Ridge Body Parts Features for Human Pose Estimation ... - IEEE Xplore

1 downloads 0 Views 1MB Size Report
proposed for human pose recognition utilizing ridge body parts features. Initially, depth silhouettes extract ridge data inside the binary edges and initialize each ...

IEEE - 33044

Ridge Body Parts Features for Human Pose Estimation and Recognition from RGB-D Video Data Ahmad Jalal

Yeonho Kim

Daijin Kim

Department of Computer Science and Engineering POSTECH Gyengbuk, Republic of Korea Email: [email protected]

Department of Computer Science and Engineering POSTECH Gyengbuk, Republic of Korea Email: [email protected]

Department of Computer Science and Engineering POSTECH Gyengbuk, Republic of Korea Email: [email protected]

Abstract—This paper addresses the issues of 3D human pose estimation, tracking and recognition from RGB-D video sequences using a generative structured framework. Most existing approaches focus on these issues using discriminative models. However, a discriminative model has certain drawbacks: a) it requires expensive training steps and large amount of training samples for covering inherently wide pose space, and (b) not suitable for real-time applications due to its slow algorithmic inferences. In this work, a real-time tracking system has been proposed for human pose recognition utilizing ridge body parts features. Initially, depth silhouettes extract ridge data inside the binary edges and initialize each body joints information using predefined pose. Then, body parts tracking incorporates appearance learning to handle occlusions and manage body joints features. Lastly, Support Vector Machine is used to recognize different poses. Experimental results demonstrate that the proposed method is reliable and efficient in recognizing human poses at different realistic scenes. Keywords—Human pose estimation, RGB-D image, Connected Component Labeling (CCL), Human Pose Recognition (HPR), Support Vector Machine (SVM)

I.

I NTRODUCTION

Human pose tracking and recognition is important for several potential applications including video games, markerless motion capture, understanding human action, security and health-care systems [1]–[7]. For example, in person monitoring application, human poses are analyzed to monitor person’s activity, its interaction with people and environment to prevent uncertain events in the monitored scenes. Analysis of human poses involve estimating the pose parameters (i.e., position, angle and length) of the human body components such as head, torso and limbs that fit to the input image. At present, human pose tracking and recognition systems are mainly deal with wearable sensors (i.e., accelerometers, 3D motion sensor and gyroscopes) and vision sensors (i.e., video cameras) [8], [9]. Using wearable sensors for pose estimation, Schwarz et al. [8] proposed Gaussian process regression to learn the person-specific functional relationship between sensors measurement and full-body poses. However, sensors consumed high electric power and quite insecure to the subject by keeping the sensors on their body for long time. In [10], Lin and Kulic described a kinematic-based approach to estimate human poses from wearable sensors. However, these sensors are quite confuse in case of similar patterns

of different poses and it is difficult to estimate data in case of fast human motion. While, using vision sensors for pose estimation, Shotton et al. [11] proposed human body pose estimation which segment different human body parts using a random forest classifier. Due to the above observations, wearable sensors have certain difficulties which make them not practical in real world applications. Thus, our work is focused on utilizing depth vision sensors which extract features from human poses. This work presents the ridge body parts features (RBPF) based on human pose recognition (HPR) system which utilizes the depth imaging data. Firstly, depth maps are utilized to extract human depth silhouettes from noisy background. These silhouettes are used to extract ridge data based on binary edge extraction acting as skeleton shape of human body. Secondly, pose estimation is applied to initialize each body part using Y-shape pose and position of each body part is tracked by considering the orientation of known parameters, search area localization and forward kinematics mechanism. Then, body joints information along with RBPF are used for training of the HPR engine. Finally, after training, the system recognizes learned poses via the trained Support Vector Machine (SVM) for HPR. This paper is organized as follows. Section II presents the system architecture of proposed HPR. The subsequent section defines human silhouettes detection with subject’s identification and used for feature extraction as ridge data. Then, the detailed design process of pose initialization, body parts tracking and training/recognition using SVM is considered. In section III, the experimental results and analysis are provided to demonstrate the effectiveness of the proposed approach over existing approach. Finally, conclusions are presented in section IV. II.

P ROPOSED HPR M ETHODOLOGY

The proposed HPR system consists of preprocessing of depth images to extract human silhouettes in video, feature extraction via RBPF and initialization which further proceed to body parts tracking. Then, these features used vector quantization algorithm followed by modelling, training and recognition using SVM. Fig. 1 shows the overall architecture of the proposed HPR system.

5th ICCCNT - 2014 July 11 - 13, 2014, Hefei, China

IEEE - 33044

color) as shown in Fig. 2(c). Finally, all segmented output silhouettes are assigned as global label ID (Skt ) which indicate a differentiation among different subjects as Fig. 2(d).     ID Skt = argmin d Sit−1 − Skt (1) i∈ID (Skt−1 )

Manually Labeled Joints Feature Extraction

Ridge-Body Joint Feature

Train SVM

Depth Data Training Testing

Depth Data

Body Parts Tracking

Ridge-Body Joint Feature

Recognized Human Pose

Trained SVM

Figure 1: System architecture of the proposed HPR system

D21 D22 D23 · · · . .. D31 D32 D33 · · · ...

D11 D12 D13 · · · . ..

.. .

.. . · · · D2n

· · · D1n

(b)

ID5

ID3

(c)

ID2 ID4

ID1

B. Ridge Body Parts Features Ridge body parts features (RBPF) consist of binary edge extraction and ridge data for feature extraction. In binary edge extraction, we extract features from binary edges using depth silhouettes. These edges are processed by distance transform to produce distance maps. While, in ridge data generation, these maps are computed to find local maximal which produce one or more ridge data inside binary edges [14], [15].

n · · · D3

.. .

(a)

where Skt is the current tracked silhouettes compared with the remaining tracked silhouettes to assign a unique identification.

1) Binary Edge Extraction: Binary edge extraction BEdge (I) consists of binary edge information around depth silhouettes using window searching mechanism to measure local statistical values of the intensities of their nearest neighbors. It produces proper edge connectivity and enclosed body structure. BEdge (I) = {Xc ∈ I |∃Xi , |d (Xi ) − d (Xc )| > δE } , Xi ∈ {Xc−1 , Xc+1 , Xc−w , Xc+w , }

(2) where Xc is the center depth pixel compared with the respective neighboring pixels Xi to evaluate intensity value. In addition, binary edges are further processed by the distance transform which provide distance maps.

(d)

Figure 2: Human detection process (a) Original scene contains different objects with noisy background, (b) 3D-CCL identified different objects, (c) segmented objects, and (d) multiple human silhouettes are detected with different subject’s identification

A. Pose Recognition Video Preprocessing From every frame of depth video clip, we extract different depth value objects from an unrestricted environment having noisy background using background subtraction routine [12], [13]. These depth objects from the scenes are labeled their candidates pixels using 3D connected component labeling (3DCCL) method. In 3D-CCL, the variation of pixel intensity in an image is observed using raster scanning. In addition, the chromatic variation and intensity values of background are quite less, thus, we remove all the non-object components acting as background by entropy analysis. While, in case of monitoring different connected components, every depth pixel of the connected component has depth values which help to divide the 3D features based data into two classes (i.e., connected components representing as human or materials and background) as shown in Fig. 2(b). While, the disparity segmentation are employed to find the target human silhouette candidates (green and yellow in color) and separated them from non-silhouette objects (red in

2) Ridge Data Generation: In ridge data generation, these distance maps are used to calculate the local maximal of respective edges and produce a chain of the pixels as ridge data [16]. Such ridge data RData (I) are enclosed by the binary edges acting as skeleton shape [17] of a human body.     N D (X )  i=1 T i (3) < δR RData (I) = Xc ∈ I   N · DT (Xc ) where DT is the distance map values which evaluate center points value with their neighboring pixel values. Fig. 3 shows the diagrammatic representation of the binary edge silhouettes and ridge data based on distance maps. Such ridge data could represent the location of skeleton and eliminate the noisy data around the edge data. C. Pose Estimation and Tracking In this section, the proposed pose estimation and tracking method that satisfies the following two fundamental principles: a) pose initialization is used to configure the human body model, and b) body parts tracking is used to recognize final pose in each frame. 1) Pose Initialization: During initialization, the initial pose is estimated by considering Y-shape pose with the arms extended sideways for human body configuration [18]. However, different body constraints are used for initialization as shown in Fig. 4(a).

5th ICCCNT - 2014 July 11 - 13, 2014, Hefei, China

IEEE - 33044 wT T PH rH ˜t X RD

T PT

θLS θRS

˜t X H

lH,T

lLS,RS

˜t X R RS

˜t X LS L

t ˜ RE X

hT ˜t X LE

(a)

BMT

(b)

t ˜ LD X

Figure 3: Feature extraction using ridge body parts features (a) binary edge extraction and (b) ridge data generation based on distance maps

t ˜ RK X

lL ˜t X LK

˜t X RF

˜t X LF

In body constraints, both torso and head constraints are foremost process during initialization. For torso initialization, the torso is detected based on rectangular box generation, where the box is generated using line scanning process to position the box using 4 degrees of freedom (x, y, height hs , width ws ). To find torso height hT and torso width wT , the box size is calculated as hT = c T · w T ,

w T = rT − l T + 1

(a)

(b)

Figure 4: Diagramatic representation of (a) pose initialization and (b) body parts tracking

data or depth values) around the head position as (4)

where rT , lT are the right, left torso width extremes values during column-wise scanning and cT is the aspect ratio of torso size respectively [19]. For head detection, the scanned data is restricted in-between the torso width extremes values. While, torso top tT are used to generate the radius of the head rH . Both these constraints are used to identify the position of shoulders and hips joint points [20] by considering the aspect ratio information of both hT and wT . However, to find frontal face during head tracking, face detector can be used to detect face [21]–[23]. While, to detect limbs constraints for initialization, both arms and legs body parts are used. For arm detection, Hough line detection [24] is used to estimate the arm positions (i.e., left/right) by representative lines produced from the ridge data. While, junction point between two lines acting as elbow joint points. For leg detection, both (left/right) legs are estimated by considering the distance from the hips joint points to the floor center. 2) Body Parts Tracking: During body parts tracking, we tracked body parts which provide the location of head, torso and limbs which help to recognize final pose in each frame. In addition, we analyze the human position and the body pose direction from the detected body parts information. Body part tracking is mainly focused on tracking rigid body motion from previous frame to current frame [25] with respect to body parts location as shown in Fig. 4(b). Initially, to detect t , we localize the search area for head by head portion XH considering the orientation of known parameters as wT and  t ← X t−1 + lH,T along with pre-defined values of rH and X T T t ΔXT respectively. These parameter values are used to isolate the head portion which locate the feature values (i.e., ridge

t XH

=

⎧ |RH (I)|

⎪ ⎪ 1 ⎪ Xn ⎪ ⎨ |RH (I)|

, if |RH (I)| > 0

n=1 |DH (I)|



⎪ ⎪ 1 ⎪ Xn ⎩ |D (I)| H

(5) , otherwise

n=1

To detect torso portion, the torso portion is surrounded by five joint points as left/right shoulder, torso center and left/right hip joint points. Thus, to track the exact position of torso center joint point XTt , we utilize the length lH,T and rotation matrix RST (θS ) in between the torso center and both shoulder joint points as XTt

St| | t XS − XH 1  =  t  lH,T RST (θS )  t ,  S XS − X H S∈S t S t ∈ {LS, RS} .

(6)

Finally, to detect left/right shoulder and hips joint points are detected around the search area of torso portion as ⎛ ⎜ ⎜ ⎜ ⎝

t XLS t XRS t XLH t XRH





⎜ ⎟ ⎜ ⎟ ⎟ = RTt (θTt ) · ⎜ ⎝ ⎠

t−1 t−1 XLS − XH t−1 t−1 XRS − XH t−1 t−1 XLH − XH t−1 t−1 XRH − XH

⎞ ⎟ ⎟ t . ⎟ + XH ⎠

(7)

Finally, limbs tracking contains a chain of arm joint points and a chain of legs joint points which used the characteristics of forward kinematics mechanism. Thus, the joint space consists of all possible values of the tracked body joints information. Fig. 5 represents the overall process of body parts initialization and body parts tracking.

5th ICCCNT - 2014 July 11 - 13, 2014, Hefei, China

IEEE - 33044

Feature Representation Ridge Data

Initialized?

No

Depth Data

Torso Initializatoin Torso Width/Height

Yes Head Detection

Yes Head Detected?

Head Initialization

Limb Initialization

Head Radius

Arm Length

Shoulder Parameters

Leg Length

Torso Tracking

Yes Limb Detected?

Limb Detection

No

No

Head Tracking

Limb Tracking

Final Pose

Figure 5: Flow diagram of pose initialization and tracking

D. Dimension Reduction via PCA After pose estimation and tracking, the RBPF space become a combination of ridge data and joint points information. Since each depth silhouette has a different number of ridge data, we subsampled the ridge data to make a same-sized features as 200 points of ridge data. However, the role of PCA here is to approximate the higher features data into lower dimensional features [26], [27]. Thus, the principal components (PCs) of the 200 × 3 dimensional subsampled i · Ve , where Pi is the ridge data are expressed as Pi = X PCA projection on the RBPF, Ve is the top eigenvectors and i is the zero mean vector of RBPF. Thus, after applying X PCA on the poses, 3 PC were taken from the ridge data along with joint information and projecting each silhouette vector of RBPF with the size of 1 × 48 (1 × 3 for PCs of ridge data and 1 × 45 for 15 joints). For feature discrimination, we used fisher discriminant analysis. E. Pose Training and Recognition using SVM SVM is state-of-the-art large margin classifier which performs classification using linear decision hyperplanes in the feature space [28]. It is based on finding the optimal separating hyperplane that maximizes the margin of the training data. During training, the hyperplanes are calculated to separate the training data with different labels. While, separating hyperplanes for linear classification can be calculated as y = sgn (w · x + b)

(8)

To maximize margin separating hyperplane within the classes, the condition for classification without training error is y (w · x + b) ≥ 1. Also, we can minimize the complexity 2 term during learning method as min 12 w , which formulate w,b

the optimal hyperplane using Lagrange multipliers. III.

E XPERIMENTAL R ESULTS

In this study, we built our own depth silhouettes database based on nine different poses. We collected 5,822 depth silhouettes from six different subjects for recognition and performed 10-fold cross validation. In the recognition process, each depth silhouette with its size of 320 × 240 (i.e., 1 × 76, 800 acting

(a)

(b)

(c)

Figure 6: Experimental results of our database (a) RGB images along with their corresponding skeleton of different poses, (b) depth silhouettes with background subtraction, and (c) skeleton informtion rendered from various viewpoints

as raw feature dimension) are processed to produce RBPF. These features are combined effort of both ridge data and joint points information, yielding a feature vector of 1 × 48 apply to be used in Linear SVM for recognition. Fig. 6 shows the experimental results as (a) the skeleton results along with RGB images having noisy background, (b) depth images using background subtraction, and (c) the tracked body parts positions are represented in 3D coordinate system. We compare the performance of the proposed features with the conventional features [29] (i.e., normalized color histogram features). Fig. 7(a) shows the 2D plot representation of conventional features using depth silhouettes of all poses. In this plot, the features donot seem to be well separable. While, Fig. 7(b) provides 2D representation of the depth silhouettes where RBPF from all nine different poses are well separated. Finally, we evaluate the recognition results of all nine pose classes using both proposed method along with the conventional method. As shown in Table 1, the mean recognition results of normalized color histogram features are 65.46% using depth poses silhouettes. While, in Table 2, the recognition results of our proposed ridge body parts features showed high recognition rate 91.19% over the conventional method. IV.

C ONCLUSIONS

In this work, we have presented a novel ridge body parts features based on HPR system using depth videos.

5th ICCCNT - 2014 July 11 - 13, 2014, Hefei, China

IEEE - 33044

! !

 ! !! ! !!!  ! #  # 

 ! #





 $   !"  "# 









ï

ï

ï

Table 1: Recognition results of normalized color histogram features using depth silhouettes (RA: Raise Arms, BT: Bend Torso, CA: Cross Arms, CL: Cross Legs, SA: Shake Arms, TH: Touch Hip, TD: Touch Head, HA: Hide Arms, CR: Crouch) Poses

RA

BT

CA

CL

SA

TH

TD

HA

CR

RA

0.71

0.02

0.01

0.00

0.05

0.00

0.00

0.00

0.22

BT

0.01

0.79

0.01

0.00

0.00

0.00

0.00

0.00

0.20

CA

0.04

0.04

0.48

0.00

0.01

0.00

0.00

0.00

0.42

CL

0.08

0.05

0.06

0.59

0.03

0.00

0.00

0.00

0.19

SA

0.05

0.01

0.01

0.01

0.65

0.00

0.00

0.00

0.27

TH

0.04

0.00

0.00

0.00

0.08

0.15

0.00

0.04

0.69

TD

0.10

0.00

0.00

0.00

0.02

0.00

0.51

0.02

0.36

HA

0.00

0.00

0.03

0.00

0.06

0.03

0.06

0.45

0.36

CR

0.01

0.07

0.00

0.00

0.02

0.00

0.00

0.00

0.90

Mean Recognition Rate(%) = 65.46 ï

ï

ï

ï ï     $   !"  "# 







Table 2: Recognition results of ridge body parts features using depth silhouettes

(a)

Poses

RA

BT

CA

CL

SA

TH

TD

HA

CR



RA

0.97

0.01

0.01

0.00

0.01

0.00

0.00

0.00

0.01

BT

0.01

0.99

0.00

0.00

0.00

0.00

0.00

0.00

0.00



CA

0.01

0.01

0.91

0.00

0.00

0.00

0.00

0.00

0.06

CL

0.02

0.02

0.02

0.95

0.00

0.00

0.00

0.00

0.00

SA

0.19

0.01

0.04

0.00

0.76

0.00

0.00

0.00

0.00

TH

0.00

0.00

0.00

0.00

0.00

1.00

0.00

0.00

0.00

TD

0.02

0.00

0.00

0.00

0.00

0.07

0.92

0.00

0.00

HA

0.00

0.00

0.03

0.00

0.00

0.00

0.09

0.88

0.00

CR

0.03

0.02

0.00

0.01

0.01

0.00

0.00

0.02

0.90

 %!#"#$!





ï

ï

Mean Recognition Rate(%) = 91.19 " !" !"

!"" !"

!"""  !" $ $  !"

!$

ï

ï

ï ï

ï

ï

ï

ï    %!#"#$!









(b)

Figure 7: 2D plots representation of (a) conventional features (i.e., normalized color histogram features) and (b) proposed features (i.e., ridge body parts features) using depth silhouettes

ACKNOWLEDGMENT Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(NRF-2010-0019523) This research was supported by the MSIP(Ministry of Science, ICT and Future Planning), Korea, under the IT Consilience Creative Program (NIPA-2014-H0201-14-1001) supervised by the NIPA(National IT Industry Promotion Agency) R EFERENCES [1]

Our proposed HPR system utilizes a combination of body pose estimation and tracking using ridge body parts features from the joints points of the skeleton model and modeling, training and pose recognition using SVM. Experimental results showed some promising performance of the proposed HPR technique, achieving the mean recognition rate of 91.19% over the conventional method as 65.46%. Also, it handles selfocclusion, overlapping among people, and hidden body parts prediction which significantly track complex poses and improve recognition rate. We believed that the proposed system is useful for many applications including automatic surveillance, 3D games and human movement analysis.

[2]

[3]

[4]

[5]

A. Adam, E. Rivlin, I. Shimshoni, and D. Reinitz, “Robust real-time unusual event detection using multiple fixed-location monitors,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 30, no. 3, pp. 555–560, 2008. L. Pishchulin, A. Jain, M. Andriluka, T. Thormahlen, and B. Schiele, “Articulated people detection and pose estimation: Reshaping the future,” in Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pp. 3178–3185. IEEE, 2012. A. Jalal, J. T. Kim, and T.-S. Kim, “Human activity recognition using the labeled depth body parts information of depth silhouettes,” in Procedding of the 6th international symposium on Sustainable Healthy Buildings, pp. 1–8, 2012. A. Jalal and I. Uddin, “Security architecture for third generation (3g) using gmhs cellular network,” in Emerging Technologies, 2007. ICET 2007. International Conference on, pp. 74–79. IEEE, 2007. A. Jalal and M. A. Zeb, “Security enhancement for e-learning portal,” International Journal of Computer Science and Network Security, vol. 8, no. 3, pp. 41–45, 2008.

5th ICCCNT - 2014 July 11 - 13, 2014, Hefei, China

IEEE - 33044

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24] [25]

[26]

H. Lu, X. Shao, and Y. Xiao, “Pose estimation with segmentation consistency,” Image Processing, IEEE Transactions on, vol. 22, no. 10, pp. 4040–4048, 2013. A. Jalal and M. A. Zeb, “Security and qos optimization for distributed real time environment,” in Computer and Information Technology, 2007. CIT 2007. 7th IEEE International Conference on, pp. 369–374. IEEE, 2007. L. A. Schwarz, D. Mateus, and N. Navab, “Discriminative human fullbody pose estimation from wearable inertial sensor data,” in Modelling the Physiological Human, pp. 159–172. Springer, 2009. A. Jalal, J. T. Kim, and T.-S. Kim, “Development of a life logging system via depth imaging-based human activity recognition for smart homes,” in Sustainable Healthy Buildings, International Symposium on, pp. 91–95, 2012. J. F. Lin and D. Kuli´c, “Human pose recovery using wireless inertial measurement units,” Physiological measurement, vol. 33, no. 12, p. 2099, 2012. J. Shotton, T. Sharp, A. Kipman, A. Fitzgibbon, M. Finocchio, A. Blake, M. Cook, and R. Moore, “Real-time human pose recognition in parts from single depth images,” Communications of the ACM, vol. 56, no. 1, pp. 116–124, 2013. M. Piccardi, “Background subtraction techniques: a review,” in Systems, man and cybernetics, 2004 IEEE international conference on, vol. 4, pp. 3099–3104. IEEE, 2004. A. Jalal, M. Z. Uddin, and T.-S. Kim, “Depth video-based human activity recognition system using translation and scaling invariant features for life logging at smart home,” Consumer Electronics, IEEE Transactions on, vol. 58, no. 3, pp. 863–871, 2012. A. Jalal and S. Kim, “Advanced performance achievement using multialgorithmic approach of video transcoder for low bit rate wireless communication,” ICGST International Journal on Graphics, Vision and Image Processing, vol. 5, no. 9, 2005. Z. Zhang, H. S. Seah, C. K. Quah, and J. Sun, “Gpu-accelerated realtime tracking of full-body motion with multi-layer search,” Multimedia, IEEE Transactions on, vol. 15, no. 1, pp. 106–119, 2013. S. Malassiotis and M. G. Strintzis, “Robust real-time 3d head pose estimation from range data,” Pattern Recognition, vol. 38, no. 8, pp. 1153–1165, 2005. A. Jalal, S. Lee, J. T. Kim, and T.-S. Kim, “Human activity recognition via the features of labeled depth body parts,” in Impact Analysis of Solutions for Chronic Disease Prevention and Management, pp. 246– 249. Springer, 2012. G. R. Bradski and J. W. Davis, “Motion segmentation and pose recognition with motion history gradients,” Machine Vision and Applications, vol. 13, no. 3, pp. 174–184, 2002. A. Jalal, N. Sarif, J. T. Kim, and T.-S. Kim, “Human activity recognition via recognized body parts of human depth silhouettes for residents monitoring services at smart home,” Indoor and Built Environment, vol. 22, no. 1, pp. 271–279, 2013. M. Dantone, J. Gall, C. Leistner, and L. Van Gool, “Human pose estimation using body parts dependent joint regressors,” in Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on, pp. 3041–3048. IEEE, 2013. A. Jalal and S. Kim, “Global security using human face understanding under vision ubiquitous architecture system.” Enformatika, vol. 13, 2006. J.-G. Wang, E. T. Lim, and R. Venkateswarlu, “Stereo head/face detection and tracking,” in Image Processing, 2004. ICIP’04. 2004 International Conference on, vol. 1, pp. 605–608. IEEE, 2004. A. Jalal and A. Shahzad, “Multiple facial feature detection using vertexmodeling structure,” in Interactive computer aided learning, 2007. ICL’07. IEEE Computer Society Conference on. IEEE, 2007. R. Okoda, “Discriminative generalized hough transform for object detection,” in ICCV, pp. 2000–2005, 2009. A. Jalal and Y. A. Rasheed, “Collaboration achievement along with performance maintenance in video streaming,” in Interactive Computer Aided Learning, IEEE Conference on, 2007. A. Jalal, M. Z. Uddin, J. T. Kim, and T.-S. Kim, “Recognition of human home activities via depth silhouettes and r transformation for smart

homes,” Indoor and Built Environment, vol. 21, no. 1, pp. 184–190, 2012. [27] A. Jalal, S. Kim, and B. Yun, “Assembled algorithm in the real-time h. 263 codec for advanced performance,” in Enterprise networking and Computing in Healthcare Industry, 2005. HEALTHCOM 2005. Proceedings of 7th International Workshop on, pp. 295–298. IEEE, 2005. [28] Y. Li, S. Gong, J. Sherrah, and H. Liddell, “Support vector machine based multi-view face detection and recognition,” Image and Vision Computing, vol. 22, no. 5, pp. 413–427, 2004. [29] C. Nakajima, M. Pontil, B. Heisele, and T. Poggio, “Full-body person recognition system,” Pattern recognition, vol. 36, no. 9, pp. 1997–2006, 2003.

5th ICCCNT - 2014 July 11 - 13, 2014, Hefei, China

Suggest Documents