Modeling Human Motion for Predicting Usage of Hospital Operating

Room

Ilyes Sghir and Shishir Shah

Department of Computer Science, University of Houston, 4800 Calhoun Road, Houston, Texas, U.S.A.

Keywords:

Hospital Operating Room Analysis, Human Motion Modeling.

Abstract:

In this paper, we present a system that exploits existing video streams from a hospital operating room (OR) to

infer OR usage states. We deﬁne OR states that are relevant for assessing OR usage efﬁciency. We adopt a

holistic approach that involves the combination of two meaningful human motion features: gestures or upper

body movements computed using optical ﬂow and whole body movements computed through motion trajecto-

ries. The two features are independently modeled for each of the deﬁned OR usage states and eventually fused

to obtain a ﬁnal decision. Our approach is tested on a large collection of videos and the results show that the

combination of both human motion features provide signiﬁcant discriminative power in understanding usage

of an OR.

1 INTRODUCTION

The Operating Room (OR) is by far the most com-

plex and expensive environment within any hospital.

With the advent of technology and the increase in the

number of minimally invasive surgeries, ORs have

become high costs / high revenues assets. Nonethe-

less, their effective utilization hasn’t been fully real-

ized. Although no published formal data assessing

their performance can be found, it was estimated in

2003 that ORs generated almost half of a hospital’s

revenues while running at only 68% of their capac-

ity (Association et al., 2003). Assessing workﬂow

performance would signiﬁcantly improve quality of

healthcare delivery and increase ﬁnancial outcomes

for a hospital.

Unplanned events, inefﬁcient supply chain man-

agement, but most importantly, lack of operational

discipline can highly affect OR performance. In fact,

start-time delays (Ciechanowicz and Wilson, 2011;

Does et al., 2009; Schuster et al., 2013), as well as,

unregulated turnover time (Kodali et al., 2014; As-

sociation et al., 2003) have been identiﬁed as major

causes of OR inefﬁciency. Does et al. (Does et al.,

2009) focused on the start-time delay of the ﬁrst op-

eration of the day and harvested 4-weeks of data from

13 hospitals in Belgium and the Netherlands. By

deﬁning the start-time as the time of the ﬁrst inci-

sion, they concluded that delays range from 25 mins

to 103 mins (Does et al., 2009). Turnover time or

the time-lapse between 2 different surgeries lasts 30

mins on average, while in best practice it should last

only 15 mins (Association et al., 2003). Macario es-

timated in 2010 that, in US hospitals, a running OR

costs about $20/min in material supplies while gener-

ating on average $60/min in revenue (Macario, 2010).

If we approximate start-time morning delays to be 60

mins and the cost of an OR to be $2000/hour, then

a hospital with 10 ORs running 250 days a year, can

potentially save 5 million dollars each year.

According to Ciechanowicz and Wil-

son (Ciechanowicz and Wilson, 2011), regular

local audit of OR usage is important to optimize the

clinical processes within the OR and the perioperative

environment. Nonetheless, studies performed until

now have been primarily based on manual data

acquisition by nurses. Daily and automated infor-

mation about OR efﬁciency would be of high value

at the administrative level for continuous quality

improvement. In this paper, we present a system

that exploits existing video streams from a hospital

operating room (OR) to infer OR usage states. We

deﬁne OR states that are relevant for assessing OR

usage efﬁciency. We adopt a holistic approach that

involves the combination of two meaningful human

motion features: gestures or upper body movements

computed using optical ﬂow and whole body move-

ments computed through motion trajectories. The

two features are independently modeled for each of

the deﬁned OR usage states and eventually fused to

328

Sghir, I. and Shah, S.

Modeling Human Motion for Predicting Usage of Hospital Operating Room.

DOI: 10.5220/0005678003280335

In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 328-335

ISBN: 978-989-758-175-5

 2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reser ved

obtain a ﬁnal decision. Our approach is tested on a

large collection of videos and the results show that the

combination of both human motion features provide

signiﬁcant discriminative power in understanding

usage of an OR.

2 RELATED WORK

The more general problem of workﬂow monitor-

ing is already being addressed in more constrained

industrial environments such as car manufactur-

ing (Voulodimos et al., 2011; Veres et al., 2011;

Arbab-Zavar et al., 2014). In 2014, Arbab-Zavar et

al. (Arbab-Zavar et al., 2014) exploited shape and mo-

tion features extracted from an overhead video in or-

der to identify highly structured tasks and activities

within a car manufacturing plant. A Markov tempo-

ral structure based decision system has been proposed

in (Behera et al., 2014) to model spatio-temporal re-

lationships during object manipulations tasks and has

been tested for continuous activity recognition in as-

sembling a pump system. Yet, in ORs, dozens of

tasks are carried out by many different people and

cannot be deﬁned as easily as in strictly designed in-

dustrial environments. Various solutions have been

proposed in the literature for enhancing OR through-

put by facilitating its management. In 2007, one of

the systems used at the MIT General Hospital was

the OR-Dashboard, which is a solution offered by a

company called LiveData (the NYP Wall of Knowl-

edge and manager, ). OR-Dashboard displays infor-

mation about the patient and the surgical procedure.

Other commercial solutions can be found such as OR-

BIT (Lange et al., 2010) or AwareMedia (Bardram

et al., 2006). More recently, in 2011, Niu et al. (Niu

et al., 2011) proposed a simulation model for perfor-

mance analysis of the OR. Unfortunately, all these so-

lutions rely on human intervention and manual data

entry.

To address this inconvenience, alternative ap-

proaches consist of leveraging electronic signals

present in the OR in order to identify automatically

its usage state without human intervention. In 2005,

Xiao et al. (Xiao et al., 2005) proposed to use patient’s

vital signs in order to monitor when the subject is in

the OR. Later on, in 2007, Bhatia et al. (Bhatia et al.,

2007) designed a system analyzing video streams to

automatically recognizing the OR state using Ma-

chine learning algorithms (SVM and HMM). In 2009,

Padoy et al. (Padoy et al., 2009) exploited a multiple-

camera system for extracting low level 3D motion fea-

tures that are ultimately fed into a workﬂow-HMM.

In 2010, Lange et al. (Lange et al., 2010) proposed

a phase recognition system using sensor technology.

In 2011, Nara et al. (Nara et al., 2011) introduced an

ultrasonic location aware system that tracks contin-

uously the 3D position of the surgical staff in order

to recognize discriminant human motion patterns. Fi-

nally, in 2013, Lea et al. (Lea et al., 2013) recorded

depth videos in Intensive Care Units (ICUs) using an

Xbox Kinect in order to identify tasks such as docu-

menting, checking-up on a patient, and performing a

procedure.

Unlike previous solutions, we do not introduce ad-

ditional sensors but, instead, we aim at exploiting ex-

isting cameras that are placed in modern ORs to fa-

cilitate observations and training of other physicians

and residents. The computation of 3D velocity val-

ues as suggested by Padoy et al. (Padoy et al., 2009)

would require an additional camera or a 3D sensor,

which is not commonly available in an OR. The effec-

tive utilization of video streams within the OR hasn’t

been fully realized. In fact, we opt to use motion

cues that can be computed from video obtained us-

ing a single camera. Unlike Bhatia et al. (Bhatia

et al., 2007), we do not deﬁne our OR states based

on the presence of objects in the scene (second bed,

drape on and off, etc.). We exploit physically mean-

ingful features capturing discriminant human motion

patterns. Instead of using a large ultrasonic location-

aware system like Nara et al. (Nara et al., 2011), we

take advantage of a detection algorithm based on a

discriminatively trained part-based upper-body model

developed using Felzenswalb et al.’s object detection

framework (Felzenszwalb et al., 2010b; Felzenszwalb

et al., 2010a).

3 PROPOSED APPROACH

3.1 OR Usage-state Model

Typically, when a patient is brought to an OR,

an anesthesiologist starts administrating anesthesia.

Once the patient is ready, surgeons proceed to make

the ﬁrst incision (Schuster et al., 2013). At the

end of the surgical procedure, all the instruments are

wrapped up, the surgical staff proceeds to clean up,

and the patient is transferred to the recovery room.

In this paper, we propose a three-stage usage-state

transition model. Human motion patterns vary across

these states within the OR. This simple observation is

the motivation for the states in our model as shown in

ﬁgure 1.

In addition, recognizing these states can provide

additional information about the usage efﬁciency of

the OR. Time taken in each state can provide a holis-

Modeling Human Motion for Predicting Usage of Hospital Operating Room

329

Figure 1: OR usage state transition model.

tic understanding of how the OR is utilized. Time

taken for surgeries can be compared to understand

variations is procedures. Similarly, additional usage

metrics can be derived from these recognized states

and their time data. Hence, the three-stage usage-state

model proposed is a good starting point to understand

OR usage.

3.2 Overview

The following is an overview of the proposed system.

Given a camera in an OR, we aim to model each of the

three states based on features computed from a single

video camera. We propose to use two features: the

upper body motion feature and the motion trajectory

feature. The ﬁrst intends to capture small movements

or gestures perform while standing at a position in the

OR, while the second is intended to capture exhibited

walking motions in the OR. An overview of the fea-

ture computation and the estimation of a model for

each of the states is shown in Figure 2) and the use of

the models for deciding on the OR state given a new

video input is shown in Figure 3. In the following, we

provide further details about each of the modules of

the proposed approach.

Figure 2: Feature computation and model estimation.

Figure 3: Inference using proposed upper and lower body

features.

3.3 Motion Trajectory Feature

In OR videos, feet and faces are often occluded de-

pending on one’s position and orientation. Consider-

ing an upper-body detector instead of a face detector

or a human-body detector is therefore extremely rel-

evant. Obviously, image parts or upper-body features

such as gloves, masks or head protections are speciﬁc

to OR environments. Using pre-trained human de-

tectors for such an environment tend to be erroneous

and result in large number of false or missed detec-

tions (ﬁgure 4(a)). Therefore, we trained our own

upper-body model based on Felzenswalb et al.’s part-

based detector (Felzenszwalb et al., 2010b; Felzen-

szwalb et al., 2010a). Training was done over a man-

ually deﬁned set consisting of 400 negative samples

and 800 positive samples extracted from OR videos in

each state, to obtain improved detections (ﬁgure 4(b)).

Having obtained the detected bounding boxes in

each frame, camera calibration is used to estimate

their position on the ground plane.

(a)

(b)

Figure 4: Upper-body detections using (a) a generic detec-

tor, and (b) our OR-trained model.

The method we use is based on Criminisi et al.

and Hoiem et al.’s (Criminisi et al., 2000; Hoiem

et al., 2008) work on Single View Metrology. The

reference plane is considered to be the ground, and

the parallel plane, the one that would contain the top

of the head of each person.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

330

Reference plane calibration consists of computing

a homography matrix H

re f

that allows us to project

the ground points from the image plane onto the

ground plane. For that, we carefully select an im-

age of an empty OR from our dataset. Selecting an

image that offers as many lines on the ﬂoor as pos-

sible is best. A minimum of 4 corresponding points

are needed to estimate 8 unknowns from the homog-

raphy matrix. We identiﬁed 6 corresponding points,

as shown in ﬁgure 5(a), and ﬁnally computed our ho-

mography matrix H

re f

using the least square min-

imization method presented by Hartley and Zisser-

man (Hartley and Zisserman, 2003). Our next step

is to ﬁnd an estimate of the image coordinate of a per-

son’s feet



, v



on the ﬂoor based on the detected

upper-body bounding box. Hoiem et al. (Hoiem

et al., 2008) offers a solution that allows us to get,

knowing the image coordinates of a pixel that lies

on the parallel plane, its corresponding image coor-

dinates once projected on the reference plane. Deﬁn-

ing top



, v



and bottom



, v



points of known

objects allows us to retrieve the camera height (ﬁg-

ure 5). Indeed, if we know the height of the object,

then the camera height y

can be approximated as

follows, where v

is the horizon line computed from

the homography matrix H

re f

(Criminisi et al., 2000):

= h

−v

)

−v

)

. We selected an image that contains one

or several height references such as tables or beds.

We computed several camera heights and ﬁnally com-

puted the average to be y

= 2.3 m. We also assume

that the average height of a person is y

= 1.65 m. As

suggested by Hoiem et al. (Hoiem et al., 2008), we

estimate the focal length as being 1.4 times the image

height and we adjust v

. Finally, we compute an es-

timate of the image coordinate v

of the person’s feet

as follows: v

A+v

A+y

where A =

(1+

−v

)(v

−v

)

Using camera calibration information, we project the

detections in each frame to estimate the correspond-

ing ground position.

Having estimated the ground plane position for

each detected person, the next step is to obtain tra-

jectories from multiple detections. In doing so, we

ﬁrst compute tracklets by solving a frame to frame as-

signment problem using the Hungarian algorithm. A

mathematical formulation is presented by Pentico in

his survey on assignment problems (Pentico, 2007).

If we consider 2 consecutive frames, l with n detec-

tions and l + 1 with m detections, we can compute

a distance matrix C = (c

i j

) where c

i j

represents the

distance between object i in frame l and object j in

frame l + 1. The Hungarian algorithm then solves

the problem by minimizing the objective function

∑

i=1

∑

j=1

i j

under the constraints where x

i j

= 1

Figure 5: Calibration, (a) Ground plane outline for refer-

ence plane calibration, (b) Height references for parallel

plane calibration,(c) OR dimensions (in meters) and result-

ing table projected on the ground plane.

if the bounding box i in frame l is assigned to bound-

ing box j in frame l + 1, and x

i j

= 0 if not. Hence,

IDs are assigned to one or several bounding boxes

as they move along time. Nonetheless, due to mis-

detections, the data association solution can result in

multiple tracks for the same person. In order to deal

with that, we further cluster tracklets using the DB-

SCAN clustering algorithm (Ester et al., 1996).

We’ve chosen this algorithm as it has a physical

meaning when it comes to clustering points. In fact,

DBSCAN, Density-Based Spatial Clustering, ﬁnds

clusters based on density reachability. Two parame-

ters have to be speciﬁed: minPts, the minimum num-

ber of points that belong to a cluster and ε the radius

around a point that the algorithm has to look at for

merging. Centroids – that is, mean positions over time

of data points associated to each single track – are

Modeling Human Motion for Predicting Usage of Hospital Operating Room

331

considered for clustering. We set the minimum num-

ber of centroids to form a cluster to be minPts = 1 and

the radius to be ε = 0.5 meters. This technique allows

us to reduce the number of tracks and hence obtain

motion trajectories.

3.3.1 Model Estimation - Bivariate Gaussian

Kernel Density Estimation (GKDE)

If we consider the 2D histogram representing the spa-

tial distribution of points, we can account for the

fact that there are areas where people stay the most

or simply move through. A Kernel Density Estima-

tor (Bishop et al., 2006) provides a non-parametric es-

timate of the probability density function (pdf) g

(X)

over each track X



. . . x



associated to a

state S = i as follows, where states 1, 2 and 3 are re-

spectively ”Setting-Up”, ”Patient-Preparation”, and

”Ongoing-Surgery”:

(X) =

∑

n=1

2πh

exp



−



X −x





(1)

Basically, x

’s are successively occupied ground

position throughout time by an individual in the OR.

Each and every one of them lie at the center of an

hypercube (here a square) of side h to which we asso-

ciate a kernel function. Choosing a Gaussian kernel

function results in a smoother density model where

h represents the standard deviation of the Gaussian

components. The bandwidth h is selected as sug-

gested by Bowman and Azzalini (Bowman and Az-

zalini, 2004).

The K previously computed pdfs g

(X) are then

combined to give the pdf f

(X) characterizing the us-

age state S = i as follows, where

∑

k=1

is the total

number of data points associated to state S = i and π

the number of points in track X

∀i = 1 ... 3, f

(X) = p(X|S = i) =

∑

k=1

∑

k=1

(X)

(2)

Estimated models from training data for each us-

age state is shown in Figure 6. As seen in Figure 6(a),

occupancy is spread all over the room except for the

upper left corner of the room due to the presence of

diagnostic tools. In Figure 6(b), individuals tend to

have a patient centered activity, and one can easily

notice someone positioned behind the OR table. Fi-

nally, in Figure 6(c), one can discern 2 individuals on

either sides of the OR table and an individual on the

lower right exhibiting constrained motion patterns.

(a) Setting-Up (b) Patient-Preparation

LARGE(c) Ongoing-Surgery

Figure 6: Estimated distribution of trajectories over the

three states (as seen on colorbar: High Occupancy in red,

Low Occupancy in blue).

3.4 Upper-body Movement Feature

To obtain an estimate of the upper body movements,

we extract 2D optical ﬂow (Brox and Malik, 2011)

from localized regions over the observed space. We

split our image into a set of 3 ×3 uniformly spaced

cells. Magnitude weighted probability density func-

tions of 2D motion orientation are computed by Gaus-

sian Kernel Density Estimation in 1D (Bishop et al.,

2006) . Optical ﬂow estimation results in N magni-

tudes and orientations (µ

, θ

) within a cell. Each one

of them is considered a sample from the underlying

distribution of motions. A Gaussian kernel, where h is

selected as suggested by Bowman and Azzalini (Bow-

man and Azzalini, 2004) (equation 3) is used to ob-

tain an estimate of the overall unknown distribution

as shown in Figure 7.

p(θ) =

∑

n=1

Nµ

∑

n=1

√

2πh

exp



−

(θ −θ

)



(3)

This results in 9 probability distribution functions,

one associated with each cell. These are concatenated

into a single vector that is then reduced, using PCA, to

a ”projected feature vector”. The criteria chosen for

dimensionality reduction is that 90% of information is

retained in the new orthogonal basis. Finally, to build

a model for each state i, we accumulate the projected

feature vectors by computing their overall means M

over the training data for each state, respectively.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

332

Figure 7: (a) 3 ×3 Splitting, (b) 2D Optical ﬂow, (c) Local

motion probability density functions.

3.5 Usage State Decisions

Having learned the features models, given a new in-

put video stream, we need to compute a state estimate.

This is done by computing features from a set of con-

secutive image frames and matching the respective

features to each of the pre-computed models. In the

case of motion trajectory features, we compute likeli-

hood probabilities. The idea is that, assuming that we

know the model for each state, new temporal obser-

vations X = O

can be used to obtain evidence about

the underlying state they characterize.

p(S = i|X = O

1:n

) ∝ p(X = O

1:n

|S = i)×p(S = i) (4)

If we assume p(S = i) to be same for all states i, and

that observations are conditionally independent, then:

p(S = i|X = O

1:n

) ∝ p(X = O

1:n

|S = i)

∏

t=1

p(X = O

|S = i) =

∏

t=1

)

(5)

As a result, for each state i, our upper-body fea-

ture is reduced into a likelihood probability value

= p(X = O

1:n

|S = i) that will later on be used for

state inference. Having learned the models, given a

new input video stream, we compute cosine similari-

ties on sub-windows of the data. In the case of upper

body movement features, given new temporal obser-

vations X = O

1:n

over the time interval [[1 : n]], we

compute the mean

X. and compare it to each pre-

trained model M

using cosine similarity (equation 6).

d(S = i|X = O

1:n

) =

X.M

XkkM

(6)

Finally, for each state i, our upper-body feature is re-

duced to a cosine similarity value d

= d(S = i|X =

1:n

) that can be used for state inference.

The obtained likelihood and similarity values

across the three states from each of the two fea-

tures are ﬁrst normalized and then combined for ﬁnal

state prediction. The combination is based in simply

adding the normalized values for the respective states

and computing:

S = argmax



¯p

−1



, (7)

where ( ¯p

) are the normalized likelihood and nor-

malized cosine distances, respectively.

4 EXPERIMENTS

4.1 Dataset

Results presented in this paper are based on videos

taken by a single camera over different days in the

same OR. The videos were captured at a rate of

10 frames/sec, and segments identifying different us-

age states were manually annotated by the hospital

staff. The video available for each usage-state of the

OR is shown in Table 1.

Table 1: Video Database.

States Setting-Up Patient-Preparation Ongoing-Surgery

Time length 12min31sec 75min12sec 57min35sec

Number of frames 7510 45125 34552

To evaluate the accuracy of our approach, all re-

sults presented in the following are based on 10-fold

cross validation performed by considering 60% of our

data for training and 40% for testing. Nonetheless, to

compute motion trajectories as well as optical ﬂow,

consecutive frames are needed from the video. To

facilitate training and testing, we consider a set of

frames in our data to be of length L. We randomly

select an integer n in the 40

percentile. Then, we

consider to be our training set the following interval:

[n, bn + 60% ×Lc]. The remaining data is then used

for testing.

4.2 Results

In making a decision for the OR usage state, a mini-

mal number of consecutive frames are needed to com-

pute necessary features prior to matching against the

feature models. Hence, shown in Figure 8 is the accu-

racy of the system as a function of consecutive frames

(ω) prior to a decision. As seen, the total accuracy in-

creases with the increase in ω and reaches ≈80% with

ω = 100. This would be equivalent to making a deci-

sion every ε = 10 seconds.

To evaluate the contribution of each of the two fea-

tures, we also evaluated the accuracy obtained when

using the individual features. As can be seen in Fig-

ure 9, different usage states are better differentiable

based on one of the two features. For example, the

motion trajectories clearly are more useful for iden-

tifying ”Ongoing Surgery” as opposed to ”Setting-

Up”, which is better recognized based on upper body

movements.

We also considered the use of a smoothing win-

dow to rule out intermediate erroneous decisions. If

we consider a smoothing window of size δ, then for

each data point S(t) at time t we deﬁne our smoothing

Modeling Human Motion for Predicting Usage of Hospital Operating Room

333

Figure 8: 10-fold Cross validation (60% training /40% test-

ing) - Accuracy as a function of number of consecutive

frames used prior to making a decision.

(a) Upper body movement feature (b) Motion trajectory feature

Figure 9: 10-fold Cross validation (60% training /40% test-

ing) - Accuracy obtained by use of the individual features

as a function of number of consecutive frames used prior to

making a decision. (a) upper body movement feature, (b)

motion trajectory feature.

window as the consecutive data points [[S(t); S(t +δ)]].

The value of S(t) is then replaced by the most oc-

curring state within the window. Figure 10 shows an

overall accuracy of ≈88% for smoothing size δ = 10

and window size ω = 60. We evaluated other combi-

nations of smoothing size and window size, but with

the chosen values, a decision of the usage state is ob-

tained every 60 seconds, which is a reasonable rate.

5 DISCUSSION AND

CONCLUSIONS

In this paper, we presented a system that exploits ex-

isting video streams from an OR to infer OR usage

states. We deﬁned OR states that are relevant for as-

sessing OR usage efﬁciency. For this purpose, we

adopted a holistic approach that involves the com-

bination of two meaningful human motion features.

We took advantage of a detection algorithm as well

as a data association algorithm to reconstruct motion

trajectories. We evaluated discriminant occupancy

patterns using a kernel based method. We incorpo-

rated gesture or upper body movement information

by computing local motion histograms from 2D op-

tical ﬂow. Our system achieved encouraging results

with an overall accuracy of ≈90%.

We evaluated the contribution of each of the two

Figure 10: 10-fold Cross validation (60% training /40%

testing) - Accuracy as a function of smoothing window size.

The number of consecutive frames = 60.

features separately and considered the number of con-

secutive frames needed to capture sufﬁcient motion

information to differentiate the deﬁned usage states.

We found that each of the feature provides useful

information to reduce ambiguity. The upper body

movement feature from optical ﬂow provides extra el-

ements of information about the ”Setting-Up” state.

Basically, human motion trajectories estimate motion

of the feet. As a result, discriminant information

about the activity happening over the OR table is dis-

carded since people are typically not moving in the

OR at this state. However, the optical ﬂow feature

is able to capture the upper body movements includ-

ing hands over the OR table. Further, hand activity

over the OR table differs when a patient is over the

table compared to that of arranging instruments and

equipment. As a result, it succeeds in discriminat-

ing the ”Patient-Preparation” state from the ”Setting-

Up” state.

Future work will involve enhancing trajectory re-

construction by exploiting image features. Further,

the independence assumption made when inferring

the usage state is rather simplistic. Therefore, tak-

ing advantage of the established usage state transition

models and exploiting time-dependent decision mod-

els will be further investigated.

REFERENCES

Arbab-Zavar, B., Carter, J. N., and Nixon, M. S. (2014). On

hierarchical modelling of motion for workﬂow analy-

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

334

sis from overhead view. Machine vision and applica-

tions, 25(2):345–359.

Association, H. F. M. et al. (2003). Achieving operating

room efﬁciency through process integration. Health-

care ﬁnancial management: journal of the Healthcare

Financial Management Association, 57(3):suppl–1.

Bardram, J. E., Hansen, T. R., and Soegaard, M. (2006).

Awaremedia: a shared interactive display supporting

social, temporal, and spatial awareness in surgery. In

Proceedings of the 2006 20th anniversary conference

on Computer supported cooperative work, pages 109–

118. ACM.

Behera, A., Cohn, A., and Hogg, D. (2014). Real-time ac-

tivity recognition by discerning qualitative relation-

ships between randomly chosen visual features. In

Proceedings of the British Machine Vision Confer-

ence. BMVA Press.

Bhatia, B., Oates, T., Xiao, Y., and Hu, P. (2007). Real-time

identiﬁcation of operating room state from video. In

AAAI, volume 2, pages 1761–1766.

Bishop, C. M. et al. (2006). Pattern recognition and ma-

chine learning, volume 4. springer New York.

Bowman, A. W. and Azzalini, A. (2004). Applied smooth-

ing techniques for data analysis. Clarendon Press.

Brox, T. and Malik, J. (2011). Large displacement optical

ﬂow: descriptor matching in variational motion esti-

mation. Pattern Analysis and Machine Intelligence,

IEEE Transactions on, 33(3):500–513.

Ciechanowicz, S. and Wilson, N. (2011). Delays to operat-

ing theatre lists: observations from a uk centre. The

Internet Journal of Health, 12(2).

Criminisi, A., Reid, I., and Zisserman, A. (2000). Single

view metrology. International Journal of Computer

Vision, 40(2):123–148.

Does, R. J., Vermaat, T. M., Verver, J. P., Bisgaard, S., and

Van den Heuvel, J. (2009). Reducing start time delays

in operating rooms. Journal of Quality Technology,

41(1):95–109.

Ester, M., Kriegel, H.-P., Sander, J., and Xu, X. (1996).

A density-based algorithm for discovering clusters in

large spatial databases with noise. In Kdd, volume 96,

pages 226–231.

Felzenszwalb, P. F., Girshick, R. B., and McAllester, D.

(2010a). Discriminatively trained deformable part

models, release 4.

Felzenszwalb, P. F., Girshick, R. B., McAllester, D., and

Ramanan, D. (2010b). Object detection with discrim-

inatively trained part based models. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

32(9):1627–1645.

Hartley, R. and Zisserman, A. (2003). Multiple view geom-

etry in computer vision. Cambridge university press.

Hoiem, D., Efros, A. A., and Hebert, M. (2008). Putting

objects in perspective. International Journal of Com-

puter Vision, 80(1):3–15.

Kodali, B. S., Kim, D., Bleday, R., Flanagan, H., and Ur-

man, R. D. (2014). Successful strategies for the re-

duction of operating room turnover times in a tertiary

care academic medical center. Journal of Surgical Re-

search, 187(2):403–411.

Lange, P. M., Nielsen, K. L. G., Petersen, S. T., and

Bardram, J. E. (2010). Phase recognition in an op-

erating room using sensor technology. IT-university

of Copenhagen.

Lea, C., Facker, J., Hager, G., Taylor, R., and Saria, S.

(2013). 3d sensing algorithms towards building an in-

telligent intensive care unit. AMIA Summits on Trans-

lational Science Proceedings, 2013:136.

Macario, A. (2010). What does one minute of operat-

ing room time cost? Journal of clinical anesthesia,

22(4):233–236.

Nara, A., Izumi, K., Iseki, H., Suzuki, T., Nambu, K., and

Sakurai, Y. (2011). Surgical workﬂow monitoring

based on trajectory data mining. In New Frontiers in

Artiﬁcial Intelligence, pages 283–291. Springer.

Niu, Q., Peng, Q., El Mekkawy, T., Tan, Y. Y., Bruant, H.,

and Bernaerdt, L. (2011). Performance analysis of the

operating room using simulation. Proceedings of the

Canadian Engineering Education Association.

Padoy, N., Mateus, D., Weinland, D., Berger, M.-O., and

Navab, N. (2009). Workﬂow monitoring based on

3d motion features. In Computer Vision Workshops

(ICCV Workshops), 2009 IEEE 12th International

Conference on, pages 585–592. IEEE.

Pentico, D. W. (2007). Assignment problems: A golden

anniversary survey. European Journal of Operational

Research, 176(2):774–793.

Schuster, M., Pezzella, M., Taube, C., Bialas, E., Diemer,

M., and Bauer, M. (2013). Delays in starting morning

operating lists: An analysis of more than 20 000 cases

in 22 german hospitals. Deutsches

Arzteblatt Interna-

tional, 110(14):237.

the NYP Wall of Knowledge, L. and manager, L. P.-O. Or-

dashboard.

Veres, G., Grabner, H., Middleton, L., and Van Gool, L.

(2011). Automatic workﬂow monitoring in indus-

trial environments. In Computer Vision–ACCV 2010,

pages 200–213. Springer.

Voulodimos, A., Kosmopoulos, D., Veres, G., Grabner, H.,

Van Gool, L., and Varvarigou, T. (2011). Online clas-

siﬁcation of visual tasks for industrial workﬂow mon-

itoring. Neural Networks, 24(8):852–860.

Xiao, Y., Hu, P., Hu, H., Ho, D., Dexter, F., Mackenzie,

C. F., Seagull, F. J., and Dutton, R. P. (2005). An

algorithm for processing vital sign monitoring data to

remotely identify operating room occupancy in real-

time. Anesthesia & Analgesia, 101(3):823–829.

Modeling Human Motion for Predicting Usage of Hospital Operating Room

335