Studying the motion of cheetahs – especially in the wild – is a technically challenging endeavour that pushes the limits of field biomechanics methodology. Consequently, it provides an interesting example of the scientific symbiosis that exists between experimental biology and the technological disciplines that support it. This article uses cheetah motion research as a basis to review the past, present and likely future of field biomechanics. Although the focus is on a specific animal, the methods and challenges discussed are broadly relevant to the study of terrestrial locomotion. We also highlight the external factors contributing to the evolution of this technology, including recent advancements in machine learning, and the influx of interest in cheetah biomechanics from the legged robotics community.

Speed is a valuable asset for any predator, but for the cheetah (Acinonyx jubatus), it is essential for survival. While other cats depend on stealth, cooperation or brute strength to capture prey, cheetahs tend to be solitary hunters that operate in plain sight, relying instead on the ability to chase down any animal (Thompson, 1998).

Researchers have been fascinated by the unparalleled speed and agility of cheetahs for as long as the species has been known to recorded science. The seemingly simple question, ‘how does the cheetah run so fast?’ has drawn contributions from various scientific fields. Anatomists answer in terms of the physiological structures that facilitate the cheetah's speed, such as its light, streamlined body, powerful leg muscles and unretractable claws. Biomechanists extend the answer to the dynamics of the motion and the forces driving it. Ethologists and ecologists consider the interactions with other animals and environment that motivate the motion. Neuromechanists include all of the above, and how they integrate into a feedback system governed by the animal's central nervous system (see Fig. 1). Roboticists answer practically, using cheetah-inspired mechanisms to improve the speed and manoeuvrability of mobile robots.

Fig. 1.

Block diagram representation of cheetah locomotion as a combination of negative feedback loops. Whole-body motion of the animal [P(t), where t is time] is driven by the muscles, under the control of the nervous system [C(t)] in response to various sensory inputs. Data describing various sensory perception processes [S(t)], muscle activation [B(t)] and external factors including prey interactions and contact with the environment are all required for a complete understanding of its locomotion.

Fig. 1.

Block diagram representation of cheetah locomotion as a combination of negative feedback loops. Whole-body motion of the animal [P(t), where t is time] is driven by the muscles, under the control of the nervous system [C(t)] in response to various sensory inputs. Data describing various sensory perception processes [S(t)], muscle activation [B(t)] and external factors including prey interactions and contact with the environment are all required for a complete understanding of its locomotion.

The ability to answer from each of these perspectives is constrained by the data that can be obtained from the animal with existing technology. Tracking the motion of nature's fastest sprinter in its natural habitat is a formidable technical challenge that reveals the limitations of established methods. A review of cheetah motion research is therefore also an account of how field biomechanics methodology has evolved over the relevant decades.

In recent years, this evolution has been accelerated by increased interest in the animal from the legged robotics community, which has drawn more engineering attention towards these technical boundaries (Ijspeert, 2014). The rapid advancement of machine learning has also produced seismic shifts in data processing, vision-based motion tracking and other remote sensing methods (LeCun et al., 2015). This makes cheetah motion research a fascinating case study in the interactions between experimental biology and the emerging technologies commonly heralded as constituents of the Fourth Industrial Revolution (Schwab, 2017).

In this article, we will examine the past, present and future of field biomechanics through the lens of cheetah motion research. Despite our focus on a particular animal, the methods, trends and challenges discussed are broadly relevant to field studies of wildlife. We begin with a brief history of cheetah motion research to set the context for the review. We then establish the current state-of-the-art technologies used to obtain motion, force and muscle activation data in the field, highlighting the factors contributing to their development. Finally, we speculate on how cheetah motion research – and, in turn, these technologies – might progress in the future, given established trends and the field's unsolved problems.

The relationship between cheetah motion research and the technology supporting it is bidirectional: sometimes, technological advancements have motivated studies into new aspects of cheetah motion, while at other times, the challenges of studying this animal in its natural habitat have motivated technological advancements.

Before quantitative descriptions of the cheetah's motion were possible, naturalists attempted to explain its speed through its morphology. As an example, an early description of the features separating the cheetah from other felids by Pocock (1927) postulates contributions that each might make to its motion: ‘All its more obvious external features – its small, light head, narrow chest and body, long, thin, sinewy legs, and powerful hind quarters – are obviously adapted to [running]. So, too, with the feet. The protruding claws, hard, pointed, digital pads, ridged plantar pads, deeply emarginate webs, and wide hind feet are all better fitted for securing a firm hold upon hard or sandy ground and for traversing it swiftly and surely than are the softer more pliable feet of other members of the cat tribe; and I think it is probable that the long, rigid, and sharp carpal pad, when jammed against the soil, aids in arresting the headlong rush when a rapid turn after the pursued quarry is required. The long heavy tail, too, probably acts as a balance in wheeling at full speed’.

The first detailed analysis of the cheetah's gait was made possible by footage of a sprinting cheetah captured for the 1955 documentary film, The African Lion (Walt Disney Productions True-Life Adventures series). Hildebrand (1959) compared this motion to earlier footage of a galloping horse by Muybridge (1882), and subsequently supplemented this study with additional footage of captive cheetahs. This showed that the galloping gait of the cheetah follows a different footfall sequence to the equine gallop. Various mechanisms, including the cat's flexible spine, were posited to explain its speed advantage over the horse, but these theories would require data that could not be obtained precisely from film to confirm.

Hildebrand (1959) also hypothesised that the cheetah's size imposes the upper limit on its endurance, suggesting that ‘its muscles can stand the strain long enough for the animal to run the necessary 400 to 600 yds., so greater efficiency is not needed’. Internal data could give more information on muscle activation and energy expenditure to elucidate this aspect of the animal's locomotion, but such studies were largely confined to the laboratory at this time. Treadmill studies by Taylor and Rowntree (1973), for example, were the first to document body temperature in running cheetahs. This study produced the conclusion that sprint duration in cheetahs was primarily limited by the possibility of overheating, which was later found to be incorrect (Hetem et al., 2013).

With film and direct observation remaining the primary methods of studying cheetah motion for most of the 20th century, even the top speed of the animal was difficult to confirm. Speeds of wild cheetahs had been subjectively estimated in comparison to the speeds of cars, or approximated from triangulated camera positions or calibration grids, but these could not be considered reliable readings (Sharp, 1997). Estimated speeds for wild cheetahs also exceeded those reported in captive animals (Hudson et al., 2012). The first reliably measured confirmation of its status as the fastest terrestrial animal was a timed run by a rewilded cheetah recorded by Sharp (1997), which indicated that it reached 29 m s−1.

Following this, the development of smaller, lighter electronics made it possible to track the motion of the animal using sensors mounted on collars. This allowed the velocity of cheetahs to be tracked over time, which also gave the first measurements of the animal's acceleration. Importantly, these collars were light enough to be tolerated by wild cheetahs, so these measurements could be obtained during hunting. Studying wild cheetahs is necessary to give the truest indication of the animal's maximum performance, as captive cheetahs typically lack the fitness and motivation to match the speeds of their wild counterparts. The resulting landmark study by Wilson et al. (2013a) showed that wild cheetahs rarely achieved speeds matching the fastest recorded during hunting, but instead relied on rapid acceleration to catch prey.

Small dataloggers could also be surgically implanted inside the cheetah's body. Besides movement data, these implants were able to collect more accurate measurements of statistics such as body temperature and heart rate, allowing established hypotheses regarding the animal's endurance to be tested in wild cheetahs. Using this technique, Hetem et al. (2013) disproved the previous theory regarding overheating in running cheetahs.

Improvements in the ‘power-to-weight ratio’ of computers also made the development of legged robots feasible, which increased interest in the mechanics and control of legged locomotion. Naturally, the cheetah became an important inspiration for engineers hoping to improve the speed and manoeuvrability of these robots. In addition to especially direct examples such as the MIT Cheetah series of agile quadrupeds, some robots have imitated specific physiological features of the cheetah, such as flexible spines (Ijspeert, 2014; Tang et al., 2020; Kamimura et al., 2021) to increase stride length, or tails to stabilise the body during high-speed manoeuvres (Briggs et al., 2012; Patel and Braae, 2013, 2014; Patel and Boje, 2015). In this way, roboticists have continued the work of early naturalists by confirming and quantifying the contributions of the cheetah's anatomy to its motion.

The recent interest in the biomechanics of the cheetah from the robotics field has driven a demand for increasingly detailed motion data. Understanding the mechanics and control of its locomotion to an extent that it can be imitated by a robot requires information about the movement of the whole body, rather than just a single point as captured by collar-mounted dataloggers. It also requires information about the forces propelling that motion, and the muscle activity and ground contact interactions producing those forces.

Although the process of obtaining this data can be simplified to some extent by studying captive animals rather than wild ones, doing so remains a technical challenge. Internal or on-animal sensors are invasive methods that should only be applied with careful consideration (Daley and Biewener, 2003), but factors such as dynamic outdoor environments, and low contrast between the cheetah and the surroundings it has evolved to camouflage in tend to confound remote approaches (Joska et al., 2021). Methods that confine the cheetah's motion to a particular area, such as force plates or static camera traps, also become expensive when extended to the large area necessary to ensure unpredictable high-speed manoeuvres are captured (Hudson et al., 2012).

The current state of cheetah motion research and, more generally, field biomechanics, can thus be summarised as the pursuit of the following objectives: (i) increasing the quantity and detail of the data that can be obtained; and (ii) decreasing the invasiveness, and (iii) the cost of the methods used to obtain these data.

The influx of engineering and computer science expertise into this research area due to the crossover with robotics has already motivated technological innovation towards these goals. This progress accelerated dramatically throughout the late 2010s as a consequence general advances in areas such as mobile communication, artificial intelligence and ‘big’ data processing.

Improvements in the detail of the data can be quantified by considering the number of ‘keypoints’ on the animal tracked in different studies. We loosely define a keypoint as any point on the animal's body about which locomotion data (position or force) is obtained. Collars, on-animal sensors and tracking markers or landmarks are regarded as keypoints, as are all four paws in studies using force plates. Fig. 2, which maps the invasiveness of the methods used versus the number of keypoints tracked on the animal for relevant biomechanical studies of quadrupeds since 2000, clearly shows that the majority of recent studies are not yet attaining the goal of remote, whole-body tracking. In this section, we explore recent developments in the field, and attempt to capture a snapshot of the dynamically shifting state of the art.

Fig. 2.

The invasiveness of methods versus keypoints monitored on the animal for recent biomechanical studies of terrestrial quadrupeds. Force plates contributed +4 to the number of key points, as they give data about all feet, while each inertial measurement unit (IMU) or GPS sensor was considered to add a single keypoint. Data from Bragança et al. (2017), Davies et al. (2019), Hudson et al. (2012), Joska et al. (2021), Patel et al. (2017), Wilson et al. (2013a,b, 2018) and Witte et al. (2004, 2006).

Fig. 2.

The invasiveness of methods versus keypoints monitored on the animal for recent biomechanical studies of terrestrial quadrupeds. Force plates contributed +4 to the number of key points, as they give data about all feet, while each inertial measurement unit (IMU) or GPS sensor was considered to add a single keypoint. Data from Bragança et al. (2017), Davies et al. (2019), Hudson et al. (2012), Joska et al. (2021), Patel et al. (2017), Wilson et al. (2013a,b, 2018) and Witte et al. (2004, 2006).

Motion tracking

Two approaches have been used to record the kinetic motion of terrestrial animals: on-animal sensors and remote sensing.

On-animal sensors

Locomotion data is most easily gathered in a laboratory setting, but besides not reflecting the natural locomotor conditions or external motivations experienced by animals in the field, wild animals like the cheetah are not amenable to indoor tests. Rapid manoeuvres such as sudden braking and turning are particularly difficult to capture, as they are typically unplanned, or require a large area to perform. As such, on-animal capture devices have been the preferred method for wildlife studies, effectively enabling an infinite capture area.

The first instances of these were electronic tags (Lord et al., 1962), but these required extensive manual effort to find and record the animal location. In the early 2000s, GPS technology became the most prevalent option for animal tracking, enabling the movement of animals to be observed over a long duration (Kays et al., 2015). GPS sensors on their own lack the temporal resolution (∼1 Hz update) and are generally only accurate to around 5 m (Wilson et al., 2013b). In the early 2010s, it was shown that fusing GPS with inertial measurement units (IMUs) could provide increased position resolution. Tan and Wilson (2011), for example, used this technology to track the motion of racehorses to within 20 cm. Further, these lightweight modules are suitable for use on wild or sensitive subjects, as demonstrated by studies applying them to wild echidna (Clemente et al., 2016) and free-roaming domestic cats (Galea et al., 2021). IMU data have also proved valuable for recording the behaviour of wild animals, as in Tatler et al. (2018), where three-axis accelerometer data were used to identify which of 14 possible activities tracked dingoes were engaged in.

These ideas were extended in a ground-breaking paper by Wilson et al. (2013a), which measured the locomotion dynamics of hunting cheetahs at an unprecedented level using GPS-IMU collars that identified when hunting was taking place. These GPS-IMU collars have since enabled the study of the hunting behaviour of cheetahs in response to prey movement (Wilson et al., 2013b, 2018).

The main drawback of GPS-IMU collars is that the animal is only treated as single point, disregarding the articulation of the body. On-animal cameras provide more insight into the motion of isolated body parts, and were used to study the head motion of falcons during hunting (Kane and Zamani, 2014). This approach was subsequently used on cheetahs by Patel et al. (2017), who developed a harness that utilised a pair of stereo cameras in conjunction with a GPS-IMU system to measure the spine and tail motion of captive cheetahs during high-speed locomotion. The photo in Fig. 3 shows the camera harness on a cheetah with the markers placed on its spine and tail for tracking. Though this study obtained promising results, many of the test cheetahs ultimately rejected the harness or showed decreased interest in running when wearing it, so the method was deemed too disruptive to test on wild cheetahs.

Fig. 3.

Cheetah wearing a stereo camera harness to track its tail and spine. The camera harness was designed by Patel et al. (2017). The photograph shows tracking markers placed on the animal's back and tail.

Fig. 3.

Cheetah wearing a stereo camera harness to track its tail and spine. The camera harness was designed by Patel et al. (2017). The photograph shows tracking markers placed on the animal's back and tail.

Remote sensors

While on-animal sensors are being progressively miniaturised, they still tend to affect animal behaviour – especially when used on sensitive, wild subjects such as the cheetah. It is also difficult to record more detailed joint movements or whole-body locomotion without overloading the animal with sensors. The remote sensing approach has been introduced in response to these challenges. The traditional method is racing traps, such as those used by Williams et al. (2009) to record the speed of greyhounds and horses through a calibrated space.

Initially, the recording of movement within a specific space was assisted with markers attached to the animal. For example, the TurfTrax tracking system used in horse racing achieves a wireless radiolocation tracking system by attaching a radio-transmitter tag to the horse and placing receiving antennas around it (Spence et al., 2008; Williams et al., 2009). When used in conjunction with IMU modules, the system can simultaneously analyse more global data such as location information, and microscopic data such as body inclination. These data have provided extensive insight into the dynamic constraints on turning in horses (Tan and Wilson, 2011). These systems only measure the motion of the tag itself, so their ability to capture details such as the motion of joints is limited. While typically lighter and smaller than on-animal sensors, these markers are also subject to the same invasiveness considerations, which limit their usefulness with wild animals.

For a motion capture system to truly satisfy the demands of current cheetah research, it must be able to record whole-body movement without assistance from sensors or markers on the animal. Before the development of deep learning, markerless position and attitude acquisition systems based on vision sensors had limited application, or were only used to analyse information in two dimensions because of the limitations of hand digitisation (Fry et al., 2003; Tian et al., 2006; Hedrick and Biewener, 2007; Straw et al., 2011). Deep learning significantly improved image processing techniques, however, leading to various frameworks for motion analysis being proposed (Günel et al., 2019; Ray and Stopfer, 2022; Pereira et al., 2022). This technology has facilitated 3D posture analysis using multiple camera systems (Nath et al., 2019; Karashchuk et al., 2021; Joska et al., 2021), which has been further supported by the development of stereo camera traps for 3D multi-object tracking of wildlife (Klasen and Steinhage, 2022).

Force, contact and muscle activation

A complete biomechanical understanding of the cheetah also requires information about the external and internal forces that produce its movement. The external forces are the ground reaction forces (GRF) acting on the animal's feet, so analysis of these forces also necessitates accurate assessment of the foot contact state. The internal forces originate from the animal's musculoskeletal system, and include the contributions of muscles, tendons and ligaments. Altogether, this makes the estimation of both the external and internal forces a particularly difficult endeavour, especially when the additional challenges of high speed and wild environments are considered.

Many developments in this area were initially intended for clinical studies, where gait analysis is used in controlled experiments to diagnose musculoskeletal diseases in animals (Merkens et al., 1993; Andrada et al., 2017). These experiments are confined to a laboratory environment, so the resulting methods are not always suitable for use with animals in the wild. The applicable methods vary widely in invasiveness and the quantity and quality of data they can obtain.

Surface electromyography

Surface electromyography (sEMG) is a direct method that measures muscle activation with electrical probes, and therefore indirectly determines the muscle force. For this reason, careful consideration is needed during calibration and interpretation of the EMG results (Roberts and Gabaldon, 2008). Recently, it has been used for walking gait analysis in olive baboons (Druelle et al., 2021). This requires shaving the animal to place the probes on the skin – a procedure that involves an anaesthetic dose. Consequently, its use in wild animals is limited.

Inertial measurement units

In addition to tracking the motion of the animal, IMUs can be used to infer information about foot–ground contact based on measured accelerations (Witte et al., 2004, 2006, Bragança et al., 2017). This method has been applied in horses to estimate contact timings in unconstrained environments over multiple strides – information that would be difficult and expensive to obtain using force plates. The comparative disadvantage is a reduction in the accuracy and precision of the acquirable data.

Force platforms

Force plates are a less invasive method of obtaining GRF and contact data than either of the previously discussed options, but have the drawback of a small capture area and ambiguous readings in cases where multiple contacts occur within a single plate. Despite this, they have been used extensively to gather GRF data from cheetahs (Hudson et al., 2012), horses (Davies et al., 2019) and other wild terrestrial mammals (Shine et al., 2017; Basu et al., 2019a).

The data obtained from force plates can be used to infer loading on the limb joints using an inverse dynamics optimisation process (Brown et al., 2020). The GRF alone also yields information about important characteristics of the gait biomechanics. Hudson et al. (2012), for example, determined the factors contributing to the acceleration of the cheetah by comparing kinematic and kinetic data captured using a runway of eight force plates against the same data for racing greyhounds. However, the expense, laborious setup process and constraints of the capture area have prevented the use of force plates with wild cheetahs.

Simulation models

In response to the limitations of current force-sensing technologies, computer simulation methods have been proposed as an alternative to estimate these forces. This is also an especially accessible option because of the availability of open-source packages such as OpenSim (Seth et al., 2018). The early ancestors of these simulations are mathematical approximations of forces based on conservation of momentum, such as those performed by Alexander et al. (1979), which were confirmed to be consistent with later force measurements for horse locomotion (Witte et al., 2004).

In the intervening time, vast advances in computing power and anatomical knowledge have made it possible to construct far more versatile and accurate motion models. Data based on previous studies, such as kinematics from motion capture, kinetics from force platforms, and morphology from digital scans, can be used as inputs to build computer models of the subject of interest. The detailed musculoskeletal model used by Hutchinson et al. (2015) to estimate muscle moments and capacity during ostrich locomotion are an example. The model was built using previously measured joint kinematics and GRF, while computed tomography (CT) scans were used to estimate bone volumes. More recently, similar work has been done on dogs (Stark et al., 2021), providing promising results for a quadruped model, but the technique has yet to be extended to the cheetah.

Advances in machine learning have allowed more detailed motion data to be captured from cheetahs more remotely, but these techniques have thus far been demonstrated exclusively on captive cheetahs, with expansion to wild cheetahs constrained by the capture area and lifespan of the supporting camera systems. Similar constraints apply to the established methods of capturing force, and related contact and muscle activity data. In this final section, we highlight emerging technologies that could allow cheetah motion research to truly go wild.

Improved on-animal methods

Despite being more invasive than state-of-the-art remote motion capture methods, on-animal sensors such as GPS-IMU collars will likely continue to have a role in long-term studies of wild cheetahs because of their low power consumption, robustness and relatively low cost. However, later iterations of these sensors should provide greater insight beyond rigid-body motion. Recent information on gait abnormality in humans obtained from on-body sensors shows the potential of the technology to go beyond single-point motion tracking (Han et al., 2019). In a similar vein, researchers have demonstrated automated movement classification using machine learning in cheetahs (Grünewälder et al., 2012) and muscle efficiency estimation for wildebeest (Curtin et al., 2018). Invasiveness will also be less of a factor for future dataloggers, as they are likely to be far smaller and lighter than current versions (Perlmutter and Breit, 2016).

New sensors

New sensing technologies have the potential to produce new insights into animal locomotion. Millimetre-wave radar (mmWave) systems have become lower in cost and have successfully been applied to human skeletal estimation with comparable accuracy to camera-based systems (Sengupta et al., 2020). A mmWave radar is unaffected by lighting conditions and occlusion (Zhao et al., 2018), which could potentially produce more robust in-field animal pose estimation. This is especially important when tracking wild cheetahs, as they often hide themselves in long grass when approaching prey.

LiDAR (sometimes called 3D laser scanning) can acquire 3D point cloud data, and also has the potential to realise long-range motion capture. Currently, the use of LiDAR is limited to measuring the animal's body and observing its general behaviour in the environment (Davies and Asner, 2014; Pezzuolo et al., 2019), but long-range motion capture systems using LiDAR have been proposed and large datasets published in the context of human motion capture (Li et al., 2022). Although both technologies have big potential for animal motion capture, they are currently limited to the measurement of vital signs (Wang et al., 2019; Azmy et al., 2012) and species identification (van Eeden et al., 2018).

Event-based cameras are a biology-inspired camera technology that asynchronously measures per-pixel brightness changes and outputs these as a stream of ‘events’ (Gallego et al., 2020). Compared with conventional cameras, event cameras have extremely high temporal resolution (of the order of microseconds) and, leveraging this, researchers have produced high-speed (1 kHz) human skeletal motion capture (Calabrese et al., 2019). This technology could enable high-speed motion capture of animals in the field.

Sparse inertial sensors have also achieved promising results for 3D motion capture as well as force estimation in humans (Yi et al., 2022 preprint) and gait analysis in canines (Jenkins et al., 2018; Zhang et al., 2022), but they have yet to be adopted for similar use in wild animals.

Moving cameras

Although multi-camera systems are becoming technically mature, their use in wild environments is limited because the area they can observe is limited by the number of cameras available and the locations in which they can be set up. A seemingly obvious way to extend the capture area of a given camera system is to give the cameras the ability to move as done with human motion capture (Von Marcard et al., 2018).

The possibilities of this approach have been drastically expanded by the emergence of unmanned aerial systems (UAS) technology. A low-cost implementation by Basu et al. (2019b), using a commercial remote-controlled camera drone, achieved promising estimates of the speed of running giraffes, while Haagensen et al. (2022) incorporated a similar device into a multi-camera study of rapid turning in dogs that applied a setup otherwise resembling the grounded array of static, commercial-grade cameras used to obtain the footage for AcinoSet (Joska et al., 2021). Although many current products can only acquire single-view camera images, they can be combined with model-based approaches to estimate the dynamics of wildlife (Krause et al., 2017).

Although this is certainly a remote motion capture method, it is not yet a non-invasive one, as visual and auditory interference from similar aerial systems has been shown to provoke elevated alertness and even flight responses in various terrestrial species when they are brought within the range at which usable footage could feasibly be captured (Bennitt et al., 2019). The responses of wild cheetahs to these systems have yet to be formally surveyed, but if a viral video of the animals attacking a remote-controlled quadcopter is any indication (Hey Nadine, ‘CHEETAH vs DRONE // South Africa’: https://www.youtube.com/watch?v=TlRkxQgMJJI), they are likely to be too disruptive to use with cheetahs in their current form.

Monocular motion capture

Another factor limiting the area that current camera systems can cover is the requirement that many cameras overlap the same capture area, to provide the multiple points of view required for 3D reconstruction. Emerging methods of monocular 3D motion capture, which uses a single-camera view of a subject to produce a 3D skeleton, could reduce this problem. This approach has been successfully demonstrated for humans (Mehta et al., 2017) and even extended to laboratory animals (Gosztolai et al., 2021). Monocular video has also been used to estimate the contact forces and body torques of human subjects using trajectory optimisation (Gärtner et al., 2022).

Big data

Many of the emerging technologies in cheetah motion research are data driven, requiring large datasets to successfully train the machine learning algorithms involved. This is especially true when studying wild animals, as wide variation in the conditions and resulting data quality demands a much larger pool of diverse training examples to overcome. By far the biggest difficulty facing field biomechanics is access to large-scale datasets of ‘ground truth’ which modern deep learning is built on (LeCun et al., 2015).

Camera-based human motion capture has leveraged large, open datasets for several years for various pose estimation tasks (Andriluka et al., 2014; Ionescu et al., 2013) and new sensing technologies have followed a similar approach (Calabrese et al., 2019; Sengupta et al., 2020). For domestic or laboratory animals, this has been the approach as well, with several open datasets, including ones on mice (Sun et al., 2022), horses (Mathis et al., 2021), macaques (Labuguen et al., 2021) and Drosophila (Günel et al., 2019). Ground truth data for wildlife remains scarce by comparison, as it is so much more difficult to obtain (Jiang et al., 2022).

There are several promising approaches to circumvent this, however. (i) The combination of 2D markerless pose estimation (Mathis et al., 2018) – training data for which are easier to create – with algorithmic approaches such as Kalman filters and trajectory optimization was used by Joska et al. (2021) to produce AcinoSet – the first 3D motion dataset obtained from captive cheetahs. (ii) The skinned multi-animal linear model (SMAL) method, which leverages 3D scans of toys (Zuffi et al., 2017), has been used to estimate the 3D pose of zebras in the wild Zuffi et al. (2019). (iii) Generative adversarial networks (GANs) are combinations of two competing networks (a generator and a discriminator) that compete to produce robust pose estimates (Mueller et al., 2018). These have yet to be leveraged for wildlife applications. (iv) Synthetic data generation – for example, the use of artificial training images – is also an emerging method that could potentially bridge the data gap for field biomechanics (Tremblay et al., 2018).

Conclusions

Cheetah motion research provides a revealing example of the interactions between research method and application. Besides giving an overview of the best current technologies for studying fast terrestrial animals in the wild, and the likely avenues for future development in this area, the purpose of examining the evolution of field biomechanics from this point of view is to give insight into the factors that propel innovation in how experimental biology is conducted.

Throughout the literature discussed, we find instances of new technology driving new research and of new research demands driving new technology. Film, microelectronics and deep learning are predominant technological breakthroughs that facilitated major shifts in motion capture methods. Key developments often also result from novel combinations of established technologies, such as the fusion of GPS and IMU sensors to produce tracking collars with better spatial resolution. The importance of interdisciplinary research is strongly evidenced, with many recent advances being motivated by overlapping interests with the legged robotics field, and facilitated by the related increase in engineering attention given to these problems.

Partly because of the influence of robotics, current research questions in cheetah biomechanics require detailed information about the whole-body motion of the animal and how it is actuated. Established technologies are close to obtaining this for captive cheetahs, but cost, invasiveness and the quantity and quality of attainable data remain primary obstacles to achieving the same in the wild. There are, however, promising emerging approaches – many of which have been demonstrated in human biomechanics or with domestic animals – that indicate how this gap might be closed in future. By forcing breaking technologies to adapt to wild environments, chasing the cheetah advances not only our understanding of terrestrial locomotion at its performance limits but also our ability to study locomotion throughout nature.

The authors thank Cheetah Outreach, Ann van Dyk Cheetah Centre, and South African National Parks (SANParks) for allowing access to their cheetahs.

Funding

The research supporting this Review was supported by South African National Research Foundation (grant no. 137762), Google and Oppenheimer Memorial Trust.

Alexander
,
R. M.
,
Maloiy
,
G.
,
Hunter
,
B.
,
Jayes
,
A.
and
Nturibi
,
J.
(
1979
).
Mechanical stresses in fast locomotion of buffalo (Syncews coffer) and elephant (Loxodonta africana)
.
J. Zool.
189
,
135
-
144
.
Andrada
,
E.
,
Reinhardt
,
L.
,
Lucas
,
K.
and
Fischer
,
M. S.
(
2017
).
Three-dimensional inverse dynamics of the forelimb of beagles at a walk and trot
.
Am. J. Vet. Res.
78
,
804
-
817
.
Andriluka
,
M.
,
Pishchulin
,
L.
,
Gehler
,
P.
and
Schiele
,
B
. (
2014
).
2D human pose estimation: new benchmark and state of the art analysis
.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
.
Azmy
,
S. N.
,
Sah
,
S. A. M.
,
Shafie
,
N. J.
,
Ariffin
,
A.
,
Majid
,
Z.
,
Ismail
,
M. N. A.
and
Shamsir
,
M. S.
(
2012
).
Counting in the dark: non-intrusive laser scanning for population counting and identifying roosting bats
.
Sci. Rep.
2
,
524
.
Basu
,
C.
,
Wilson
,
A. M.
and
Hutchinson
,
J. R.
(
2019a
).
The locomotor kinematics and ground reaction forces of walking giraffes
.
J. Exp. Biol.
222
,
jeb159277
.
Basu
,
C. K.
,
Deacon
,
F.
,
Hutchinson
,
J. R.
and
Wilson
,
A. M.
(
2019b
).
The running kinematics of free-roaming giraffes, measured using a low cost unmanned aerial vehicle (UAV)
.
PeerJ
7
,
e6312
.
Bennitt
,
E.
,
Bartlam-Brooks
,
H. L. A.
,
Hubel
,
T. Y.
and
Wilson
,
A. M.
(
2019
).
Terrestrial mammalian wildlife responses to unmanned aerial systems approaches
.
Sci. Rep.
9
,
2142
.
Bragança
,
F. S.
,
Bosch
,
S.
,
Voskamp
,
J.
,
Marin-Perianu
,
M.
,
van der Zwaag
,
B. J.
,
Vernooij
,
J. C.
,
van Weeren
,
P. R.
and
Back
,
W.
(
2017
).
Validation of distal limb mounted inertial measurement unit sensors for stride detection in warmblood horses at walk and trot
.
Equine Vet. J.
49
,
545
-
551
.
Briggs
,
R.
,
Lee
,
J.
,
Haberland
,
M.
and
Kim
,
S
. (
2012
).
Tails in biomimetic design: Analysis, simulation, and experiment
.
2012 IEEE/RSJ International Conference on Intelligent Robots and Systems
.
Brown
,
N. P.
,
Bertocci
,
G.
,
States
,
G. J. R.
,
Levine
,
G. J.
,
Levine
,
J. M.
and
Howland
,
D. R.
(
2020
).
Development of a canine rigid body musculoskeletal computer model to evaluate gait
.
Front. Bioeng. Biotechnol.
8
,
150
.
Calabrese
,
E.
,
Taverni
,
G.
,
Awai Easthope
,
C.
,
Skriabine
,
S.
,
Corradi
,
F.
,
Longinotti
,
L.
,
Eng
,
K.
and
Delbruck
,
T
. (
2019
).
Dhp19: Dynamic vision sensor 3d human pose dataset
.
Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops
.
Clemente
,
C. J.
,
Cooper
,
C. E.
,
Withers
,
P. C.
,
Freakley
,
C.
,
Singh
,
S.
and
Terrill
,
P.
(
2016
).
The private life of echidnas: using accelerometry and GPS to examine field biomechanics and assess the ecological impact of a widespread, semi-fossorial monotreme
.
J. Exp. Biol.
219
,
3271
-
3283
.
Curtin
,
N. A.
,
Bartlam-Brooks
,
H. L.
,
Hubel
,
T. Y.
,
Lowe
,
J. C.
,
Gardner-Medwin
,
A. R.
,
Bennitt
,
E.
,
Amos
,
S. J.
,
Lorenc
,
M.
,
West
,
T. G.
and
Wilson
,
A. M.
(
2018
).
Remarkable muscles, remarkable locomotion in desert-dwelling wildebeest
.
Nature
563
,
393
-
396
.
Daley
,
M. A.
and
Biewener
,
A. A.
(
2003
).
Muscle force-length dynamics during level versus incline locomotion: a comparison of in vivo performance of two guinea fowl ankle extensors
.
J. Exp. Biol.
206
,
2941
-
2958
.
Davies
,
A. B.
and
Asner
,
G. P.
(
2014
).
Advances in animal ecology from 3D-LiDAR ecosystem mapping
.
Trends Ecol. Evol.
29
,
681
-
691
.
Davies
,
Z. T.
,
Spence
,
A. J.
and
Wilson
,
A. M.
(
2019
).
Ground reaction forces of overground galloping in ridden Thoroughbred racehorses
.
J. Exp. Biol.
222
,
jeb204107
.
Druelle
,
F.
,
Supiot
,
A.
,
Meulemans
,
S.
,
Schouteden
,
N.
,
Molina-Vila
,
P.
,
Rimbaud
,
B.
,
Aerts
,
P.
and
Berillon
,
G.
(
2021
).
The quadrupedal walking gait of the olive baboon, Papio anubis: an exploratory study integrating kinematics and EMG
.
J. Exp. Biol.
224
,
jeb242587
.
Fry
,
S. N.
,
Sayaman
,
R.
and
Dickinson
,
M. H.
(
2003
).
The aerodynamics of free-flight maneuvers in Drosophila
.
Science
300
,
495
-
498
.
Galea
,
N.
,
Murphy
,
F.
,
Gaschk
,
J. L.
,
Schoeman
,
D. S.
and
Clemente
,
C. J.
(
2021
).
Quantifying finer-scale behaviours using self-organising maps (SOMs) to link accelerometery signatures with behavioural patterns in free-roaming terrestrial animals
.
Sci. Rep.
11
,
1
-
10
.
Gallego
,
G.
,
Delbrück
,
T.
,
Orchard
,
G.
,
Bartolozzi
,
C.
,
Taba
,
B.
,
Censi
,
A.
,
Leutenegger
,
S.
,
Davison
,
A. J.
,
Conradt
,
J.
,
Daniilidis
,
K.
et al. 
(
2020
).
Event-based vision: A survey
.
IEEE Trans. Pattern Anal. Mach. Intell.
44
,
154
-
180
.
Gärtner
,
E.
,
Andriluka
,
M.
,
Xu
,
H.
,
Sminchisescu
,
C
. (
2022
).
Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video
.
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition
.
Gosztolai
,
A.
,
Günel
,
S.
,
Lobato-Ríos
,
V.
,
Pietro Abrate
,
M.
,
Morales
,
D.
,
Rhodin
,
H.
,
Fua
,
P.
and
Ramdya
,
P.
(
2021
).
LiftPose3d, a deep learning-based approach for transforming two-dimensional to three-dimensional poses in laboratory animals
.
Nat. Methods
18
,
975
-
981
.
Grünewälder
,
S.
,
Broekhuis
,
F.
,
Macdonald
,
D. W.
,
Wilson
,
A. M.
,
McNutt
,
J. W.
,
Shawe-Taylor
,
J.
and
Hailes
,
S.
(
2012
).
Movement activity based classification of animal behaviour with an application to data from cheetah (Acinonyx jubatus)
.
PLoS ONE
7
,
e49120
.
Günel
,
S.
,
Rhodin
,
H.
,
Morales
,
D.
,
Campagnolo
,
J.
,
Ramdya
,
P.
and
Fua
,
P.
(
2019
).
DeepFly3D, a deep learning-based approach for 3D limb and appendage tracking in tethered, adult Drosophila
.
eLife
8
,
e48571
.
Haagensen
,
T.
,
Gaschk
,
J. L.
,
Schultz
,
J. T.
and
Clemente
,
C. J.
(
2022
).
Exploring the limits to turning performance with size and shape variation in dogs
.
J. Exp. Biol.
225
,
jeb244435
.
Han
,
Y. C.
,
Wong
,
K. I.
and
Murray
,
I.
(
2019
).
Gait phase detection for normal and abnormal gaits using imu
.
IEEE Sens. J.
19
,
3439
-
3448
.
Hedrick
,
T. L.
and
Biewener
,
A. A.
(
2007
).
Low speed maneuvering flight of the rose-breasted cockatoo (Eolophus roseicapillus). I. Kinematic and neuromuscular control of turning
.
J. Exp. Biol.
210
,
1897
-
1911
.
Hetem
,
R. S.
,
Mitchell
,
D.
,
de Witt
,
B. A.
,
Fick
,
L. G.
,
Meyer
,
L. C.
,
Maloney
,
S. K.
and
Fuller
,
A.
(
2013
).
Cheetah do not abandon hunts because they overheat
.
Biol. Lett.
9
,
20130472
.
Hildebrand
,
M.
(
1959
).
Motions of the running cheetah and horse
.
J. Mammal.
40
,
481
-
495
.
Hudson
,
P. E.
,
Corr
,
S. A.
and
Wilson
,
A. M.
(
2012
).
High speed galloping in the cheetah (Acinonyx jubatus) and the racing greyhound (Canis familiaris): spatio-temporal and kinetic characteristics
.
J. Exp. Biol.
215
,
2425
-
2434
.
Hutchinson
,
J. R.
,
Rankin
,
J. W.
,
Rubenson
,
J.
,
Rosenbluth
,
K. H.
,
Siston
,
R. A.
and
Delp
,
S. L.
(
2015
).
Musculoskeletal modelling of an ostrich (Struthio camelus) pelvic limb: influence of limb orientation on muscular capacity during locomotion
.
PeerJ
3
,
e1001
.
Ijspeert
,
A. J.
(
2014
).
Biorobotics: using robots to emulate and investigate agile locomotion
.
Science
346
,
196
-
203
.
Ionescu
,
C.
,
Papava
,
D.
,
Olaru
,
V.
and
Sminchisescu
,
C.
(
2013
).
Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments
.
IEEE Trans. Pattern Anal. Mach. Intell.
36
,
1325
-
1339
.
Jenkins
,
G. J.
,
Hakim
,
C. H.
,
Yang
,
N. N.
,
Yao
,
G.
and
Duan
,
D.
(
2018
).
Automatic characterization of stride parameters in canines with a single wearable inertial sensor
.
PLoS ONE
13
,
1
-
15
.
Jiang
,
L.
,
Lee
,
C.
,
Teotia
,
D.
and
Ostadabbas
,
S.
(
2022
).
Animal pose estimation: a closer look at the state-of-the-art, existing gaps and opportunities
.
Comput. Vis. Image Underst.
222
,
103483
.
Joska
,
D.
,
Clark
,
L.
,
Muramatsu
,
N.
,
Jericevich
,
R.
,
Nicolls
,
F.
,
Mathis
,
A.
,
Mathis
,
M. W.
and
Patel
,
A
. (
2021
).
AcinoSet: A 3D pose estimation dataset and baseline models for cheetahs in the wild
.
2021 IEEE International Conference on Robotics and Automation (ICRA), Xi'an, China, 2021, pp.
13901
-
13908
.
Kamimura
,
T.
,
Aoi
,
S.
,
Higurashi
,
Y.
,
Wada
,
N.
,
Tsuchiya
,
K.
and
Matsuno
,
F.
(
2021
).
Dynamical determinants enabling two different types of flight in cheetah gallop to enhance speed through spine movement
.
Sci. Rep.
11
,
9631
.
Kane
,
S. A.
and
Zamani
,
M.
(
2014
).
Falcons pursue prey using visual motion cues: new perspectives from animal-borne cameras
.
J. Exp. Biol.
217
,
225
-
234
.
Karashchuk
,
P.
,
Rupp
,
K. L.
,
Dickinson
,
E. S.
,
Walling-Bell
,
S.
,
Sanders
,
E.
,
Azim
,
E.
,
Brunton
,
B. W.
and
Tuthill
,
J. C.
(
2021
).
Anipose: A toolkit for robust markerless 3D pose estimation
.
Cell Rep.
36
,
109730
.
Kays
,
R.
,
Crofoot
,
M. C.
,
Jetz
,
W.
and
Wikelski
,
M.
(
2015
).
Terrestrial animal tracking as an eye on life and planet
.
Science
348
,
aaa2478
.
Klasen
,
M.
and
Steinhage
,
V.
(
2022
).
Improving wildlife tracking using 3D information
.
Ecological Informatics
68
,
101535
.
Krause
,
D. J.
,
Hinke
,
J. T.
,
Perryman
,
W. L.
,
Goebel
,
M. E.
and
LeRoi
,
D. J.
(
2017
).
An accurate and adaptable photogrammetric approach for estimating the mass and body condition of pinnipeds using an unmanned aerial system
.
PLoS ONE
12
,
e0187465
.
Labuguen
,
R.
,
Matsumoto
,
J.
,
Negrete
,
S. B.
,
Nishimaru
,
H.
,
Nishijo
,
H.
,
Takada
,
M.
,
Go
,
Y.
,
Inoue
,
K.-i.
and
Shibata
,
T.
(
2021
).
Macaquepose: A novel “in the wild” macaque monkey pose dataset for markerless motion capture
.
Front. Behav. Neurosci.
14
,
581154
.
LeCun
,
Y.
,
Bengio
,
Y.
and
Hinton
,
G.
(
2015
).
Deep learning
.
Nature
521
,
436
-
444
.
Li
,
J.
,
Zhang
,
J.
,
Wang
,
Z.
,
Shen
,
S.
,
Wen
,
C.
,
Ma
,
Y.
,
Xu
,
L.
,
Yu
,
J.
and
Wang
,
C
. (
2022
).
LiDARCap: long-range marker-less 3D human motion capture with LiDAR point clouds
.
2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, 2022,
pp.
20470
-
20480
.
Lord
,
R. D.
,
Bellrose
,
F. C.
and
Cochran
,
W. W.
(
1962
).
Radiotelemetry of the respiration of a flying duck
.
Science
137
,
39
-
40
.
Mathis
,
A.
,
Mamidanna
,
P.
,
Cury
,
K. M.
,
Abe
,
T.
,
Murthy
,
V. N.
,
Mathis
,
M. W.
and
Bethge
,
M.
(
2018
).
Deeplabcut: markerless pose estimation of user-defined body parts with deep learning
.
Nat. Neurosci.
21
,
1281
-
1289
.
Mathis
,
A.
,
Biasi
,
T.
,
Schneider
,
S.
,
Yuksekgonul
,
M.
,
Rogers
,
B.
,
Bethge
,
M.
and
Mathis
,
M. W
. (
2021
).
Pretraining boosts out-of-domain robustness for pose estimation
.
Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, 2021
, pp.
1858
-
1867
.
Mehta
,
D.
,
Rhodin
,
H.
,
Casas
,
D.
,
Fua
,
P.
,
Sotnychenko
,
O.
,
Xu
,
W.
and
Theobalt
,
C
. (
2017
).
Monocular 3D human pose estimation in the wild using improved CNN supervision
. In
2017 International Conference on 3D Vision (3DV), Qingdao, China, 2017
, pp.
506
-
516
.
Merkens
,
H. W.
,
Schamhardt
,
H. C.
,
van Osch
,
G. J. V. M.
and
van den Bogert
,
A. J.
(
1993
).
Ground reaction force patterns of dutch warmblood horses at normal trot
.
Equine Vet. J.
25
,
134
-
137
.
Mueller
,
F.
,
Bernard
,
F.
,
Sotnychenko
,
O.
,
Mehta
,
D.
,
Sridhar
,
S.
,
Casas
,
D.
and
Theobalt
,
C
. (
2018
).
GANerated hands for real-time 3D hand tracking from monocular RGB
.
018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 2018,
pp.
49
-
59
.
Muybridge
,
J.
(
1882
).
The horse in motion
.
Nature
25
,
605
.
Nath
,
T.
,
Mathis
,
A.
,
Chen
,
A. C.
,
Patel
,
A.
,
Bethge
,
M.
and
Mathis
,
M. W.
(
2019
).
Using DeepLabCut for 3D markerless pose estimation across species and behaviors
.
Nat. Protoc.
14
,
2152
-
2176
.
Patel
,
A.
and
Boje
,
E.
(
2015
).
On the conical motion of a two-degree-of-freedom tail inspired by the cheetah
.
IEEE Trans. Robot.
31
,
1555
-
1560
.
Patel
,
A.
and
Braae
,
M
. (
2013
).
Rapid turning at high-speed: Inspirations from the cheetah's tail
.
2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, Tokyo, Japan, 2013
, pp.
5506
-
5511
.
Patel
,
A.
and
Braae
,
M
. (
2014
).
Rapid acceleration and braking: Inspirations from the cheetah's tail
. In
2014 IEEE International Conference on Robotics and Automation (ICRA), Hong Kong, China, 2014
, pp.
793
-
799
.
Patel
,
A.
,
Stocks
,
B.
,
Fisher
,
C.
,
Nicolls
,
F.
and
Boje
,
E.
(
2017
).
Tracking the cheetah tail using animal-borne cameras, GPS, and an IMU
.
IEEE Sens. Lett.
1
,
1
-
4
.
Pereira
,
T. D.
,
Tabris
,
N.
,
Matsliah
,
A.
,
Turner
,
D. M.
,
Li
,
J.
,
Ravindranath
,
S.
,
Papadoyannis
,
E. S.
,
Normand
,
E.
,
Deutsch
,
D. S.
,
Wang
,
Z. Y.
et al. 
(
2022
).
SLEAP: A deep learning system for multi-animal pose tracking
.
Nat. Methods
19
.
Perlmutter
,
M.
and
Breit
,
S
. (
2016
).
The future of the mems inertial sensor performance, design and manufacturing
. In
2016 DGON inertial sensors and systems (ISS), Karlsruhe, Germany, 2016
, pp.
1
-
12
.
Pezzuolo
,
A.
,
Giora
,
D.
,
Guo
,
H.
,
Ma
,
Q.
,
Guercini
,
S.
and
Marinello
,
F.
(
2019
).
A comparison of low-cost techniques for three-dimensional animal body measurement in livestock buildings
.
IOP Conf. Ser. Earth Environ. Sci.
275
,
012015
.
Pocock
,
R. I
. (
1927
).
Description of a new species of cheetah (Acinonyx)
. In
Proceedings of the Zoological Society of London
, Vol.
97
.
Wiley Online Library
.
Ray
,
S.
and
Stopfer
,
M. A.
(
2022
).
Argos: A toolkit for tracking multiple animals in complex visual environments
.
Method. Ecol. Evol.
13
,
585
-
595
.
Roberts
,
T.
and
Gabaldon
,
A.
(
2008
).
Interpreting muscle function from emg: lessons learned from direct measurements of muscle force
.
Integr. Comp. Biol.
48
,
312
-
320
.
Schwab
,
K
. (
2017
).
The Fourth Industrial Revolution
.
Portfolio Penguin
.
Sengupta
,
A.
,
Jin
,
F.
,
Zhang
,
R.
and
Cao
,
S.
(
2020
).
mm-pose: Real-time human skeletal posture estimation using mmWave radars and CNNs
.
IEEE Sens. J.
20
,
10032
-
10044
.
Seth
,
A.
,
Hicks
,
J. L.
,
Uchida
,
T. K.
,
Habib
,
A.
,
Dembia
,
C. L.
,
Dunne
,
J. J.
,
Ong
,
C. F.
,
DeMers
,
M. S.
,
Rajagopal
,
A.
,
Millard
,
M.
et al. 
(
2018
).
Opensim: Simulating musculoskeletal dynamics and neuromuscular control to study human and animal movement
.
PLoS Comput. Biol.
14
,
e1006223
.
Sharp
,
N.
(
1997
).
Timed running speed of a cheetah (Acinonyx jubatus)
.
J. Zool.
241
,
493
-
494
.
Shine
,
C. L.
,
Robbins
,
C. T.
,
Nelson
,
O. L.
and
McGowan
,
C. P.
(
2017
).
Grizzly bear (Ursus arctos horribilis) locomotion: forelimb joint mechanics across speed in the sagittal and frontal planes
.
J. Exp. Biol.
220
,
1322
-
1329
.
Spence
,
A. J.
,
Tan
,
H.
and
Wilson
,
A.
(
2008
).
Accuracy of the TurfTrax Racing Data System for determination of equine speed and position
.
Equine Vet. J.
40
,
680
-
683
.
Stark
,
H.
,
Fischer
,
M.
,
Hunt
,
A.
,
Young
,
F.
,
Quinn
,
R.
and
Andrada
,
E.
(
2021
).
A three-dimensional musculoskeletal model of the dog
.
Sci. Rep.
11
,
11335
.
Straw
,
A. D.
,
Branson
,
K.
,
Neumann
,
T. R.
and
Dickinson
,
M. H.
(
2011
).
Multi-camera real-time three-dimensional tracking of multiple flying animals
.
J. R. Soc. Interface
8
,
395
-
409
.
Sun
,
J.
,
Wu
,
J.
,
Liao
,
X.
,
Wang
,
S.
and
Wang
,
M.
(
2022
).
A large-scale mouse pose dataset for mouse pose estimation
.
Symmetry
14
,
875
.
Tan
,
H.
and
Wilson
,
A. M.
(
2011
).
Grip and limb force limits to turning performance in competition horses
.
Proc. R. Soc. B
278
,
2105
-
2111
.
Tang
,
Y.
,
Chi
,
Y.
,
Sun
,
J.
,
Huang
,
T.-H.
,
Maghsoudi
,
O. H.
,
Spence
,
A.
,
Zhao
,
J.
,
Su
,
H.
and
Yin
,
J.
(
2020
).
Leveraging elastic instabilities for amplified performance: Spine-inspired high-speed and high-force soft robots
.
Sci. Adv.
6
,
eaaz6912
.
Tatler
,
J.
,
Cassey
,
P.
and
Prowse
,
T. A.
(
2018
).
High accuracy at low frequency: detailed behavioural classification from accelerometer data
.
J. Exp. Biol.
221
,
jeb184085
.
Taylor
,
C. R.
and
Rowntree
,
V.
(
1973
).
Temperature regulation and heat balance in running cheetahs: a strategy for sprinters?
Am. J. Physiol.
224
,
848
-
851
.
Thompson
,
S. E
. (
1998
).
Built for Speed: the Extraordinary, Enigmatic Cheetah
.
Twenty-First Century Books
.
Tian
,
X.
,
Iriarte-Diaz
,
J.
,
Middleton
,
K.
,
Galvao
,
R.
,
Israeli
,
E.
,
Roemer
,
A.
,
Sullivan
,
A.
,
Song
,
A.
,
Swartz
,
S.
and
Breuer
,
K.
(
2006
).
Direct measurements of the kinematics and dynamics of bat flight
.
Bioinspir. Biomim.
1
,
S10
-
S18
.
Tremblay
,
J.
,
Prakash
,
A.
,
Acuna
,
D.
,
Brophy
,
M.
,
Jampani
,
V.
,
Anil
,
C.
,
To
,
T.
,
Cameracci
,
E.
,
Boochoon
,
S.
and
Birchfield
,
S
. (
2018
).
Training deep networks with synthetic data: Bridging the reality gap by domain randomization
. In
Proceedings of the IEEE conference on computer vision and pattern recognition workshops, Salt Lake City, UT, USA, 2018
, pp.
1082
-
10828
.
van Eeden
,
W. D.
,
de Villiers
,
J. P.
,
Berndt
,
R. J.
,
Nel
,
W. A. J.
and
Blasch
,
E.
(
2018
).
Micro-Doppler radar classification of humans and animals in an operational environment
.
Expert Syst. Appl.
102
,
1
-
11
.
Von Marcard
,
T.
,
Henschel
,
R.
,
Black
,
M. J.
,
Rosenhahn
,
B.
and
Pons-Moll
,
G.
(
2018
).
Recovering accurate 3D human pose in the wild using imus and a moving camera
. In
Proceedings of the European Conference on Computer Vision (ECCV)
.
Wang
,
P.
,
Zhang
,
Y.
,
Ma
,
Y.
,
Liang
,
F.
,
An
,
Q.
,
Xue
,
H.
,
Yu
,
X.
,
Lv
,
H.
and
Wang
,
J.
(
2019
).
Method for distinguishing humans and animals in vital signs monitoring using IR-UWB radar
.
Int. J. Environ. Res. Public Health
16
,
4462
.
Williams
,
S. B.
,
Tan
,
H.
,
Usherwood
,
J. R.
and
Wilson
,
A. M.
(
2009
).
Pitch then power: limitations to acceleration in quadrupeds
.
Biol. Lett.
5
,
610
-
613
.
Wilson
,
A. M.
,
Lowe
,
J. C.
,
Roskilly
,
K.
,
Hudson
,
P. E.
,
Golabek
,
K. A.
and
McNutt
,
J. W.
(
2013a
).
Locomotion dynamics of hunting in wild cheetahs
.
Nature
498
,
185
-
189
.
Wilson
,
J. W.
,
Mills
,
M. G.
,
Wilson
,
R. P.
,
Peters
,
G.
,
Mills
,
M. E.
,
Speakman
,
J. R.
,
Durant
,
S. M.
,
Bennett
,
N. C.
,
Marks
,
N. J.
and
Scantlebury
,
M.
(
2013b
).
Cheetahs, Acinonyx jubatus, balance turn capacity with pace when chasing prey
.
Biol. Lett.
9
,
20130620
.
Wilson
,
A. M.
,
Hubel
,
T. Y.
,
Wilshin
,
S. D.
,
Lowe
,
J. C.
,
Lorenc
,
M.
,
Dewhirst
,
O. P.
,
Bartlam-Brooks
,
H. L. A.
,
Diack
,
R.
,
Bennitt
,
E.
,
Golabek
,
K. A.
et al. 
(
2018
).
Biomechanics of predator–prey arms race in lion, zebra, cheetah and impala
.
Nature
554
,
183
-
188
.
Witte
,
T. H.
,
Knill
,
K.
and
Wilson
,
A. M.
(
2004
).
Determination of peak vertical ground reaction force from duty factor in the horse (Equus caballus)
.
J. Exp. Biol.
207
,
3639
-
3648
.
Witte
,
T. H.
,
Hirst
,
C.
and
Wilson
,
A. M.
(
2006
).
Effect of speed on stride parameters in racehorses at gallop in field conditions
.
J. Exp. Biol.
209
,
4389
-
4397
.
Yi
,
X.
,
Zhou
,
Y.
,
Habermann
,
M.
,
Shimada
,
S.
,
Golyanik
,
V.
,
Theobalt
,
C.
and
Xu
,
F
. (
2022
).
Physical inertial poser (PIP): Physics-aware real-time human motion tracking from sparse inertial sensors
.
arXiv
Zhang
,
X.
,
Jenkins
,
G.
,
Hakim
,
C.
,
Duan
,
D.
and
Yao
,
G.
(
2022
).
Four-limb wireless imu sensor system for automatic gait detection in canines
.
Sci. Rep.
12
,
4788
.
Zhao
,
M.
,
Li
,
T.
,
Alsheikh
,
M. A.
,
Tian
,
Y.
,
Zhao
,
H.
,
Torralba
,
A.
and
Katabi
,
D
. (
2018
).
Through-wall human pose estimation using radio signals
. In
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition
.
ISBN 978–1-5386-6420-9
.
Zuffi
,
S.
,
Kanazawa
,
A.
,
Jacobs
,
D.
and
Black
,
M. J
. (
2017
).
3D menagerie: Modeling the 3D shape and pose of animals
. In
IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 2017
, pp.
5524
-
5532
.
Zuffi
,
S.
,
Kanazawa
,
A.
,
Berger-Wolf
,
T.
and
Black
,
M. J
. (
2019
).
Three-D safari: Learning to estimate zebra pose, shape, and texture from images ‘in the wild
’. In
Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Korea (South), 2019
, pp.
5358
-
5367
.

Competing interests

The authors declare no competing or financial interests.