You are on page 1of 11

| |

Received: 2 May 2018    Revised: 18 September 2018    Accepted: 9 October 2018

DOI: 10.1111/aje.12573

SPECIAL SECTION : CAMERA TRAPPING IN AFRICA

Are camera traps fit for purpose? A rigorous, reproducible and


realistic test of camera trap performance

Peter Apps  | John Weldon McNutt

Botswana Predator Conservation Trust,


Maun, Botswana Abstract
Camera traps are now widely used to produce the data that inform critical, large‐
Correspondence
Peter Apps, Botswana Predator scale conservation strategies, but insufficient attention has been paid to evaluating
Conservation Trust, Private Bag X13, Maun, or validating their performance. A realistic, reproducible and rigorous test of camera
Botswana.
peterjapps@gmail.com trap performance has been developed, using a real animal target moving in different
directions and at different distances from sets of up to six cameras. The test clearly
Funding information
Wild Entrust International; Tusk Trust; shows differences in performance between camera trap brands and models that af‐
Private donors; Self funded fect whether they are fit for research purposes, and the effects on camera perfor‐
mance of the small differences in mounting height and angle that are likely to arise
when camera traps are used in the field.
Résumé
Les pièges photographiques sont aujourd’hui largement utilisés pour produire des
données qui informent des stratégies de conservation critiques et à grande échelle,
mais l’on accorde trop peu d’attention à l’évaluation ou à la validation de leurs perfor‐
mances. Un examen réaliste, répétable et rigoureux des performances des pièges
photographiques a été mis au point, en utilisant une véritable cible animale qui se
déplace dans différentes directions et à différentes distances d’ensembles qui
comptent jusqu’à six appareils photo. Le test montre clairement des différences de
performances selon la marque et le modèle d’appareil, qui influencent le fait qu’ils
soient, ou non, adaptés aux besoins d’une recherche, et les effets sur les perfor‐
mances d’un appareil de petites différences dans sa hauteur et de l’angle de vue, qui
sont susceptibles d’apparaître lorsque l’on utilise des pièges photographiques sur le
terrain.

KEYWORDS
camera trap, capture‐recapture, fit for purpose, occupancy, trail camera, validation

1 | I NTRO D U C TI O N for purpose of different camera trap brands and models, verifying the
performance of individual cameras, measuring the effect on perfor‐
The use of camera traps in field biology has exploded since about 2002 mance of how camera traps are set up, and validating camera trap‐
(Meek & Pittet, 2012; Wearn & Glover‐Kapfer, 2017), and they are now ping methods, so that results and conclusions are robust and reliable.
widely used to generate data that inform critical conservation man‐ Testing cameras before they are deployed to the field is recommended
agement decisions at ecosystem and landscape scale (Swanson et al., by Wellington, Bottom, Merrill, and Litvaitis (2014), and reviews of
2015; TEAM, 2011). Since camera trapping results inform such import‐ camera trapping (e.g. Burton et al., 2015) have stressed the importance
ant decisions, there needs to be a rigorous way of evaluating the fitness of camera performance to the collection of quality data that generate

710  |  wileyonlinelibrary.com/journal/aje


© 2018 John Wiley & Sons Ltd Afr J Ecol. 2018;56:710–720.
APPS and McNUTT |
      711

results that are fit for purpose. Nonetheless, although there has been cameras that do well with humans as targets are a bitter disap‐
some consideration of camera traps’ limitations (Burton et al., 2015; pointment to field biologists who are trying to detect and record
Meek, Ballard, & Fleming, 2015), comparison between camera models the presence of animals and their behaviour in the wild. Also, un‐
has been “rudimentary” (Meek et al., 2015), and the performance of like the humans in Internet review walk tests, wild animals do not
some cameras is disturbingly poor and disturbingly variable between always walk neatly across the field of view at right angles to the
units (Ballard, Meek, Doak, Fleming, & Sparkes, 2014; Hamel et al., camera’s direction of aim, but Glen, Cockburn, Nichols, Ekanayake,
2013; Hughson, Darby, & Dungan, 2010; Lyra‐Jorge, Ciocheti, Pivello, and Warburton (2013) had screens in front of the cameras that
& Meirelles, 2008; Meek et al., 2015; Newey et al., 2015; Pease, constrained their test targets to walk across the fields of view, and
Nielsen, & Holzmueller, 2016; Swann, Hass, Dalton, & Wolf, 2004). Rowcliffe, Carbone, Jansen, Kays, and Kranstauber (2011) consid‐
There is no accepted protocol for testing camera traps that is realistic ered only the distance from the camera and angular displacement
enough to predict their performance in the field, and most investiga‐ from the direction of aim, not the direction of the animals’ move‐
tors simply assume that camera traps are fit for purpose, and then use a ments. A test target that more closely resembles the animals that
proliferation of increasingly elaborate models to process their data (e.g. biologists are interested in, and that moves around like they do,
review O’Brien, 2011), which may not improve the quality of results if would improve how well the tests predict performance in the field.
the model’s assumptions are violated (Burton et al., 2015). Field tests with wild animals as targets (e.g. Hughson et al., 2010)
Camera traps cannot always be set up optimally—they may need avoid the problem of the target and test conditions being unrealistic,
to be up trees to avoid being stolen by humans (Meek, Ballard, & but suffer from the equally serious problem that the only record
Falzon, 2016) or high on a pole to avoid being chewed by hyaenas of the cameras’ performance is the images that they capture—the
(Apps, Claase, Yexley, & McNutt, 2017), and if they are fixed to trees true positives. The false negatives of animals that were not captured
or other natural supports, their heights and angles relative to local are unknown unknowns that cannot be counted, and without the
topography will always vary. The effect of large differences in height denominator of the number of potential image captures, we can‐
and angle has been investigated (Meek et al., 2016; Nichols, Glen, not tell whether a camera captured all, most, or hardly any of the
Garvey, & Ross, 2017; Smith & Coulson, 2012; Swann et al., 2004), animals that passed in front of it. Where camera traps have been
but not the effects of the smaller differences imposed by the prac‐ tested against independent detections such as spoor, continuous
ticalities of fieldwork. There are calls for the fine details of camera video or time‐lapse images, the cameras have been shown to be
placement (height, distance to target, angle, etc.) to be included remarkably poor at capturing medium‐sized carnivores and smaller
in published accounts of camera trap studies (Burton et al., 2015; animals (Damm, Grand, & Barnett, 2010; Glen et al., 2013; Hamel et
Meek et al., 2014). While this allows the set up to be repeated with al., 2013; Lyra‐Jorge et al., 2008; Pirie, Thomas, & Fellowes, 2016),
the same make and model of camera trap, it has limited practical and if camera traps are catching only 30%–50% of potential targets,
value in reproducing the camera traps’ detection performance with then correspondingly more resources are required than with cam‐
other makes and models because the impact of mounting on camera eras that catch everything. If cameras catch only half of the pass‐
performance is known only in very broad terms (Meek et al., 2016; ing animals, they have to be deployed for twice as long to capture
Nichols et al., 2017; Smith & Coulson, 2012; Swann et al., 2004). For sufficient images for analysis, and the longer the trapping goes on,
data to be comparable between studies, we need a way of reproduc‐ the more the assumptions of closure that underpin many models
ing camera performance rather than repeating camera set‐ups, and are likely to be violated. Validating camera performance requires a
that calls for data that directly compare the performance of different known number of opportunities for the cameras to capture images.
brands and models, and the impact on their performance of different Tests in the field suffer from the further disadvantage that they
mounting heights and angles and operating settings. In addition to cannot be rigorously repeated or reproduced; which animals happen
the large differences in height and settings that have been investi‐ to pass through the camera’s detection zone varies with time and
gated so far, we need data on the effect of the small differences that from site to site, even within a given habitat with a given suite of
arise in practice when camera traps are set up in the field. target species. This limits comparisons between cameras to side‐by‐
In addition to a test to validate the performance of camera side shootouts. In the field, data yield rate in terms of images per
traps and the results they generate, there is a need for a stan‐ unit time is low; depending on species, it may take days or weeks
dardised, realistic, quantitative performance test that directly to capture a single image, and even when cameras are left out for
compares camera trap brands and models and detects their weeks to generate useful sample sizes (Hamel et al., 2013; Hughson
strengths and weaknesses. With camera trap prices differing by a et al., 2010; Nichols et al., 2017; O’Connor et al., 2017; Rowcliffe
factor of ten, such a test will save researchers from wasting their et al., 2011), there is no systematic coverage of their fields of view
money on cameras that are not fit for purpose, or spending more or the directions in which animals move. Cameras in the field also
than they need to get the job done. While it is straightforward to capture a large number of images without the targets of interest in
test camera traps by walking in front of them (Seccombe, 2017; them, which wastes time during data processing. A practical test of
see also TrailcamPro https://www.trailcampro.com/pages/trail- camera performance needs to generate useful quantities of data in
camera-shootout), a walking human is a lot taller than nearly all minutes rather than weeks, with systematic, equal coverage of op‐
mammals and has easily detectable areas of warm, bare skin. Some erating variables, and without large numbers of extraneous images.
|
712       APPS and McNUTT

Rovero, Zimmermann, Berzi, and Meek (2013) called for stan‐ At the centre of one end of the arena, a 50 × 50 mm steel square tube
dardised camera trap testing, and there is a need for a test of camera was set vertically into rammed gravel, and sets of up to six test cameras
performance that uses a realistic target, can be reproduced at will, and were mounted using their tripod screw sockets to a 40 × 40 mm angle
quickly generates rigorously comparable data with a known expecta‐ iron bar bolted to the pole either vertically or horizontally. Camera angle
tion of the number of images that will be captured. To realistically sim‐ from the horizontal (angle of tilt and dip) could be varied continuously
ulate the movement of wild animals in the field, the target needs to be by tilting the cameras relative to the horizontal mounting bar (Figure 1b)
at different distances from the camera, moving in different directions or changing the angle between the vertical mounting bar and the pole
with respect to the camera’s direction of aim, and in different places (Figure 1c). Mounting points were at 20 cm vertical increments, with the
within its field of view. Using a trained domestic dog as a real animal exact height of the camera sensors determined by the sensors’ heights
target provides a test that meets these criteria, and that is straightfor‐ relative to the tripod sockets. Heights were measured to 0.5 cm with a
ward, cheap and quick to carry out. Here, how the test is conducted is steel tape, vertical and horizontal were determined with spirit levels, and
described and examples of the performance comparisons and valida‐ angles were measured by a plumb line against a protractor (Figure 1c).
tion data that it generates are given. Full reports on the performance of Cameras faced approximately south, and all tests were carried out in the
each model of camera trap will be published elsewhere. early morning or evening when shade temperatures were below 26°C.
At the edges of the arena, pairs of stakes 50 cm apart and 2.5, 5,
7.5, 10, 12.5 and 15 m from the cameras had thin cotton rope strung
2 |  M E TH O DS between them just off the ground to mark pathways for the target
to walk along. One set of paths ran at right angles to the cameras’
A test arena was marked out on an area of short grass, partly shaded direction of aim (Figure 1d) and the other set diagonally across the
by large trees (Figure 1a). Distances were measured by surveyor’s tape. field of view and down its centre line (Figure 1e).

F I G U R E 1   (a) The camera trap test


arena; image from a Cuddeback C camera
trap mounted 52 cm off the ground,
upright and level (dip and tilt both 0°)
with Toffee the target dog walking left
to right 7.5 m from the camera. (b) Six
Reconyx Hyperfire camera traps tilted at
2° increments on the horizontal mounting
bar. (c) Six Reconyx Hyperfire camera
traps mounted at 20 cm height increments
and dipped 6°. (d) Layout of the tangential
tracks. (e) Layout of the radial tracks
APPS and McNUTT |
      713

A large, mixed breed dog, 68 cm tall at the shoulder (about the of the camera traps. Much of the data are best presented graphically.
same as a leopard, Panthera pardus, or African wild dog, Lycaon pic‐ Because there is insufficient space here for a full set of test results,
tus, two species that we use camera traps to investigate; Figure 1) I will illustrate the utility of the tests using selected examples from
was used for all the tests considered here. He wore a light fabric three brands of camera. Full reports on the performance of each
vest with a matt black spot about 16 cm across on it so that image model of camera trap will be published elsewhere.
processing software could be used to measure his position in the
images (see below). For each test sequence, his handler (PA) left
3.1 | Directional bias
him standing or sitting at the beginning of a track, walked along the
track between the ropes, and turned and called him, he walked at a Controlling the direction in which the target moves generates tests
uniform speed along the track and was rewarded with petting and of whether cameras preferentially detect targets moving in certain
food treats. The handler then walked back along the same track directions. When the target dog walked tangentially across the field
and repeated the recall. Tangential tracks were walked in order of view at right angles to the direction of aim (Figure 1d) of cameras
from closest to farthest from the camera, and radial tracks in a mounted at optimum height, all the camera traps considered here
clockwise pattern starting from the track along the centre, except detected him out to at least 15 m, but the maximum range was re‐
when Cuddebacks were tested when the radial tracks were walked duced along the radial tracks, and when he walked towards the cam‐
counterclockwise. Each series was walked once (Cuddeback C eras along the centre line, he was not detected until he was 7.5 m
range), twice (Bushnell Aggressor, Reconyx Hyperfire dip and tilt from the Bushnells, 5 m from the Cuddebacks and 2.5 m from the
tangential) or three times (Reconyx Hyperfire radial, Cuddeback Hyperfires. This caused a marked reduction in the area of the detec‐
lure simulation and tilt) for each test. The dog is PA’s household tion zone compared to when animals are walking across the field of
pet who has been trained using rewards, and ethics were compli‐ view (Figure 2, Table 1).
ant with research permit 8/36/4 XXXVIII (14) from the Botswana As an example of the results from standard tests being used to
Department of Wildlife and National Parks. predict the performance of different brands and models for spe‐
The results from three camera trap models are presented cific applications, the radial track layout (Figure 1e) simulates ani‐
here: Cuddeback C, Reconyx Hyperfire PC800 and Bushnell mals approaching a scent lure placed 5 m in front of a camera, and
Aggressor (Serial numbers B170511159, B170511162, B170511190, a camera that is suitable for monitoring reactions to the lure will
B170515066, B170515120, B170515137). All the cameras were set capture an image, or start a video while the animal is approaching
to their highest sensitivity and shortest time between triggers and from any direction. Although every unit of each of the three camera
recorded one still image per trigger. They all had fresh batteries that models considered here captured an image for every replicate on
showed at least half charge at the beginning and end of each test. every radial track, only one model, the Bushnell Aggressor, is suit‐
able for monitoring a lure. The other two would not detect animals
approaching the lure in a sector about 45° wide on the opposite side
2.1 | Data analysis
to the camera (Figure 2).
The target dog’s position on radial walks was fixed relative to dis‐
tance markers, the edge of the field of view and the known positions
3.2 | Effect of camera settings
of the tracks. His position on tangential walks was measured from
the X co‐ordinate of his position in each image and the known dis‐ As an example of a test of the effects of changing camera settings,
tance of the tracks in front of the camera. X co‐ordinates were meas‐ Cuddeback Cs have a sliding shutter that blanks off the edges of the
ured by locating the spot on his jacket with the open‐source free movement detector with the aim of having the target centred on the
software ImageJ (https://imagej.nih.gov/ij/). The colour images from images. Tests were run in triplicate with the slider up or down, and
the cameras were converted to 8‐bit monochrome and then to black an image was captured for every replicate on every track in both
and white using the threshold function, and the spot was located directions. Due to the short trigger delay, the first image was usually
using the roundness and area criteria of the Analyze Particles com‐ of the dog’s head as he entered the field of view, independent of the
mand. Since each camera model’s images have their own idiosyncra‐ position of the mask, except when he was moving from right to left
sies, the threshold and particle analysis settings were adjusted to at 10 or 12.5 m from the camera (Figure 3), when raising the mask
suit in each case. X‐coordinates in pixels were calibrated to distances captured his whole body on the first images. Overall, masking the
on the ground by measuring the distance between the edges of the edges of the detector increased the number of full‐body images of
field of view at the 15‐m track. the target at 5 and 7.5 m from the camera, but not at other distances
(Figure 3).

3 |  R E S U LT S
3.3 | Mounting height and angle
Each set of tests generates a large body of detailed, precise data that The effects on image capture of mounting height and the angles at
capture the performance parameters, deficiencies and idiosyncrasies which cameras were leaned forward or to the side were tested. Dip
|
714       APPS and McNUTT

–6 m at which an animal (but not a human) target was detected, and in‐
creased dip decreased both minimum and maximum range for an
–4 m animal and maximum range for a human. The range interval over
which images could be captured got narrower, and the detection
–2 m area shrank proportionally, as cameras were mounted higher and
cuddeback dipped more, and there was a restricted subset of heights and an‐
gles that allowed detections over the full range of distances (Table 2;
Supporting Information Tables S1–S3).
2m Tilt is the angle at which a camera leans to the side (Figure 1b).
It did not affect range, but with different cameras, it has markedly
4m different effects on where the target is when images are captured.
Cuddeback Cs are robust to tilt up to at least 10°. Tilted Bushnell
6m
Aggressors capture images only after the target has penetrated
–6 m deeper into the field of view on the side away from which the camera
was tilted (corresponding to an animal walking uphill past an upright
–4 m camera). Reconyx Hyperfires are particularly sensitive; when they
are tilted, 10° images are captured only after the target has crossed
–2 m the centre line (Figure 4).
hyperfire

3.4 | The importance of realistic targets


2m
A human is bigger, taller and narrower than a leopard or its sur‐
rogate, and much easier for a camera trap to detect; and a human
4m
target generated misleadingly optimistic results for camera per‐
formance. For example, a Reconyx Hyperfire mounted straight and
6m
level with its sensor 1.2 m above the ground detected a human out
–6 m to 15 m, but the dog only at 5 m (Supporting Information Table
S2) even though the whole arena from 5 to 15 m and beyond was
–4 m in the field of view (Apps, 2018). At short range, cameras that are
mounted higher missed the animal target but detected the human.
–2 m Only a real animal target revealed dead zones where a leopard‐
bushnell sized target could walk across the field of view in both directions
without being detected, even though he was within the maximum
detection range and a human was detected every time. For exam‐
2m ple, Reconyx Hyperfires mounted upright at 80 cm had a dead zone
with no dog detections 7.5–10 m in front of them, even though the
4m
dog was detected moving in both directions at both 5 and 12.5 m
(Table 3), and at 100 cm mounted upright, and 120 cm dipped 2°,
6m
the dead zone was at 7.5 m (Supporting Information Table S2).
2.5 m 5m 7.5 m 10 m 12.5 m 15 m

F I G U R E 2   The locations at which images are captured depend


on the direction that an animal is moving relative to the camera, and 4 | D I S CU S S I O N
this affects the size of the area in which targets are recorded. The
pale grey triangle is the field of view, the area between the heavy
The camera trap test proposed here meets the need for a realistic
dotted lines is where a target moving tangentially was captured, the
area inside the heavy solid line is where the target moving radially simulation of field conditions by having a real animal as a target
was captured (see also Table 1), and the darker grey triangle is the moving in different directions at different distances from cameras
sector where a target approaching a lure 5 m in front of the camera mounted at increments of height and angle that might be imposed
would not be detected by field conditions. The track layout and camera mounting can
be reproduced at will and quickly generate rigorously compara‐
ble data with a known expectation of the number of images that
is the angle at which cameras are tipped forward to bring the fields of should be captured. In contrast with Meek et al.’s (2015) statement
view and detection lower and closer to the camera (Figure 1c, Apps that “Studies to generate (camera trap) calibration data are time
& McNutt, 2018). Increased height increased the closest distance consuming, expensive and difficult,” this test takes about 15 min
APPS and McNUTT |
      715

TA B L E 1   The areas covered by each


Areas (m2)
camera trap's field of view out to 15 m,
and the areas of the part of the field of Field of view Moving tangentially Moving radially
view in which targets moving tangentially
Cuddeback 90 59 15
and radially relative to the camera will be
captured as images Hyperfire 90 55 6
Bushnell 87 65 22

to run a sequence, and 30 min for a set of duplicates. During each (P. J. Apps, unpublished data), and there is no particular reason to
15‐min test replicate, the target dog walks across the field of view limit the targets to dogs; opportunistic trials during the dog tests
twelve times—equivalent to weeks or months of data from the showed that food treats will induce domestic cats to walk across
wild (e.g. Di Bitetti Paviolo, & Angelo, 2014; O’Connor et al., 2017) the field of view.
and data are collected equally for different distances, directions, Camera traps that are used to record behaviour (e.g. Vogt,
heights and angles. Having controlled, known target movements Zimmermann, Kölliker, & Breitenmoser, 2014; Allen, Gunther, &
is critical to finding detection dead zones; there is no way of find‐ Wilmers, 2017; Apps et al., 2017) must arguably work better, and
ing them under field conditions because the most parsimonious be set up more carefully, than those for population estimates,
explanation for a lack of images in a particular part of the field of because one missed record of, for example, an animal scent mark‐
view is that no animals moved through it. Precise and repeatable ing can confound the interpretation of later records of behaviour
aiming of the cameras is assured by mounting them all on the same at the site, and there is no prospect of using modelling on large
bars, which are bolted together; once the main pole has been ac‐ data sets to fill in the gaps. Behaviour will nearly always be re‐
curately positioned, there is no need for time‐consuming aiming of corded on video rather than by still images, and although using
individual cameras. still image trigger delay as a surrogate for video start time is con‐
Obviously, the camera traps’ performance with wild animals will venient for screening tests, video trigger delays are always longer
not be exactly the same as with dogs, but they will be a lot more than for stills and rigorous validation of video performance must
similar than with a human target. In particular, detection dead zones obviously be run with videos. For example, the Cuddeback Cs
for animals are not likely to be detected with a human as a target have very fast still image triggers but their video trigger delays
(Table 3). To participate, a dog needs only the most basic of pet obe‐ of 3–4 s are much too slow to be useful. To maximise the length
dience in the form of a recall from a wait. Dogs of different sizes can of time that the target is in shot, it needs to be detected as soon
stand in for different wild species, possibly with the tracks closer as possible after it enters the field of view, and the effects of tilt
to the cameras for small breeds and species. A variety of camera on penetration into the field of view before the camera triggers
trapping scenarios, such as monitoring responses to roadside re‐ (Figure 4) show that this needs cameras to be set up perpendicu‐
pellents, can be simulated by laying out the tracks appropriately lar to the ground surface. If the aim is to video animals reacting to

F I G U R E 3   Locations of the target


when the first image on each traverse
was captured by a Cuddeback C camera
trap with the detector mask raised
(circles) or lowered (triangles), and the
number of images showing the whole
target with the detector window slider
raised or lowered at each distance
|
716       APPS and McNUTT

TA B L E 2   The ranges of distances (m)


Dip
over which camera traps mounted at
Height (cm) 0° 2° 4° 6° 8° 10° different heights and angles captured
images of a leopard‐sized dog
Cuddeback C
36 2.5–15 2.5–15 2.5–12.5 2.5–2.5 No images No images
56 2.5–15 2.5–15 2.5–15 2.5–10 2.5–2.5 2.5–2.5
76 5–15 5–15 2.5–15 2.5–10 2.5–5 2.5–2.5
96 7.5–15 5–15 5–15 2.5–12.5 2.5–7.5 2.5–5
116 10–15 7.5–15 5–15 5–12.5 5–10 5–7.5
Hyperfire
20 2.5–15 2.5 2.5 No images No images No images
40 2.5–15 2.5–15 2.5 2.5 2.5 No images
60 2.5–15 2.5–15 2.5–15 2.5–5 2.5 2.5
80 2.5–15 2.5–15 2.5–15 2.5–10 2.5–5 2.5–5
100 5–15 5–15 5–12.5 2.5–5 2.5–5 5
120 5–15 5–15 5–10 5–10 2.5–5 2.5–5
Bushnell aggressor
22 2.5–15 2.5– 2.5–5 2.5 2.5 No images
12.5
42 2.5–15 2.5–15 2.5–10 2.5–5 2.5 2.5
62 2.5–15 2.5–15 2.5–12.5 2.5–12.5 2.5–10 2.5–5
82 2.5–15 2.5–15 2.5–15 2.5–10 2.5–7.5 2.5–5
102 5–15 2.5–15 2.5–12.5 2.5–12.5 2.5–10 2.5–7.5
122 5–15 5–15 5–15 5–12.5 2.5–10 2.5–7.5

Note. Cells with bold type have detections over the full range.

lure or repellent placed close to the camera, a camera that does the Ultrafires (Supporting Information Table S2), and 98.9% for the
not detect animals moving towards or away from it (Figure 2) is Bushnells (Supporting Information Table S3). With the Cuddebacks
unfit for purpose, and the tests with radial tracks can quickly (Supporting Information Table S1), the points from replicates over‐
eliminate certain camera trap models from consideration. lapped on plots of target position (Figures 3 and 4), allowing the
Comparisons between results are valid only if the test is re‐ number of replicates to be reduced from three to two or one. To re‐
producible and repeatable. Comparing single units of different produce the arena layout and camera mounting is a straightforward
brands and models in a side‐by‐side shootout is straightforward application of measurements of lengths and angles, but the repro‐
(Apps, 2018) but needs the cameras to be available at the same ducibility between different dogs needs to be investigated if the test
time. Comparisons on the basis of results generated at different is to be transferable.
times depend on the test being reproducible, and the three mod‐ Because missed images were recognised as such, and target dis‐
els considered here were tested at different times over a period of tances and camera mounting were varied systematically, the opti‐
9 months. The Cuddeback and Bushnell cameras are what Newey mum height and angle for maximum detections were very quickly
et al. (2015) refer to as “recreational” in terms of technical specifi‐ and clearly recognised. The results do not fully support the stan‐
cations (and price), and the Hyperfire PC800 is designated as a re‐ dard advice to mount camera traps at the target’s shoulder height.
search camera trap. Clearly, from the results presented here, these The Cuddeback achieved its full range of detection distances only
categories do not correspond to performance. when it was 20 cm below shoulder height, and although the Reconyx
Testing cameras in sets of six provide strict repeatability in and the Bushnell achieved detections over the full range at shoul‐
that each of them has exactly the same opportunities to capture der height, mounting them 20 cm lower provided the same range
images; the only difference between them is the parameter under of detection, and with a 20 cm buffer above and below. A Reconyx
test: height, dip or tilt. The most likely cause of differences in re‐ Hyperfire mounted only 12 cm above shoulder height had a de‐
sults between replicates would have been the target dog walking at tection dead zone 7.5–10 m in front of the camera. Differences of
different speeds, but in practice, his speed was so consistent that 20 cm or less in height, and 2° in dip, which are well within the range
detections on tangential replicates agreed in 95.5% of walks for of differences that might be imposed by uneven terrain or when
APPS and McNUTT |
      717

–6
cuddeback
–4

distance from centre (m)


–2

6
0 2.5 5 7.5 10 12.5 15
distance from camera (m)

–6
hyperfire
–4
distance from centre (m)

–2

6
0 2.5 5 7.5 10 12.5 15
distance in front of camera (m)

–6
bushnell
–4
distance from cetre (m)

F I G U R E 4   Shifts in target position –2


when the first image on each traverse
was captured, due to tilting camera traps 0
to the side. Circles are for the target
moving left to right (top to bottom in this 2

orientation) and squares for the target


4
moving right to left. Filled symbols are
captures with the camera upright, empty
6
symbols with the camera tilted 10° to the
0 2.5 5 7.5 10 12.5 15
side. Cameras are mounted at the centre
of the left‐hand edge distance in front of camera (m)

mounting cameras on awkwardly shaped or angled trees, for exam‐ mounting cameras high enough to be out of the reach of animals
ple, have measurable effects on performance that may jeopardise will compromise detections for all except the tallest species, unless
the comparability of data between cameras. Occupancy models can the cameras are dipped, and then, maximum range is compromised.
accommodate camera failure, but they require the detection at one The effects of height and dip are to be expected given the geometry
site to be extrapolated to others (MacKenzie et al., 2002), and these of the camera traps’ passive infrared detectors’ fields of view, but
results show that this requirement may not be met unless camera the differences in results between models, especially the minimum
mounting is more repeatable than it is reported to be. In addition to range for the Cuddeback Cs and the dead zones for Hyperfires, point
copying the variation in camera angles imposed by mounting them to differences in detector field of view that need to be taken into ac‐
on natural objects, dip and tilt also simulate cameras aimed up or count when setting up cameras; the bottom edge of the Cuddebacks’
across a slope, and the results show that for maximum detections detection zone is well above the bottom edge of the image, making
and comparable data, cameras must be angled to match the slope them more vulnerable to being mounted too high or dipped too little
(see figure 5 in Apps & McNutt, 2018). The results show clearly that and there is a gap between the two detection bands of a Hyperfire.
|
718       APPS and McNUTT

TA B L E 3   Number of images of a
Replicate 1 Replicate 2
human and a dog captured on each track
Track (m in L‐R R‐L L‐R R‐L in each direction by a Reconyx Hyperfire
front of camera trap mounted upright 80 cm off
camera) Human Dog Human Dog Human Dog Human Dog the ground
2.5 1 2 1 1 1 1 1 1
5 2 2 1 2 1 1 1 2
7.5 2 0 2 0 2 0 2 0
10 2 1 2 0 2 0 2 0
12.5 3 2 2 2 3 2 3 1
15 3 2 3 2 3 2 3 1

For all three models, their sensitivity to height and dip could make The flat, level, grassed test arena and the cool temperatures
camera trap detections selective for target height, which might lead under which the tests were run provided optimum conditions for
to spurious negative correlations in species occurrence. Mismatches camera trapping, but obviously do not cover all of the conditions
between the detector and image fields of view, and detection dead such as forest, bare sand, rock or snow where camera traps might be
zones within the field of view (Tables 2 and 3), which can be de‐ deployed. The test is straightforward enough to be adapted easily
tected only with known animal movements, mean that even though to different conditions, and different‐sized dogs could be used as
the whole of a vehicle track, for example, may be in the field of view, surrogates for different target species to validate the cameras and
only one side of it will be in a detection zone. Although missed de‐ set up for particular projects.
tections can be accommodated by occupancy models that are robust
to camera failure (MacKenzie et al., 2002), an improperly mounted
camera will create a gap in a capture‐recapture grid, violating the as‐ 5 | CO N C LU S I O N S
sumption that all individuals have a chance of being recorded (Royle,
Nichols, Karanth, & Gopalaswamy, 2009; Tobler & Powell, 2013). A Rigorously establishing the performance parameters of camera traps
camera that misses its only chance to photograph a rare animal will by standard, controlled, validated tests like those described here will
fail to record the presence of that species, which might influence the yield several advantages. The suitability of particular makes and
conservation status of an area, or lead to a failure to recognise the models for particular tasks can be quickly established without using
new or continued presence of an invasive species. up valuable field time or wasting funds on unsuitable cameras. If the
The controlled tests detected a bias against animals that are mov‐ results and conclusions that are derived from camera trap data are
ing radially with respect to the cameras that violates the requirement to be robust and reliable, data models must be robust to the vagar‐
of random encounter models that detections are random with re‐ ies of practical camera trapping, and these tests show aspects of
spect to animal movement (Lucas, Moorcroft, Freeman, Rowcliffe, & camera traps’ performance that models need to be robust to, and
Jones, 2015; Rowcliffe, Field, Turvey, & Carbone, 2008). In addition, how cameras need to be operated in order to minimise the extent
these controlled tests have shown that the area over which camera to which unavoidable assumptions are violated. The test results
traps detect targets and capture their images, which is a parame‐ provide empirical values for model parameters that are currently
ter in random encounter estimates (Cusack, Swanson, et al., 2015), estimated by further modelling, and inform best practice in cam‐
varies by a factor of up to 15 depending on which direction the ani‐ era choice and camera set‐up. Empirical test results can be added
mals are moving relative to the camera. Results from controlled tests to the study design and data modelling that are the current focus
with targets moving in different directions will allow more accurate of critiques of camera trapping methods (Tobler & Powell, 2013).
estimates of detection area than assumptions based on the angle Implementing tests like those described here as a routine compo‐
of the field of view or detections with targets moving in only one nent of camera trapping programmes will enhance the quantity and
direction. The smaller detection area for animals that are not walking quality of data, the robustness and reliability of results, and the va‐
tangentially across the field of view is very likely to contribute to the lidity of conclusions.
lower image capture rates of randomly placed cameras compared
to those on game trails (e.g. Di Bitett et al., 2014; Cusack, Dickman,
AC K N OW L E D G E M E N T S
et al., 2015), and when cameras are aimed across roads, they will
preferentially capture images of animals walking along them rather I am very grateful for loans of cameras from Tau Consultants
than across them (Di Bitetti et al., 2014). If detections are biased by (Christiaan Winterbach and Rob Thomson), WildCru (Dominik
direction of approach, then data models based on random camera Bauer) and Cheetah Conservation Botswana (Leanne Van
siting also require the cameras’ directions of aim to be randomly or der Weyde). The research was carried out under Botswana’s
uniformly distributed. Department of Wildlife and National Parks, Ministry of
APPS and McNUTT |
      719

Environment, Natural Resources, Conservation and Tourism re‐ validity of ecological inferences. Methods in Ecology and Evolution,
search permit number 8/36/4 XXXVIII (14) for which I thank the 4(2), 105–113. https://doi.org/10.1111/j.2041-210x.2012.00262.x
Hughson, D. L., Darby, N. W., & Dungan, J. D. (2010). Comparison of mo‐
government of Botswana. The Botswana Predator Conservation
tion‐activated cameras for wildlife investigations. California Fish and
Trust is funded through Wild Entrust International, Tusk Trust, and Game, 96(2), 101–109.
private donors. Lucas, T. C., Moorcroft, E. A., Freeman, R., Rowcliffe, J. M., & Jones, K. E.
(2015). A generalised random encounter model for estimating animal
density with remote sensor data. Methods in Ecology and Evolution, 6,
ORCID 500–509. https://doi.org/10.1111/2041-210X.12346
Lyra‐Jorge, M. C., Ciocheti, G., Pivello, V. R., & Meirelles, S. T. (2008).
Peter Apps  http://orcid.org/0000-0002-0129-9468 Comparing methods for sampling large‐and medium‐sized mammals:
Camera traps and track plots. European Journal of Wildlife Research,
54(4), 739. https://doi.org/10.1007/s10344-008-0205-8
REFERENCES MacKenzie, D. I., Nichols, J. D., Lachman, G. B., Droege, S.,
Andrew Royle, J., & Langtimm, C. A. (2002). Estimating
Allen, M. L., Gunther, M. S., & Wilmers, C. C. (2017). The scent of site occupancy rates when detection probabilities are
your enemy is my friend? The acquisition of large carnivore scent less than one. Ecology, 83, 2248–2255. https://doi.
by a smaller carnivore. Journal of Ethology, 35, 13–19. https://doi. org/10.1890/0012-9658(2002)083[2248:ESORWD]2.0.CO;2
org/10.1007/s10164-016-0492-6 Meek, P. D., Ballard, G., Claridge, A., Kays, R., Moseby, K., O’Brien, T., …
Apps, P. J. (2018). Why you need a real animal target to test camera traps. Townsend, S. (2014). Recommended guiding principles for reporting
Retrieved from https://www.researchgate.net/project/A-realistic- on camera trapping research. Biodiversity and Conservation, 23(9),
reproducible-and-rigorous-test-for-wildlife-camera-trap-perfor‐ 2321–2343. https://doi.org/10.1007/s10531-014-0712-8
mance/ update/5a776e65b53d2f0bba4fd589 (accessed on 2 May Meek, P. D., Ballard, G. A., & Falzon, G. (2016). The higher you go the less
2018). you will know: Placing camera traps high to avoid theft will affect
Apps, P. J., & McNutt, J. W. (2018). How camera traps work and how to detection. Remote Sensing in Ecology and Conservation, 2, 204–211.
work them. African Journal of Ecology, 56(4), 710-720. https://doi.org/10.1002/rse2.28
Apps, P., Claase, M., Yexley, B., & McNutt, J. W. (2017). Interspecific re‐ Meek, P. D., Ballard, G. A., & Fleming, P. J. (2015). The pitfalls of wildlife
sponses of wild African carnivores to odour of 3‐mercapto‐3‐methyl‐ camera trapping as a survey tool in Australia. Australian Mammalogy,
butanol, a component of wildcat and leopard urine. Journal of Ethology, 37, 13–22. https://doi.org/10.1071/AM14023
35(2), 153–159. https://doi.org/10.1007/s10164-016-0503-7 Meek, P. D., & Pittet, A. (2012). User‐based design specifications for the
Ballard, G., Meek, P. D., Doak, S., Fleming, P. J., & Sparkes, J. (2014). ultimate camera trap for wildlife research. Wildlife Research, 39(8),
Camera traps, sand plots and known events: what do camera traps 649–660. https://doi.org/10.1071/WR12138
miss?. In P. Meek, P. Fleming, G. Ballard, P. Banks, A. Claridge, J. Newey, S., Davidson, P., Nazir, S., Fairhurst, G., Verdicchio, F., Irvine, R.
Sanderson & D. Swann (Eds.), Camera trapping for wildlife man‐ J., & van der Wal, R. (2015). Limitations of recreational camera traps
agement and research (pp. 189–202). Collingwood, Vic., Australia: for wildlife management and conservation research: A practitioner’s
CSIRO Publishing, Australasian Wildlife Management Society, Royal perspective. Ambio, 44(4), 624–635. https://doi.org/10.1007/
Zoological Society of NSW. s13280-015-0713-1
Burton, A. C., Neilson, E., Moreira, D., Ladle, A., Steenweg, R., Fisher, J. T., … Nichols, M., Glen, A. S., Garvey, P., & Ross, J. (2017). A comparison of
Boutin, S. (2015). Wildlife camera trapping: A review and recommen‐ horizontal versus vertical camera placement to detect feral cats and
dations for linking surveys to ecological processes. Journal of Applied mustelids. New Zealand Journal of Ecology, 41, 145–150. https://doi.
Ecology, 52, 675–685. https://doi.org/10.1111/1365-2664.12432 org/10.20417/nzjecol.41.11
Cusack, J. J., Dickman, A. J., Rowcliffe, J. M., Carbone, C., Macdonald, D. O’Brien, T. G. (2011). Abundance, density and relative abundance: a con‐
W., & Coulson, T. (2015). Random versus game trail‐based camera ceptual framework. In A. F. O’Connell, J. D. Nichols & K. U. Karanth
trap placement strategy for monitoring terrestrial mammal commu‐ (Eds.), Camera traps in animal ecology (pp. 71–96). Tokya, Japan:
nities. PLoS ONE, 10(5), e0126373. https://doi.org/10.1371/journal. Springer.
pone.0126373 O’Connor, K. M., Nathan, L. R., Liberati, M. R., Tingley, M. W., Vokoun,
Cusack, J. J., Swanson, A., Coulson, T., Packer, C., Carbone, C., Dickman, J. C., & Rittenhouse, T. A. (2017). Camera trap arrays improve detec‐
A. J., … Rowcliffe, J. M. (2015). Applying a random encounter model tion probability of wildlife: Investigating study design considerations
to estimate lion density from camera traps in Serengeti National using an empirical dataset. PLoS ONE, 12(4), e0175684. https://doi.
Park, Tanzania. The Journal of Wildlife Management, 79, 1014–1021. org/10.1371/journal.pone.0175684
https://doi.org/10.1002/jwmg.902 Pease, B. S., Nielsen, C. K., & Holzmueller, E. J. (2016). Single‐camera trap
Damm, P. E., Grand, J. B., & Barnett, S. W. (2010). Variation in detection survey designs miss detections: Impacts on estimates of occupancy
among passive infrared triggered‐cameras used in wildlife research (Vol. and community metrics. PLoS ONE, 11(11), e0166689. https://doi.
64, pp. 125–130). In Proceedings of the Annual Conference of the org/10.1371/journal.pone.0166689
Southeastern Association of Fish and Wildlife Agencies. Pirie, T. J., Thomas, R. L., & Fellowes, M. D. (2016). Limitations to re‐
Di Bitetti, M. S., Paviolo, A., & De Angelo, C. (2014). Camera trap cording larger mammalian predators in savannah using camera traps
photographic rates on roads vs. off roads: Location does matter. and spoor. Wildlife Biology, 22(1), 13–21. https://doi.org/10.2981/
Mastozoología Neotropical, 21, 37–46. wlb.00129
Glen, A. S., Cockburn, S., Nichols, M., Ekanayake, J., & Warburton, B. Rovero, F., Zimmermann, F., Berzi, D., & Meek, P. (2013). “Which camera
(2013). Optimising camera traps for monitoring small mammals. PLoS trap type and how many do I need?" A review of camera features and
ONE, 8(6), e67940. https://doi.org/10.1371/journal.pone.0067940 study designs for a range of wildlife research applications. Hystrix,
Hamel, S., Killengreen, S. T., Henden, J. A., Eide, N. E., Roed‐Eriksen, L., 24(2), 148–156.
Ims, R. A., & Yoccoz, N. G. (2013). Towards good practice guidance Rowcliffe, J. M., Carbone, C., Jansen, P. A., Kays, R., & Kranstauber,
in using camera‐traps in ecology: Influence of sampling design on B. (2011). Quantifying the sensitivity of camera traps: An adapted
|
720       APPS and McNUTT

distance sampling approach. Methods in Ecology and Evolution, 2(5), for future studies. Biological Conservation, 159, 109–118. https://doi.
464–476. https://doi.org/10.1111/j.2041-210X.2011.00094.x org/10.1016/j.biocon.2012.12.009
Rowcliffe, J., Field, J., Turvey, S., & Carbone, C. (2008). Estimating an‐ Vogt, K., Zimmermann, F., Kölliker, M., & Breitenmoser, U. (2014). Scent‐
imal density using camera traps without the need for individual marking behaviour and social dynamics in a wild population of
recognition. Journal of Applied Ecology, 45, 1228–1236. https://doi. Eurasian lynx Lynx lynx. Behavioural Processes, 106, 98–106. https://
org/10.1111/j.1365-2664.2008.01473.x doi.org/10.1016/j.beproc.2014.04.017
Royle, J. A., Nichols, J. D., Karanth, K. U., & Gopalaswamy, A. M. Wearn, O. R., & Glover‐Kapfer, P. (2017). Camera trapping for conserva‐
(2009). A hierarchical model for estimating density in camera‐trap tion: A guide to best practices. WWF Conservation Technology Series
studies. Journal of Applied Ecology, 46(1), 118–127. https://doi. 1(1). Woking, UK: WWF‐UK.
org/10.1111/j.1365-2664.2008.01578.x Wellington, K., Bottom, C., Merrill, C., & Litvaitis, J. A. (2014). Identifying
Seccombe, S. (2017). ZSL trial camera comparison testing. London, UK: performance differences among trail cameras used to monitor for‐
Zoological Society, Conservation Technology Unit. est mammals. Wildlife Society Bulletin, 38(3), 634–638. https://doi.
Smith, J. K., & Coulson, G. (2012). A comparison of vertical and hori‐ org/10.1002/wsb.425
zontal camera trap orientations for detection of potoroos and bandi‐
coots. Australian Mammalogy, 34, 196–201. https://doi.org/10.1071/
AM11034
Swann, D. E., Hass, C. C., Dalton, D. C., & Wolf, S. A. (2004). Infrared‐ S U P P O R T I N G I N FO R M AT I O N
triggered cameras for detecting wildlife: An evaluation and review.
Wildlife Society Bulletin, 32, 357–365. https://doi.org/10.2193/0091- Additional supporting information may be found online in the
7648(2004)32[357:ICFDWA]2.0.CO;2 Supporting Information section at the end of the article.   
Swanson, A., Kosmala, M., Lintott, C., Simpson, R., Smith, A., & Packer, C.
(2015). Snapshot Serengeti, high‐frequency annotated camera trap
images of 40 mammalian species in an African savanna. Scientific
How to cite this article: Apps P, McNutt JW. Are camera
Data, 2, 150026. https://doi.org/10.1038/sdata.2015.26
TEAM Network (2011). Terrestrial Vertebrate Protocol Implementation traps fit for purpose? A rigorous, reproducible and realistic
Manual, v. 3.1. Arlington, VA: Tropical Ecology, Assessment and test of camera trap performance. Afr J Ecol. 2018;56:710–
Monitoring Network, Center for Applied Biodiversity Science, 720. https://doi.org/10.1111/aje.12573
Conservation International.
Tobler, M. W., & Powell, G. V. (2013). Estimating jaguar densities with
camera traps: Problems with current designs and recommendations

You might also like