Active Face Frontalization using Commodity Unmanned Aerial Vehicles
Nagashri Lakshminarayana, Yifang Liu, Karthik Dantu, Venu Govindaraju, Nils Napp
AActive Face Frontalization using CommodityUnmanned Aerial Vehicles
N. Lakshminarayana, Y. Liu, K. Dantu, V. Govindaraju, N. Napp
Abstract
This paper describes a system by which Unmanned Aerial Vehicles(UAVs) can gather high-quality face images that can be used in biometric identi-fication tasks. Success in face-based identification depends in large part on the im-age quality, and a major factor is how frontal the view is. Face recognition softwarepipelines can improve identification rates by synthesizing frontal views from non-frontal views by a process call frontalization . Here we exploit the high mobility ofUAVs to actively gather frontal images using components of a synthetic frontaliza-tion pipeline. We define a frontalization error and show that it can be used to guidean UAVs to capture frontal views. Further, we show that the resulting image streamimproves matching quality of a typical face recognition similarity metric. The sys-tem is implemented using an off-the-shelf hardware and software components andcan be easily transfered to any ROS enabled UAVs.
Surveillance in remote areas is difficult due to the difficulty and cost of building andoperating dense camera networks over large areas. Instead, Unmanned Aerial Vehi-cles (UAVs) can be used as mobile sensors to increase coverage [12], to identify, andto follow individuals. Biometrics, i.e. consisting biological signals, of a person haveproven to be an efficient means of automated identification [13]. Visual biometricssuch as faces, fingerprints, and iris patters, are highly discriminative. However suchmodalities show great appearance variations in nature. For example,the pixel levelcomparison of facial images from the same person can look rather different whilethe biometric signal should be constant. An adaptive biometric system could reducethe efforts for capturing and encoding all such variations. UAVs being highly mobilecan be controlled to obtain high quality biometric data. Typically, pilots in groundstations control UAVs remotely to acquire high quality data. However, this requiresextensive human effort and places constraints on scaling up such approaches, whichconflicts with the goal of covering large areas. Automation of the tracking and UAVnetworking could potentially change this [6].In the past, military has employed ”Drone stares” in war zones and border areasto monitor any malicious activities [30]. Beyond surveillance, UAVs have also beenused for search and rescue operations [28] [26].
Computer Science and Engineering, University at Buffaloe-mail: {nagashri,yifangli,kdantu,govind,nnapp}@buffalo.edu a r X i v : . [ c s . C V ] F e b N. Lakshminarayana, Y. Liu, K. Dantu, V. Govindaraju, N. Napp
Autonomous UAVs that track humans have to detect and identify individuals effi-ciently. Among various biometrics, face is one of the primary modalities for humanidentification. However, the video quality of aerial imagery poses severe limitationsto perform facial recognition. For example [30] lists instances of ”collateral dam-age” caused due to the ambiguity in identifying the person of interest from the aerialimagery.In the biometrics community facial analysis based on facial images from printmedia, on-line photo postings, or other data sources which were captured inciden-tally are called a unconstrained faces. Such data show huge appearance variabilitydue to change in pose, lighting, and occlusions. Such inconsistencies degrade thequality of automatic face detection and analysis. Although humans can efficientlyrecognize such pose variant faces, for a computer to train and learn on such noisyrepresentation requires massive data and a considerable amount time.A common approach of face recognition pipelines is to synthesize a frontal viewof the face using 3D reconstruction methods from a non-frontal image. Howeversince UAVs can be used as active sensors in biometrics to obtain a optimized viewof face, we propose active face frontalization , in which real frontalized data is gatherrather than synthesized from non-frontal views.Besides getting better recognition, frontal images of faces are more aesthetic andcan be used in the UAVs photography technology popular these days. We define ac-tive frontalisation based on the existing frontalisation methods. The contributions ofthis paper are: Firstly, we derive a frontalization error on the pose variant face basedon existing virtual fontalization techniques. We show that this metric can be effi-ciently used to adjust the flight path to obtain an improved view of face. Secondly,we evaluate face recognition performance of the resulting views using a similaritymeasure on the features obtained from a deep convolutional neural network (CNN)as commonly done in state-of-the-art biometric identification. We present an imple-mentation of this approach using inexpensive off-the-shelf components.The resulting system focuses on frontalization and is currently limited by severalpractical limitations related to safety and sensor quality. Autonomous UAVs createssafety hazards. Therefore, we chose a relatively safe UAV over more sophisticatedbut dangerous ones. Further the lack of high quality cameras and sensors in an off-the-shelf UAV needs to be overcome by using robust algorithms.More generally, the images from UAVs are not optimal for biometric analysis.The faces seen by the UAVs have two crucial issues: long range and lateral poses.Generally the perspective of UAVs capture faces at a distance and are not detectable.In the following section 3.2 we describe a simple person following algorithm todirect the UAVs into the range of face detection. Section 1 elaborates the methodused for assessment of the face quality with regard to the pose. In Section 3.5, wegive an overview of the controller. ctive Face Frontalization using Commodity Unmanned Aerial Vehicles 3
Human detection from aerial imagery was performed using thermal signatures inearly days [10] [27]. However due to popularity of computer vision based tech-niques, visual feedback from the sensors were used to track and follow humansefficiently. Pestana [24] investigated whether visual based object tracking and fol-lowing is reliable using a cheap GPS-denied UAVs while assuring safety. This workdemonstrated that current tracking algorithms, such as OpenTLD [14], are reliableto work on UAVs platform, and the proposed architecture has been able to followunmarked targets of different sizes and from a wide range of distances even whenocclusions happened. Although researchers achieved good performance for track-ing and following humans, identification was still a challenging task. Unfortunatelyneither thermal images nor detectors were sufficient for identification of people.Therefore researchers started analyzing the feasibility of using face recognition onUAVs.Unlike static cameras or high resolution cameras, UAVs have disturbance, resolu-tion limitation and complex flying environments (different illumination and weather,indoor and outdoor), current state-of-the-art face recognition methods may have lim-its and issues while they are applied on UAVs, therefore, we need to know how wellface recognition perform on UAVs given different altitude, distance and angle. HSUand CHEN [12] investigated the capability of two face recognition services (
Face ++ [19] and ReKognition [23]) in face detection and face recognition with different al-titudes, distances, and angles of depression. The results showed that the present facerecognition technologies are able to perform adequately on UAVs.Another study was made on analyzing the face quality on various platforms [16].They based their study on three commercially available face analysis algorithms 1)ViolaJones [Viola and Jones 2001] and Rowley [Rowley et al. 1998] face detec-tion algorithms; 2) QDA-based face recognition algorithm [Lu et al. 2003]; and 3)CAMSHIFT [Bradski 1998] face tracking algorithm. It was inferred that the accu-racy of these algorithms change significantly after a critical video quality. Davis etal. [7] present a modular and adaptive algorithms for facial recognition on commer-cial off the shelf UAVs . They use human visual based approach with LBP featuresto train classifiers for facial recognition systems. However such systems are not im-mune to the large pose variations of faces as seen by the UAVs. Currently, commercial following drones, such as Airdog [3], Bebop [1], Hexo+ [2]are popular. Airdog uses a wearable device called AirLeash to track and follow peo-ple. AirLeash can track movement and send control commands to AirDog. Bebophas ”Follow me” feature by using visual recognition technology and GPS trackingsystem on the smart phone. Same as Bebop drone, Hexo also uses smart phone tocontrol and make it track and follow people. For all these tracking drones, the usershave to wear a tracking device to ensure that drones follow them, they are incapableautonomously capturing face images with good quality.
N. Lakshminarayana, Y. Liu, K. Dantu, V. Govindaraju, N. Napp
Face detection is the process of identifying image areas containing faces. Facerecognition refers to matching the detected face to the reference. Thus face detec-tion is the initial stage of recognition. The first limitation for a good face detectionis the range of detection. The size of the faces in images restrict the detections andfaces seen from far are not detectable. To check the performance of face detectionat different distances, we performed an empirical study similar to [12] and obtainedsimilar results.Specifically, in Figure 1 we plot the probability of obtaining a detection with re-spect to the distance maintaining a constant height of eye level. It can be observedthat the performance degrades quickly after about 3 m. In order to navigate the UAVinto the zone of face detections we employ person following approach describedin the preceding sections. The second limitation for obtaining a good face imageFig. 1: Probability histogram for person detections at various distances from 1 m to6 m . is posed by the quality of the detected face. According to a study [21], the recom-mendation for face recognition in images are 32 pixels minimal distance betweenthe eyes and 64 pixels for a better accuracy. These distances are determined by thequality of images as well as the distance at which they were captured. Althoughthere have been efforts to increase the quality of sensors on UAVs, streaming withlimited bandwidth introduces compression artifacts. The individual video framesand images are also affected by the velocity and movement of the sensor due tomotion artifacts. Face detection algorithms perform reliably until a critical value ofcompression below which they degrade significantly [16]. Further changes in illu-mination and pose variations make the process more challenging. In this paper weaim to minimize the pose variations seen by the UAV. Through continuous assess-ment of facial pose, we adjust the flight path to obtain an optimized view of face fora given image quality once it is above the face detection threshold. ctive Face Frontalization using Commodity Unmanned Aerial Vehicles 5 The overview of our algorithm is described in Figure 2a. The communicationfrom the UAV is established using ROS (Robot Operating System). The detectornodes subscribe to the high resolution onboard video feed. The frequencies of all thecomponents of the architecture is listed in 2b. The pedestrian detector is concate-nated with the tracker to increase the frequency of person detections. The boundingbox obtained from the person detection is used to approximate the distance of thecamera from the person. The PD controller issues the control signals based on theestimated distance to drive the UAV into the zone of face detections. For every de-tection obtained by the face detector node, an assessment on the face quality interms of pose is performed and a frontalisation error is calculated. The error is usedto navigate the UAV to th region of optimized face detections.
1. Parrot Bebop : Out of the several commercially available UAVs, Bebop Parrot isrelatively safe to operate around humans. It has a dual core processor with quad-core GPU, 8GB flash memory, and GPS. Connectivity is via Wi-Fi, and maxoperating distance is 1 mile. Therefore, it is mainly designed for outdoor flight.The on-board camera has 14 mega-pixels with a fish-eye lens. The on-board im-age resolution is 1920 × × − ◦ to + ◦ from the frontal face and radiusup to 5 m was covered. The performance accuracy of face recognition using Fis-cher faces [5] was evaluated for the various poses. Person detection and tracking using vision based cues has received some researchattention in the past. With the focus on making identification more reliable, in thiswork we mainly design a simple approach using off-the-shelf, open-source detec-tors. The pre-trained pedestrian detector in OpenCV [25] is used detect people even
N. Lakshminarayana, Y. Liu, K. Dantu, V. Govindaraju, N. Napp(a) Overview of the active face alignment system. The video feedand the odometry readings are the sensor readings obtained fromBebop. The input to the UAV are the control commands generated bya PD controller.Onboard Images Odometry Pedestrian Detector Tracker Face detector1 .
75 Hz 5 Hz 0 . . . Fig. 2: System Overviewin cluttered environment. The HoG features trained on linear SVM detects peoplein a video stream with very few false positives. However, the detector has rela-tively lower frequency compared to the incoming video stream as listed in Table 2b.Therefore, in order to bridge the gap between two detectors, we use object tracking.Many state-of-the-art trackers have been proposed previously [4, 9, 14, 18]. How-ever, for deformable objects, e.g. humans, it is still a challenging problem. In thiswork, we employ [20] for tracking people. CMT is a key-points based deformablepart model. A set of keypoints are initially selected from the object to be tracked.At every time instant, the keypoints from the previous frames are matched and theconsensus of the points are used select object being tracked. CMT is initialized withbounding box from the detector as shown in Figure 2a . After every detection, CMTis updated with the bounding box. This cascade of the detector and tracker worksefficiently with very few false detections. In order to translate the position of personin image frame to the UAV frame, we make use of the dimensions of the boundingbox. The center of the bounding box is aligned with the center of the image thuskeeping the person in line of sight. Further, to estimate the depth of the camera fromthe person, we make use of the height of the bounding box. In Figure 3, we plot thecorrelation of depth with the height of the detected bounding box using the groundtruth from the AprilTags. The correspondences of the box height with the depth isapproximated to be linear. At any instant of time, the estimation of depth is used toobtain the heading. Thus the UAV can be lead into the zone of face detections. ctive Face Frontalization using Commodity Unmanned Aerial Vehicles 7
Fig. 3: Height of bounding box obtained from the pedestrian detection observed atvarious distances of person from the camera mounted on UAV. The correlation ofthe bounding box height and the distance is approximated to be linear ignoring theoutliers.
The general automated face recognition pipeline involves face detection, face align-ment, feature extraction and feature matching [29]. In a realistic scenario, thechances of detected faces to be frontal are slim. Therefore, the pose variant facesare synthesized to give a frontal view of the face. Tal Hassner et.al. [11] propose aface frontalization technique that uses a fixed reference 3D model. The open sourceface detector from the DLIB library [15] is used for cropping faces from the back-ground. As we focus on images captured from relatively low altitudes, we can safelyclaim that the HoG-based features can distinguish human faces through contrast.The frontalization is carried out by first detecting certain facial features and project-ing the corresponding pixel intensities from 2D space to fixed reference 3D model offace. In order to account for the occlusions and pose variations, the authors use visi-bility scores originally used by 3D reconstruction methods [17], [31]. The visibilityscore for each pixel in the projected 3D surface q is then estimated to be: v ( q ) = − exp ( − N q ) (1)Here N q is the number of times a query pixel q is accessed while forming corre-spondence of the 2D query image pixel with the reference 3D surface. As the faceturns away from the camera increasing the angle, fewer pixels in the query imageare mapped to the 3D pixels, hence reducing the visibility of the region facing away.Figure 4. displays the visibility on surface of faces in three different poses cap-tured during flight. The face detection algorithms renders a tight bounding box N. Lakshminarayana, Y. Liu, K. Dantu, V. Govindaraju, N. Napp
Fig. 4: Heat map showing the visibility at each pixel in different orientations offace.around all the pixels that constitute face. Since the frontalization is preceded bythe person following controller, we make sure that the center of the detected faceis aligned to the center of the UAV frame though visual servoing. The gradient ofvisibility gives a crude estimate regarding the part of the face facing the camera. Wedefine a frontalization error as the difference between the average visibility scoresof right half pixels and average visibility scores of left half pixels of face. Thus thefrontalisation error is :
Frontalisation error = ∑ q = R q v q R q − ∑ q = L q v q Lq (2)Here R q are the right pixels and L q are the left pixels. In order to analyze the effectiveness of the proposed approach in the face recogni-tion pipeline, we perform face verification along the flight path. The faces capturedthroughout the flight path must show an increasing similarity to the frontal face.Facial biometrics are highly expressive and diverse. Convolutional neural networksshow capacity to deliver an intricate representation of images. The deep cascade oflayers give a high level representation of faces. This representation of faces can beused for a one-to-one comparison of face against a registered face. In this paper, weused the Deep CNN architecture VGG-16 proposed by [11] to extract the featuresfrom the detected face.The architecture as depicted in Figure 6 consists of series of convolutional layersand max-pooling layers followed by the fully connected layer and soft-max layer.The convolutional layer extracts features using a sliding window convolving overpatches of the image. The intermediate pooling layers are used to sub sample thefeatures. The last three fully connected layers are similar to the general neural net-work architecture. The network is pre-trained on ImageNet dataset [8] consisting of15 million images of various objects. The features from the third fully connectedlayer is used for representing face. In biometrics, a similarity measure between theregistered image and the query image is used for verification. Higher similarity ofthe query face indicates more confidence of the person being verified. The cosinesimilarity for any two vectors A and B is calculated as follows: ctive Face Frontalization using Commodity Unmanned Aerial Vehicles 9 CS ( A , B ) = A . B || A |||| B || (3)The face verification scheme based on cosine similarity is represented in Figure 5.Face recognition is simply face verification over all the registered users. Howeverin our experiments, we perform face verification for a single user as the multi-userverification requires a large database of registered users.Fig. 5: A generic CNN based verification system consists of face detection,facealignment, feature extraction followed by similarity matchingFig. 6: The figure describes the VGG-16 architecture, a network used to extractfeatures for faces from a high dimensional representation of images. An overview of the system is shown in Figure7. The control architecture consists oftwo distinct parts that both take visual cues from the on-board camera to producemotion commands. When no faces are detected, the first controller looks for peoplein the image and approaches them. Once the UAS is close enough to reliably detecta face, a fontalization controller takes over.
Our person following module navigates the UAV into the region with higher prob-ability of face detection. The first person detection seen by the UAV initiates thetake-off. After the initialization, the aim is to head towards the person keeping the
Fig. 7: The overview of the vision based controllersperson in the center of the frame. The upper segment of the controller illustratesthe person following controller. From the bounding box detected, two features arecalculated, the height of the bounding box and the center of the bounding box. Theheight is used for estimating the depth as explained in Section 3.2. The UAV movesprogressively towards the person with a continuous estimation of depth. As UAVflies directly at the person, we always maintain a minimum distance of 1 . Owing to the small range of operation for the active face frontalization, the inputsfor the face frontalization unit are a set of waypoints through the course of flightpath. The system with a cascade of controllers is illustrated in the Figure 7b. Fromthe detected face, the frontalization error is calculated as explained in Section 1. Asetpoint or the target point for the UAV is set based on the fronatlization error. Basedon the current odometry of the UAV and the desired setpoint, the PD controllerissues velocities about the longitudinal and lateral axis of the UAV. However, thefrequency of the odometry is much higher than the face detections. Therefore, in the ctive Face Frontalization using Commodity Unmanned Aerial Vehicles 11 inner loop, with every estimate of current odometry and the desired setpoint, the PDcontroller issues a new velocity command. However the setpoint is updated as andwhen a face is detected in the outer loop. The outer loop also sets the yaw for theUAV creating a circling movement around the face. The yaw is issued based on thedifference between the center of the face and the center of the image. This way theUAV is oriented towards the detected face.
In order to asses frontalization error as a quantitative measure of the face detection,we perform several experiments. Firstly, we fly the UAV at different angles anddistances from the person. For each of the sampled points, frontalization accuracyis calculated as:
Frontalization accuracy = − Frontalisation error
In Figure 8 (a) and (b) frontalization accuracy as a function of distance and orienta-tion from the face is plot. Since the frontalisation process begins after aligning thefacial bounding box to the center of the UAV frame, experiments of frontalisationerror with respect to the altitude of UAV is not relevant here. During the process offrontalisation, the altitude of UAV with respect to face bounding box is maintainedconstant. (a) (b)(c) (d)
Fig. 8: (a) Frontalization accuracy vs radius for on board camera, (b) Frontalizationaccuracy vs orientation for onboard camera, (c) Average Recognition confidence vsradius from Kinect and (d) Average Recognition confidence vs orientation fromKinect
A visual representation of the velocity vectors is shown Figure . At every locationon the field of experiment we represent the strength and direction of the velocityusing arrows. all the vectors point towards the human positioned at [0,0.5]. Thevelocities are calculated with respect to the relative position of UAV with the persondetection bounding box as described in Section 3.2 We compare our results withthe empirical data obtained by performing a similar analysis on Kinect camera forface recognition. Figure 8 (c) and (d) is a plot of recognition accuracy with respectto orientation and distance observed from the Kinect camera. Figure 8 establishes aclose correspondence of frontalization error with the face recognition confidence.The accuracy is high when the camera is oriented at 0 ◦ from the face and sub-sides down at higher angles leading to a more severe pose of face. A similar resultwas observed for the radius. As explained in Section 3.5, we use a PD controller tonavigate the UAV to the center.Fig. 9: The plot visualizes the velocity commands that are calculated at eachlocation with respect to the detected faces.In Figure 9, the control signals generated using the frontalization error at anygiven point in the space is shown. The arrows converge towards the dead zone ,where the frontalization error is zero or the view of face is fully frontal.Fig. 10: UAV trajectory for a single run. The color map indicate the time lapse. Theperson is standing at the origin looking straightWe test the interface described in Section 3.5 in an indoor setting. A typical pathtraversed during active face frontalization is shown in the Figure 10. The position ofthe person and the orientation is marked by the red circle and the arrow respectively. ctive Face Frontalization using Commodity Unmanned Aerial Vehicles 13(a) Frontalization error during the flightpath. The error saturates at zero when afrontal view is obtained (b) Similarity measure for the facedetected along the proposed flight path. Fig. 11: The trajectories of frontalisation error and cosine similarity as a functionof time stepsThe yaw of the UAV at every instant is represented along the path. By taking offat different angles away from the face, we track the frontalization error along thetrajectory. As shown in Figure 11a, the error starts off at a high value, with positiveerror indicating the right side of face and negative error indicating the left part. Theerror gradually converges to zero. Any drift is efficiently sustained. In the final stageof evaluation, to justify our claim that active frontalization boosts the performance offace recognition, cosine similarity scores for the faces detected along flight path andthe registered face for different runs is calculated. In Figure 11b, it can be observedthat the similarity scores for the images along the trajectory increases along theflight path. The higher similarity yields better face recognition.
The contribution of this paper is twofold. Firstly, we provide an autonomous bio-metric information gathering system based on UAVs. Although literature suggestresearch aiming to incorporate face technology with UAVs. We make the first at-tempt of actively assessing biometric data on UAVs as it is gather and to improvethe quality of the gathered information. We make use of soft biometric cues to im-plement autonomy of motion useful to a particular class of applications like UAVsurveillance and photography. Secondly, we introduce frontalization error derivedfrom the existing methods that can be used online for adjusting flight paths to im-prove the image quality to aid verification. Through experimentation we show thatthe proposed approach leads to a better quality of faces images for recognition. Wepresent a inexpensive consumer system that is modular and can be adapted to anyROS enabled UAV platform. The software components are also off-the-shelf andcould likely be adapted improve the overall system performance.
Acknowledgements
This material is based upon work supported by the National Science Foun-dation under Grant IIP
References
1. Bebop. .2. Hexo+. https://hexoplus.com/ .3. AIRDOG. Airdog. .4. B
ABENKO , B., Y
ANG , M.-H.,
AND B ELONGIE , S. Visual tracking with online multipleinstance learning. In
Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEEConference on (2009), IEEE, pp. 983–990.5. B
ELHUMEUR , P. N., H
ESPANHA , J. A . P., AND K RIEGMAN , D. J. Eigenfaces vs. fisherfaces:Recognition using class specific linear projection.
IEEE Trans. Pattern Anal. Mach. Intell. 19 ,7 (July 1997), 711–720.6. C
HEN , H., M . W ANG , X.,
AND L I , Y. A survey of autonomous control for uav. In (Nov 2009),vol. 2, pp. 267–271.7. D AVIS , N., P
ITTALUGA , F.,
AND P ANETTA , K. Facial recognition using human visual systemalgorithms for robotic and uav platforms. In (April 2013), pp. 1–5.8. D
ENG , J., D
ONG , W., S
OCHER , R., L I , L.-J., L I , K., AND F EI -F EI , L. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09 (2009).9. F
ELZENSZWALB , P. F., G
IRSHICK , R. B., M C A LLESTER , D.,
AND R AMANAN , D. Ob-ject detection with discriminatively trained part-based models.
IEEE transactions on patternanalysis and machine intelligence 32 , 9 (2010), 1627–1645.10. G
ASZCZAK , A., B
RECKON , T. P.,
AND H AN , J. Real-time people and vehicle detection fromuav imagery. vol. 7878, pp. 78780B–78780B–13.11. H ASSNER , T., H
AREL , S., P AZ , E., AND E NBAR , R. Effective face frontalization in uncon-strained images.
CoRR abs/1411.7964 (2014).12. H SU , H.-J., AND C HEN , K.-T. Face recognition on drones: Issues and limitations. In
Pro-ceedings of the First Workshop on Micro Aerial Vehicle Networks, Systems, and Applicationsfor Civilian Use (2015), ACM, pp. 39–44.13. J
AIN , A., H
ONG , L.,
AND P ANKANTI , S. Biometric identification.
Commun. ACM 43 , 2(Feb. 2000), 90–98.14. K
ALAL , Z., M
IKOLAJCZYK , K.,
AND M ATAS , J. Tracking-learning-detection.
IEEE trans-actions on pattern analysis and machine intelligence 34 , 7 (2012), 1409–1422.15. K
ING , D. E. Dlib-ml: A machine learning toolkit.
Journal of Machine Learning Research 10 (2009), 1755–1758.16. K
ORSHUNOV , P.,
AND O OI , W. T. Video quality for face detection, recognition, and tracking. ACM Trans. Multimedia Comput. Commun. Appl. 7 , 3 (Sept. 2011), 14:1–14:21.17. K
UTULAKOS , K. N.,
AND S EITZ , S. M. A theory of shape by space carving. In
Proceedingsof the Seventh IEEE International Conference on Computer Vision (1999), vol. 1, pp. 307–314vol.1.18. M A , C., H UANG , J.-B., Y
ANG , X.,
AND Y ANG , M.-H. Hierarchical convolutional featuresfor visual tracking. In
Proceedings of the IEEE International Conference on Computer Vision (2015), pp. 3074–3082.19. M
EGVII . Face++. .20. N
EBEHAY , G.,
AND P FLUGFELDER , R. Clustering of static-adaptive correspondences fordeformable object tracking. In
Proceedings of the IEEE Conference on Computer Vision andPattern Recognition (2015), pp. 2784–2791.21. N
EUROTECHNOLOGY . Basic recommendation for face recognition. .22. O
LSON , E. AprilTag: A robust and flexible visual fiducial system. In
Proceedings of the IEEEInternational Conference on Robotics and Automation (ICRA) (May 2011), IEEE, pp. 3400–3407.23. O
RBEUS . Rekognition api. https://rekognition.com/ .ctive Face Frontalization using Commodity Unmanned Aerial Vehicles 1524. P
ESTANA , J., S
ANCHEZ -L OPEZ , J. L., S
ARIPALLI , S.,
AND C AMPOY , P. Computer visionbased general object following for gps-denied multirotor unmanned vehicles. In
AmericanControl Conference (ACC), 2014 (2014), IEEE, pp. 1886–1891.25. R
OSEBROCK , A. Pedestrian detection opencv. .26. R
UDOL , P.,
AND D OHERTY , P. Human body detection and geolocalization for uav searchand rescue missions using color and thermal imagery. In (March 2008), pp. 1–8.27. R
UDOL , P.,
AND D OHERTY , P. Human body detection and geolocalization for uav searchand rescue missions using color and thermal imagery. In (March 2008), pp. 1–8.28. SAR. Sar. http://sardrones.org/ .29. S
TAN
Z. L I , A. K. J. Handbook of Face Recognition 2005 . Springer-Verlag New York, 2005.30. W
ALL , T.,
AND M ONAHAN , T. Surveillance and violence from afar: The politics of dronesand liminal security-scapes.
Theoretical Criminology 15 , 3 (2011), 239–254.31. Z
ENG , G., P
ARIS , S., Q
UAN , L.,
AND S ILLION , F. Progressive surface reconstruction fromimages using a local prior. In