Model-based Detection and Analysis of Animal Behaviors
using Signals Extracted by Automated Tracking
Gennady Denisov, Tomoko Ohyama, Tihana Jovanic and Marta Zlatic
Janelia Farm Research Campus, Howard Hughes Medical Institute, Ashburn, Virginia, 20147, U.S.A.
Keywords:
Animal Behavior, Automated Tracking, High-throughput Detection, Model, Algorithm, Software, Signal
Processing, Statistical Analysis, Feature Extraction, Hit Detection.
Abstract:
Analysis of behaviors of model organisms has a number of applications, particularly to determination of the
function of genes and neurons. Drosophila larva is an especially convenient model system for this kind of
study because of availability of powerful genetic analysis tools and of automated tracking software that allows
high-throughput recording of animal’s shape and position characteristics as time-dependent signals. We have
developed an open source software that allows a high-throughput detection and analysis of a comprehensive set
of meaningful behaviors of this species. Using the recorded signals as input variables and a set of processing
thresholds as parameters, the software employs model-based algorithms to detect the behavioral actions with
high accuracy, typically 1-5%. For each detected action it extracts and stores meaningful quantitative features
that allow statistical discrimination of mutants from wild type animals and set stage for subsequent application
of machine learning techniques to classification of the mutants.
1 INTRODUCTION
Insights into the function of a gene or neuron can
be gained in multiple ways, including the loss-of-
function screening of genes and neurons underlying
observed behavioral phenotypes. Drosophila larva is
an especially tractable and convenient model system
for this kind of study because of the relative simplicity
of its nervous system, availability of powerful genetic
analysis tools (Pfeiffer et al., 2008), and availability
of automated tracking software, such as Multi-Warm
Tracker (MWT) (Swierczek et al., 2011). Given a
contour of a tracked animal object, the MWT makes
a guess about the position of its spinal cord (Fig. 1),
assigns a ”center of mass” to the object, computes a
variety of other metrics and stores the results as time
series data.
Figure 1: Drosophila larva objects tracked by Multi-Worm
Tracker (MWT).
Automated analysis of behaviors of Drosophila
larva has been primarily focused on study of different
types of taxis (Gomez-Marin et al., 2011), (Gomez-
Marin and Louis, 2012), (Gershow et al., 2012), (Luo
et al., 2010). The most high-throughput approach,
employing custom machine vision software, was used
in (Gershow et al., 2012). Machine vision was also
applied to analysis of social and sex behaviors in
adult Drosophila melanogaster (Branson et al., 2009),
where a machine learning classifier was used for au-
tomatic detection of animal behaviors with high ac-
curacy. A typical machine learning classifier, which
takes tracking movies as input, may employ from sev-
eral hundreds to several thousand features. This has
two drawbacks. First, the number of features that has
to be processed reduces the efficiency of behavior de-
tection. Screening of thousands of mutant lines, with
each line being represented by a sufficiently large
group of animals for making statistical conclusions,
under a variety of experimental conditions/stimuli in
order to explore different behavioral actions, will re-
quire efficient data analysis algorithms. Second, it is
hard to know which of the used features are really im-
portant. Knowing which motor patterns are actually
altered in specific mutants is a key to gaining biolog-
ical insight into the function of a gene or neuron.
For this reason, we explored a different approach,
175
Denisov G., Ohyama T., Jovanic T. and Zlatic M..
Model-based Detection and Analysis of Animal Behaviors using Signals Extracted by Automated Tracking.
DOI: 10.5220/0004235101750181
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 175-181
ISBN: 978-989-8565-36-5
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
Figure 2: Signals used as input by SALAM: (a) (crawling) speed, mm/s; (b) crabspeed (or rolling speed), mm/s; (c) midline,
mm; (d) morpwidth, mm; (e) area of the tracked object, mm
2
; (f) (head) cast, degrees; (g) x and y, mm, (h) tailvecx and
tailvecy, mm; and (i) headx and heady, mm.
which makes use of the signals extracted by MWT,
rather than of tracking movies. This approach has
been implemented in an open source software pack-
age SALAM (http://sourceforge.net/projects/salam-
hhmi), written in Python and R programming lan-
guages. By properly combining the input signals,
employing a set of specifically tuned thresholds and
model-based algorithms, our software allows detec-
tion of a comprehensive set of meaningful behavioral
actions of Drosophila larva with high accuracy and
throughput. The detected actions include, but are not
limited to, the crawling runs, head casts and the ear-
lier studied types of taxis.
SALAM has been used by our recent study
(Ohyama et al., 2012). This paper provides an
overview of its capabilities and algorithms.
2 APPROACH OVERVIEW:
SIGNALS, BIOLOGICALLY
MEANINGFUL ACTIONS AND
DATA PROCESSING
WORKFLOW
Figure 2 schematically represents several signals
output by Choreography software of MWT pack-
age (http://sourceforge.net/projects/mwt) and used by
SALAM as input. Speed, or crawling speed (a), and
crabspeed, or rolling speed (b), are defined as the
speed of the center of mass in the direction parallel
and perpendicular to the spine, respectively; (c) mid-
line is the spine length; (d) morpwidth is measured
at the center of mass perpendicular to the spine; (f)
(head) cast is defined as the angle formed between
the straight lines drawn for the top 20% and the bot-
tom 80% of the spine; (g) (x, y) is the position of the
center of mass, (h) (tailvecx, tailvecy) are components
of the vector joining the bottom and middle points of
the spine; and (i) (headx, heady) is the position of an
animal nose, i.e. the top point of the spine.
Figure 3 schematically outlines behavioral ac-
tions observed in Drosophila larva and detected by
SALAM, using its built-in algorithms, given the sig-
nals listed in Fig. 2. Crawling, Fig. 3(a), and casting,
Fig. 3(b), are the most typical actions: these are what
an animal is doing most of the time. Other actions
usually occur in response to a certain kind of stimu-
lus. For example, hunch, or contraction of animal’s
body, Fig. 3(c), and rolling, Fig. 3(d), usually oc-
cur in response to a time-dependent stimulus, such as
IR light or ultrasound. Digging, whereby an animal
pops up in an attempt to dig a hole in the support,
Fig. 3(e), and following, Fig. 3(f), typically occur on
supports possessing scratches. Animals may be pre-
vented from digging and/or following by other stim-
uli, for example by a strong air flow. Finally, taxis
action, Fig. 3(g), may be taken when there is a spatial
gradient of a stimulus. Such an action may include
two components: 1) turning an animal’s body towards
the source of stimulus (positive taxis) or from it (neg-
ative taxis), and 2) navigating (crawling) towards the
area with higher or lower intensity of the stimulus, re-
spectively.
Fig. 4 overviews a scope of operations and a
workflow performed by SALAM.
BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
176
Figure 3: Meaningful behavioral actions detected by SALAM: (a) (peristaltic) crawling, (b) cast, where the middle chart
represents the possibility of a strong cast, (c) hunch, (d) rolling, (e) digging, (f) following, i.e. crawling along a scratch
on a support, and (g) taxis, where the stimulus may be an odor (chemotaxis), a temperature (thermotaxis), a visible light
(phototaxis), or an air pressure, resulting in air flow (thigmotaxis).
Time series
data
Extracted
features
Detected
hits
Signal
processing
Show signals
and events
Show
hits
Produce
histograms
Hypotheses
testing
Figure 4: Flowchart of data processing. The upper row of boxes represents the core, high-throughput processing. The lower
row of boxes represents visualization tools, which are typically applied to a limited amount of data with the purposes of
debugging the algorithms or reviewing the results.
3 DETECTING BEHAVIORAL
ACTIONS AND EXTRACTING
FEATURES
Analysis of behavioral actions by SALAM starts from
detecting events in the individual signals used.
3.1 Detection of Individual Signal
Events
For non-oscillating signals, such as cast, crabspeed,
midline or morpwidth, events are simply the signifi-
cant peaks or wells, with amplitude above the speci-
fied thresholds. Detection of events in oscillating sig-
nals (e.g. speed, area) resulting from the peristaltic
nature of larval crawling, takes a different approach.
In either case, event detection is completed by com-
puting an event signal, which is nonzero at events
(equals the event amplitude) and zero otherwise.
Event detection in an oscillating signal is initi-
ated by identifying peaks. Peak positions are the lo-
cal maxima of a signal. Peak amplitude is a signal
value at the peak position. Peak boundaries are set at
the minima of signal on both sides of a peak. A peak
is considered good if its amplitude exceeds a speci-
fied threshold. Positions of good peaks in speed sig-
nal are shown as dashed green lines in the left plot
Model-basedDetectionandAnalysisofAnimalBehaviorsusingSignalsExtractedbyAutomatedTracking
177
of Fig. 5. A speed event is a sequence of at least
three adjacent good peaks. The event amplitude is
the mean height of all the good peaks comprising the
event. Event boundaries are set at the boundaries of
the first and last peak comprising the event. (When
detecting crawling actions, the algorithm additionally
requires that a crawling event is terminated by any
event in cast or crabspeed signal, Fig. 5).
Event detection in a non-oscillating signal. Our
procedure is an extension of the ”Schmitt trigger”
approach previously used for detection of movement
events in flies (Robie et al., 2010). It employs four
adjustable thresholds, which are specifically tuned for
each type of signal. The thresholds are: 1) the upper
and 2) the lower amplitude threshold, shown as the
solid and dashed horizontal green line, respectively,
in Fig. 5 (cast and crabspeed plots) and in Fig. 6;
3) the width threshold; and 4) the gap threshold. An
event starts when the absolute value of a signal, while
increasing as a function of time, crosses the upper am-
plitude threshold. An event ends when the absolute
value of a signal, while decreasing as a function of
time, crosses the lower amplitude threshold. Event
duration is the difference between the event end and
event start times. Event amplitude is the highest abso-
lute value of a signal during the event. A single event
of duration less than the width threshold will not be
detected (i.e., will be ignored). However, if two or
more adjacent events of the same type (all peaks or all
wells) are less than the gap threshold apart one from
another, and the time duration between the start of the
first event and the end of the last event exceeds the
width threshold, then all the events will be merged
into a single detected event, as illustrated by the crab-
speed plot in Figure 5.
3.2 Detection of Behavioral Actions
Detection of the most of actions listed in Table 1, with
the only exception for rolling, requires simultaneous
processing of more than one signal. This requirement
can be illustrated by comparing the strong cast (Fig.
3(b), middle chart), hunch, Fig. 3(c), and digging ac-
tion, Fig. 3(e). During any of these actions, a drop
in the midline signal is observed, so the shape of a
tracked animal object is close to a ball (and there-
fore the MWT may be unable to properly identify the
spine line). Thus, while the midline signal can be used
for detecting these actions, it may not be sufficient
for their reliable discrimination, so additional signals
must be used.
Our model algorithm for detection of a strong cast
action makes use of the cast, midline and morpwidth
signals, as illustrated in Fig. 6. First, we note that a
strong cast must be accompanied not only by a well
in the midline signal, but also by a peak in the mor-
pwidth signal. (The morpwidth peak is not expected
to be observed neither during hunch nor during dig-
ging action.) Second, a strong cast can only appear
as a part of the sequence of states schematically de-
picted in Fig. 3(b). Thus, one should always observe
cast signal events (peaks or wells) on both sides of a
midline well/morpwidth peak during the strong cast
action, as illustrated by Fig. 6.
Detection of a hunch action makes use of the same
three variables as detection of cast action, but employs
additional thresholds to make sure there are no signif-
icant cast and morpwidth events in a close vicinity of
the midline well.
Our algorithm for detection of a digging action, in
addition to the presence of a deep well in a midline
signal, requires that the values of both x and y sig-
nal stay constrained to a certain small region within a
certain period of time, since the tracked animal object
practically does not move while the animal digs.
A following action is detected similarly to the de-
tection of digging action, but does not require the
presence of a midline well and requires that x/y
stays below a certain small threshold for a certain
period of time, since in all the experiments we used
scratches were parallel to the y axis.
Detection of taxis actions is more sophisticated.
For example, during chemotaxis, an animal usually
performs two or more head casts, both left and right
ones, and compares the local concentrations of an
odorant at the position (headx, heady) of its nose dur-
ing each cast (Gomez-Marin et al., 2011). It then
changes its orientation, described by the tail vec-
tor with components (taivecx, tailvecy), and subse-
quently crawls for a certain time in the newly chosen
direction.
3.3 Feature Extraction
The features defined for detected actions can be sub-
divided into two categories: those computed per-
event, e.g. event amplitude, and those computed per-
animal, e.g. an event frequency, i.e. number of events
detected per unit time interval. Table 1 lists the fea-
tures extracted for detected actions. For crawling ac-
tion, only per-event features are defined: the mean
peak amplitude (height), frequency of oscillation (de-
termined by applying the Fourier transform to the por-
tion of speed signal comprising the action event, and
the event duration. For head cast, hunch and rolling
actions, the features are the event amplitude (maximal
value of a signal during the event) and event duration,
as well as the event frequency. For digging, the only
BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
178
time, s
speed, mm/s
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5
6.0
56 58 60 62 64 66 68 70
# 00493
time, s
cast
−150
−100
−50
0
50
100
150
56 58 60 62 64 66 68 70
# 00493
time, s
crabspeed, mm/s
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5
6.0
56 58 60 62 64 66 68 70
# 00493
Figure 5: Detection of crawling actions. Shown in blue are three signals recorded for tracked animal #493 in a particular
time window: speed (left plot), cast (middle plot) and crabspeed (right plot). The red line indicates an event signal, which is
nonzero only at detected events. The green lines provide auxiliary information: positions of detected good peaks in the speed
signal plot, and the upper (solid line) and lower (dashed line) amplitude thresholds used for event detection in the cast and
crabspeed signal plots.
time, s
cast
−40
−20
0
20
40
60
80
100
120
140
160
180
14 16 18 20 22 24 26 28
# 00004
time, s
midline
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
1.05
1.10
1.15
14 16 18 20 22 24 26 28
# 00004
time, s
morpwidth
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
14 16 18 20 22 24 26 28
# 00004
Figure 6: Detection of head cast actions. Shown in blue are three signals extracted by tracking animal #4 in a particular
time window: cast (left plot), midline (middle plot) and morpwidth (right plot). Here, the midline and morpwidth signals are
normalized to their median values. The meaning of the red and green lines is the same as in Fig. 5. Two detected cast actions
have been shown in the left plot: the regular cast (left) and the strong cast (right). The strong cast can only occur as a part of
a sequence of events schematically shown in Fig. 3(b). It is assigned a ”theoretical” amplitude of 180
0
, since MWT is unable
to properly determine the value of cast signal for a ball-like object.
feature is the event frequency.
For taxis events, a number of features have been
extracted, of which we will here discuss only the nav-
igation index. For experiments where a gradient of
a stimulus is constant and parallel to the x-axis, one
way to define the navigation index is (Gershow et al.,
2012)
nind =< speed
x
> / < speed > (1)
where speed
x
stands for the x-component of the speed
signal, and the angle brackets designate averaging
over time. According to this formula, the naviga-
tion index should be always within the range [-1, +1],
reaching the value +1 for animals moving strongly in
the direction of a stimulus source and -1 for animals
moving strongly in the opposite direction. While this
definition can work reasonably well in many cases, it
has two limitations. First, if an animal with strongly
positivetaxis was initially, at time t = 0, located a min-
imal possible distance away from the stimulus source,
then (1) will give nind = 0 (since the animal can
no longer move toward the source), rather than nind
= +1. Second, formula (1) is not applicable to the
settings with (multiple) point source(s) of stimulus,
i.e. with non-flat geometry, which is being the case
in studies of chemotaxis (Gomez-Marin and Louis,
2012).
We here present an alternative approach to the
computation of navigation index, which attempts to
overcome the two limitations mentioned above. Our
approach is based on comparing the times an ani-
mal spent at different distances away from a stimulus
source during experiment.
Let d(t) be a distance from an animal object to
the closest (flat or point) source of stimulus at time t.
We assume that, for a given experimental setting, the
minimal and maximal possible values of d are d
min
and d
max
, respectively. Then, the navigation index can
be computed as
Model-basedDetectionandAnalysisofAnimalBehaviorsusingSignalsExtractedbyAutomatedTracking
179
Table 1: Detected behavioral actions, signals used, accuracy of detection (false discovery rate, estimated by comparing the
predictions from the software with tracking movies of contours of larvae) and features extracted and stored for each the action.
The features are: amplitude, ampl; (crawling) frequency, freq; action event duration, dur; event frequency, efreq; fractions of
time and path length spent following, ftime and flength; and the navigation index, nind. See the text for more detailed feature
definitions. Accurate estimates of the FD rate for digging, following and taxis actions are not currently available. False
negative rate was not estimated, since it does not affect (at least, directly) the statistical characteristics of extracted features,
e.g. their mean values.
Action Signals used FD rate Extracted features
crawling speed, cast, crabspeed 5% ampl, freq, dur
head cast cast, midline, morpwidth 1% ampl, dur, efreq
hunch midline, cast, morpwidth 5% ampl, dur, efreq
rolling crabspeed 1% ampl, dur, efreq
digging x, y, midline efreq
following x, y ftime, flength
taxis tailvecvx, tailvecy, cast, headx, heady, x, y nind, ...
nind = (d
max
+ d
min
2· < d >)/(d
max
d
min
) (2)
This formula can be applied both to settings with a flat
gradient of stimulus (by simply replacing d for x) and
to settings with (multiple) point source(s) of stimulus.
In the limiting case of an animal with strongly posi-
tive taxis, which spent all the time of experiment at a
minimal possible distance from the source of stimu-
lus, so that d(t) = d
min
, formula (2) gives nind = +1.
In the opposite case of an animal with strongly neg-
ative taxis which spent all the experimental time at a
maximal possible distance from the source of stumu-
lus, it gives nind = -1. A further adjustment of this
formula is possible for non-flat geometries, by tak-
ing into account a variation in the experimental space
available for animals located at different distances d
from the closet source of stimulus.
4 HYPOTHESES TESTING AND
HIT DETECTION
To discriminate the behaviors of the wild type and
mutant animal groups, we compare their feature dis-
tributions/histograms. Our null hypothesis is that the
differences between the (possibly, noisy) histograms
are statistically insignificant, so that the groups be-
long to the same population. Whenever evidence is
found that the two distributions are different enough
so that the groups cannot belong to the same popula-
tion, we say there is a ”hit”.
The hit detection functionality of SALAM pack-
age is currently under development. At this time,
the software is only capable of detecting hits based
on analysis of one feature at a time. The particu-
lar approach to hit detection varies depending on the
type of a feature. For per-animal features, e.g. the
presence or absence of roll actions in a given animal
trace, we employ the statistical tests available for pro-
portions, notably Pearson’s χ
2
-test or Fisher’s exact
test. To compare the mean values of features with
continuous distributions, binary parametric (t-test) or
non-parametric (e.g. Wilcoxon rank sum) tests are
available. If the distribution approximately meets a
multivariate normality requirement, even more accu-
rate, multivariate test can be used, as illustrated be-
low with an example of crawling speed amplitude.
This approach, in addition to the mean, involves com-
parison of several other characteristics of the distri-
bution, all of them being computed based on statis-
tical moments: the standard deviation, the skewness
and the kurtosis. The approach borrows its formal-
ism from Hotelling’s theory for T
2
-distributions (see,
e.g. (Rencher, 2002)), but allows empiric calibration
of thresholds, based on existing wild type data, rather
than using theoretical thresholds derived for strictly
multivariate normal distributions.
Let U
(i)
be a vector of characteristics (mean, std-
dev, skewness, kurtosis) representing a histogram of
i-th run of a wild type population,
¯
U = (1/n) ·
U
(i)
be the mean of all the wild type runs, and
ˆ
Σ = (1/(n
1))·
(U
(i)
¯
U)(U
(i)
¯
U)
T
be sample (4×4) covari-
ance matrix. Then, givena desired confidence level α,
one can build an ellipsoid representing a Mahalanobis
vicinity of
¯
U
(U
(i)
¯
U)
T
ˆ
Σ
1
(U
(i)
¯
U) D
thr
(3)
with the threshold distance D
thr
being calibrated so
as to include exactly (1 α) portion of all the wild
type runs in the Mahalanobis vicinity. Now, for any
run of a mutant population with representative vec-
tor of characteristics located within the Mahalanobis
vicinity, we say that it is undistinguishable from the
wild type population with confidence α, i.e. the null
hypothesis is met. Conversely, any mutant run with
characteristics located outside this vicinity will repre-
sent a hit (Fig. 7).
BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
180
1.0 1.5 2.0 2.5 3.0
0.2 0.4 0.6 0.8
mean
stddev
positive control1
negative control1
negative control2
positive control2
positive control3
Figure 7: Multivariate hit detection procedure applied to
crawling speed amplitude. Shown in the figure are two out
of the four used variables: the mean and the standard de-
viation. The open circles represent the individual runs of
a wild type population, with each the run comprising ap-
prox. 100 animals. Multivariate normality of this data was
confirmed by the test implemented in R package ”energy”.
The dashed ellipse, drawn using R package ”chemomet-
rics”, represents a 90%-confidence Mahalanobis vicinity of
the center of wild type runs, so that α = 0.1. The lled
circles represent pooled data for two negative control lines
and three positive control lines. A hit is detected when-
ever a filled circle falls outside the Mahalanobis vicinity, so
this chart confirms biological expectations for the indicated
confidence level.
5 CONCLUSIONS
We haveprovided an overviewof SALAM, a software
package for Statistical Analysis of Larval Motions.
The input signals taken by the software, the detected
behavioral actions and the processing workflow have
been illustrated graphically. The algorithms used for
data processing have been discussed. Detection of the
two most typical behavioral actions, the crawling and
the head cast, as well as a multivariate procedure for
detection of hits, have been illustrated with examples.
This software is being used as a part of an automated
data analysis pipeline for screening mutant lines in or-
der to determine the function of genes and neurons.
ACKNOWLEDGEMENTS
The authors are grateful to Prof. Daniel Naiman for
consultation on statistical data analysis.
REFERENCES
Branson, K., Robie, A. A., Bender, J., Perona, P., and Dick-
inson, M. H. (2009). High-throughput ethomics in
large groups of drosophila. Nature Methods, 6:451–
457.
Gershow, M., Berck, M., Mathew, D., Luo, L., Kane, E. A.,
Carlson, J. R., and Samuel, A. D. (2012). Controlling
airborne cues to study small animal navigation. Na-
ture Methods, 9:290–296.
Gomez-Marin, A. and Louis, M. (2012). Active sensation
during orientation behavior in the drosophila larva:
more sense than luck. Curr. Opin. Neurobiol., 22:208–
215.
Gomez-Marin, A., Stephens, G. J., and Louis, M. (2011).
Active sampling and decision making in drosophila
chemotaxis. Nature Communications, 22:441.
Luo, L., Gershow, M., Rosenzweig, M., Kang, K., Fang-
Yen, C., Garrity, P. A., and Samuel, A. D. (2010).
Navigational decision making in drosophila thermo-
taxis. J. Neurosci., 30:4261–4272.
Ohyama, T., Jovanic, T., Denisov, G., Hoffmann, D., Kerr,
R. A., and Zlatic, M. (2012). Automated high-
throughput analysis of stimulus-evoked ethograms
in drosophila larva reveals modality-specific escape
strategies. Submitted for publication.
Pfeiffer, B. D., Jenett, A., Hammonds, A. S., Ngo, T. T.,
Misra, S., Murphy, C., Scully, A., Carlson, J. W.,
Wan, K. H., Laverty, T. R., Mungall, C., Svirskas,
R., Kadonaga, J. T., Doe, C. Q., Eisen, M. B., Cel-
niker, S. E., and Rubin, G. M. (2008). Tools for
neuroanatomy and neurogenetics in drosophila. Proc.
Natl. Acad. Sci. USA, 105:9715–9720.
Rencher, A. (2002). Methods of Multivariate Analysis. John
Wiley and Sons, Inc., second edition.
Robie, A. A., Straw, A. D., and Dickinson, M. H. (2010).
Object preference by walking fruit ies, drosophila
melanogaster, is mediated by vision and gravipercep-
tion. J. Exp. Biol., 213:2494–2506.
Swierczek, N. A., Giles, A. C., Rankin, C. H., and Kerr,
R. A. (2011). High-throughput behavioral analysis in
c. elegans. Nature Methods, 8:592–598.
Model-basedDetectionandAnalysisofAnimalBehaviorsusingSignalsExtractedbyAutomatedTracking
181