A low‐cost, long‐running, open‐source stereo camera for tracking aquatic species and their behaviours

Videos are being increasingly used to record animal behaviour and species interactions in natural environments over extended periods of time (Dell et al., 2014; Smith & Pinter-Wollman, 2021). These approaches have been facilitated by advancements camera technology, allowing ecologists capture the abundance, diversity remote field locations for multiple hours, if not days (Burton 2015). In addition, advances automated image-based tracking methods, will result large numbers videos autonomously analysed quantify presence, biomass. Moreover, individuals' body parts, positions postures (Couzin Heins, 2022; Graving 2019; Karashchuk 2021) can also allow individual behavioural motifs movements be identified. some cases, however, quantifying from video data requires three-dimensional (3D) positional data. captured cameras using stereophotogrammetry, inferences about size, orientation position 3D space recorded (e.g. Straw 2011). From these data, animals' Satterfield 2023; Vivancos Closs, 2015), behaviours Ballesta Janisch 2021), physiology Schiettekatte 2022) inferred. Therefore, collecting has broad applications fields behaviour, applied conservation ecology. While capturing long-term animals their with a high-resolution high-sampling frequency is more straightforward terrestrial laboratory settings 2021; Letessier 2013; Wang Zhu Weng, 2007), equivalent aquatic studies challenging, particularly where constant power cannot supplied. Underwater stereo-video fish biomass Boldt 2018; Langlois 2020) feeding behaviour: Lester 2015, swimming Williams 2010, avoidance Watson Harvey, 2007). However, stereo collected often limited ~1.5 h continuous footage owing many operating short battery lives. To increase likelihood, decrease effort of, rarer full repertoires, should ideally run continuously hours across days. underwater longer lives 10 plus) developed, two-dimensional which suitable tracking, image/video sequences predetermined timepoints one photo every 15 min: Bilodeau 2022, twice daily photos: Greene 2020), or at low frame rates frames per second, Mouy 2020). There need, therefore, system that multi-day, situ, stereo-footage. By filming 3D, we infer locations, orientations, movement trajectories space. could then through pose estimation, e.g. Mathis Mathis, 2020 and/or inferring like attraction/aggression via relative fine-scale movement-trajectories individuals, Nathan 2022). Such yet widely utilised field-based systems but hold great potential Here, present complete hardware software designs an open-source camera, behaviours/interactions high resolution days, costing £1337. We provide step-by-step details design, its running flat-packed, travel-friendly, rig design. guide on how deploy calibrate cameras, illustrate 2D situ estimation individuals (using software). between species. tested equipment Galapagos reefs, deployed up 2 providing in-depth including concerning weighting, deployment focusing, this facilitate use research. Each recording device (hereafter camera) (two total rig) consisted five main components: Raspberry Pi computer, real clock, external hard drive, USB portable bank (Figure 1a; Table 1). An optional microphone added audio (Table 1; Figure 1a). Build guides control codes found online (Dunkley 2023). take ~3 assemble. The modular nature design components), means parts replaced cost 1) breakages occur. For all except (which soldering), they function plugging them into computer. Following al. (2020), designed our around components. becoming popular tool biological offering range (see Jolles, Cameras were Zero. Compared other models 4B), Zero lower processing meaning it draws current (0.8 W idle versus 4B ~3.4 W, Zwetsloot, 2019), thus less extending time. 1920 × 1080 pixels 30 second (adjustable depending requirements, see code). GoPro Hero 10) now film 4K, p sufficient observations 1c) reduces file sizes videos. Recorded stored drive during 1): 1 TB store ~130 (at fps). Contrasting whose saved Pi's MicroSD card possible here shorter deployments), meant storage did limit length. powered banks trialled two sizes, standard 26,800 mAh/99 Wh larger (max size currently available) 30,000 mAh/150 bank. Standard mAh day deployments (and required charge times use), while deployments. longest deployment, turned 31.5 (18 13.5 mode), packs (30,000 mAh) still had ~50% rigs deployed, batteries: 49%–52% remaining). With ~22 (recording land, users set specific windows code, below). capacity likely increase, further increasing capabilities rig. Power removed charged Velcro strips, although future upgrades include charging cables need this. banks, specialised costly) chargers/housings lithium-ion batteries previously Favaro 2012; 2020; Purser was housed four-inch pressure housing Blue Robotics 1, depth rated 100 m). It however construct housings tubing as lower-cost alternative PVC acrylic Bergshoeff 2017; so fit dome port rather than flat-port (camera mounted centrally dome). Dome ports advantageous flat does restrict camera's view image refraction, steps focussing below) (She 2019). uses lightweight (~12 kg) quickly assembled (<1 h) dismantled, facilitating transport sites/institutions 1b, Dunkley 2023 files). Full ~20 prepare prior first (without outsourcing hole drilling printing). This timing dictated speed printing (hole takes h). centre gravity height 500 mm substrate) withstand strong currents swell. includes adjustable legs, facilitates adjustment position, angle view. adjusted incorporate bait pole baited required. distance angles depends what aim capture, view, far subjects camera. More precise spatial measurements objects obtained when placed apart each (Boutros scene ~15 m max (dependent water visibility, etc), horizontal ~90° vertical ~70°, fixed 700 inward 7°. mounting plates, screwed 1b), maintained attached. Further holes drilled change study requirements. controlled Pi. Control python based require basic coding experience) modified user requirements automatically start, stop (in 30-min date stamped sequential chunks) specified points start 6:30 am, finish 5 pm). started manually ON/OFF switch any point), time, When ON recording, remained mode, minimal would restart Camera installed clock holds ±2 million equating drift 0.17 s (Datasheets.Maximintegrated.Com/En/Ds/DS3231.Pdf) even connected internet underwater). Any accounted flashing light visible both Video quality, resolution, contrast, brightness, white balance) (volume/sensitivity) separately merged post FFMPEG). code number failsafes issue detected crashed lens (by altering lens' sensor). focusing land hyperfocal distance), manual consideration. important housing, creates ‘virtual image’ front ports, focus much closer air focal length diameter dome. information trial-and-error approach links guidance). focused object 17 cm away sensor. measurement 87 3.56 (F2.5). Since designing new developed (Arducam 16MP Module), allows cameras' code. different values use, most optimal Stereo retrieved surface ropes there uncertainty (as placement guaranteed). recommend attempting behaviours, guided one-third benthos, two-thirds pelagic). reefs (March–May 18.8 m, average ~12 n = 32 deployments, ~480 footage, under National Park research permit number: PC-06–22), SCUBA. methods retrieve lowering rope before diver placement, lifting/moving weighted lift bag) efficient, minimalist method involved divers descending/ascending carrying unweighted swim find depend study. If retrieval diving (i.e. limits may reached strong), upon buoy attached rig, pulled hand. opportunistic drops, maximises minimises costs only dive two), carefully organised obstruct changing tides), snag benthos. Rope buoys disturbed passing boats risk theft. always rope. deeper beyond SCUBA depths), feasible, surface. additional weights ensure remain position. presented weigh kg marginally negatively buoyant seawater, sinking slowly. fix place, 12 (6 weights) moderate (~2 knots). weight quantities stronger currents), conditions. independently, offset unsynchronised must calculated reconstruction. calculate temporal pairs: microphones 1a,b). Audio streams cross correlating peaks waveforms (Hasler 2009). A light, preferably synchronise left right images (Harvey 2003) video. 2023), involves pinpointing (for starts calculating difference cameras. flashes calibration stage repeated within account minimum. Before synchronised positioning/tracking, intrinsic properties distortion, length), positioning (extrinsic properties), needs calculating. Many utilise specialist commercial presents cost. softwares available StereoMorph: open-source, reconstructions anatomy, Olsen Westneat, 2015) do work well video, open source Calibration toolbox Matlab, Bouguet, 2004). GoPro) difficult calibrate, significantly refraction. effect refraction modelled. radial distortion component three terms polynomial (Shortis, six extreme improve accuracy. script calibrates incorporating necessary move object, known dimensions checkerboard cuboid), overlapping angles. identified viewpoints software. Whilst cuboid calibrations accurate, stable method, A3 checkerboards low-cost (average proportional error: 1.7% checkerboard, 0.5% cuboid, 11.6% A4 Boutros perform step deploying/retrieving practical involving cuboids. printed laminated, rigid board keep flat. eight row nine columns (with square measuring 33.5 mm, 1d). observed orientations completed ~10 min. repeat slight deviations tube tubes charging, download) alters error. Our produced error 1.1 (median, IQR 1.77) test Once parameters determined (following 2023) extrinsic objects/individuals 1e). achieved triangulation programme SEBASTES (Williams 2016). long environments, fine detailed ecological collection. (30 fps) (1080 p) resolutions. offers researchers opportunity abundance inform (more accurate calibrated organisms. quantification postures. ultimately social aggressive Although map conditions Bala Günel combination field. effects experimental manipulations rarely observed. accompanying post-processing provides source, travel friendly accessible analyse period set-up avoids trade-offs associated gathering sampling presence (including rare species), environment another finer, detail. answer questions, relating actions decisions shape ecosystem function. artificial intelligence autonomous analysis sets facilitated. picture health ecosystems. Katie idea James Herbert-Read wrote manuscript support Herbert-Read. Andrew built hardware. Drewnicki metal rigs. support, supervision mentorship Inti Keith. All authors contributed revisions. funded FSBI Postdoctoral International Travelling Fellowship, Christ's College University Cambridge Islands Fund (both awarded Dunkley). supported Whitten Lectureship Marine Biology, Swedish Research Council Grant (2018-04076). publication contribution 2479 Charles Darwin Foundation Islands. thank hosting Dunkley, fieldwork. granting us authorization carry out investigation (research PC-06–22). Wilson Iniguez, William Bensted-Smith, Emily McFarling, Julio Delgado, Mikel Goñi Franklin Teran help deploying Dr Sam Matchette holders, discussions regarding calibration. Daniel Matterson advice. no conflicts interest. peer review history article https://www.webofscience.com/api/gateway/wos/peer-review/10.1111/2041-210X.14151. files uploaded project dedicated github page: https://github.com/katiedunkley/Underwater-stereo-camera- Open Science Framework