y, Basso J, Melle M: Adaptive smoothing in realtime image stabilization. In Visual Information Processing XXI. Baltimore; 24–25 April 2012.Google Scholar
 Wang Y, Hou Z, Leman K, Chang R: Realtime video stabilization for unmanned aerial vehicles. In MVA2011 IAPR Conference on Machine Vision Applications. Nara; 13–15 June 2011:336339.Google Scholar
 Wang Y, Chang R, Chua T: Video stabilization based on high degree bspline smoothing. In 2012 21st Conference on Pattern Recognition (ICPR). Tsukuba; 11–15 Nov 2012:31523155.Google Scholar
 Ryu YG, Roh HC, Chung MJ: Longtime video stabilization using pointfeature trajectory smoothing. In 2011 IEEE International Conference on Consumer Electronics (ICCE). Las Vegas; 9–12 Jan 2011:189190.View ArticleGoogle Scholar
 Jing D, Yang X: Realtime video stabilization based on smoothing feature trajectories. Appl. Mech. Mater 2014, 519—520(1662–7482):640643.Google Scholar
 Mikolajczyk K, Schmid C: Scale & affine invariant interest point detectors. Int. J. Comput. Vis 2004, 60(1):6386.View ArticleGoogle Scholar
 Faugeras O, Luong QT, Papadopoulou T: The Geometry of Multiple Images: the Laws That Govern the Formation of Images of a Scene and Some of Their Applications. MIT Press, Cambridge; 2001.Google Scholar
 Forsyth DA, Ponce J: Computer Vision: a Modern Approach. Prentice Hall, Upper Saddle River; 2002.Google Scholar
 Hartley R, Zisserman A: Multiple View Geometry in Computer Vision. Cambridge University Press, New York; 2003.Google Scholar
 Aguilar WG, Angulo C: Estabilización robusta de vídeo basada en diferencia de nivel de gris. Proceedings of the 8th Congress of Science and Technology ESPE 2013 Sangolquí, 5–7 June 2013.Google Scholar
 Aguilar WG, Angulo C: Robust video stabilization based on motion intention for lowcost micro aerial vehicles. In 2014 11th International Multiconference on Systems, Signals and Devices (SSD). Barcelona; 11–14 Feb 2014:16.View ArticleGoogle Scholar
 Vazquez M, Chang C: Realtime video smoothing for small RC helicopters. In IEEE International Conference on Systems, Man and Cybernetics, 2009. SMC 2009. San Antonio; 11–14 Oct 2009.Google Scholar
 Bailey SW, Bodenheimer B: A comparison of motion capture data recorded from a Vicon system and a Microsoft Kinect sensor. In Proceedings of the ACM Symposium on Applied Perception. SAP ’12. ACM, New York; 2012:121121.View ArticleGoogle Scholar
 Xu J, Chang HW, Yang S, Wang M: Fast featurebased video stabilization without accumulative global motion estimation. Consum. Electron. IEEE Trans 2012, 58(3):993999.View ArticleGoogle Scholar
 Niskanen M, Silven O, Tico M: Video stabilization performance assessment. In 2006 IEEE International Conference on Multimedia and Expo. Toronto; 9–12 July 2006:405408.View ArticleGoogle Scholar
 Kang SJ, Wang TS, Kim DH, Morales A, Ko SJ: Video stabilization based on motion segmentation. 2012 IEEE International Conference on Consumer Electronics (ICCE) 2012, 416417.View ArticleGoogle Scholar
 Fang CL, Tsai TH, Chang CH: Video stabilization with local rotational motion model. In 2012 IEEE Asia Pacific Conference on Circuits and Systems (APCCAS). Kaohsiung; 2–5 Dec 2012:551554.View ArticleGoogle Scholar
 Yilmaz A, Javed O, Shah M: Object tracking: a survey. ACM Comput. Surv 2006., 38(13):Google Scholar
 Zhang Z: A flexible new technique for camera calibration. Pattern Anal. Mach. Intell. IEEE Trans 2000, 22(11):13301334. 10.1109/34.888718View ArticleGoogle Scholar
 Aguilar WG, Costa R, Angulo C, Molina L: Control autónomo de cuadricópteros para seguimiento de trayectorias. In Proceedings of the 9th Congress of Science and Technology ESPE 2014. Sangolquí; 28–30 May 2014.Google Scholar
 Aguilar WG, Angulo C: Estabilización de vídeo en micro vehículos aéreos y su aplicación en la detección de caras. In Proceedings of the 9th Congress of Science and Technology ESPE 2014. Sangolquí; 28–30 May 2014.Google Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.