Opening the black box of bird-window collisions: passive video recordings in a residential backyard

Study area

We filmed in the backyard of a residential property in London, Ontario, Canada from May to October in 2019 and in 2020. The property is located immediately adjacent to a mature eastern deciduous forest nested within a suburban neighbourhood. We expected there to be a relatively high abundance of resident and migratory bird species in study area based on the adjacent forest’s designation as Significant Wildlife Habitat by the Province of Ontario, results of long-term window building monitoring conducted within a few kilometers of the site, and the property owners’ observations over several years of feeding birds in the backyard. The property owners maintain a variety of bird feeders and bathing stations in the backyard that were included in the setup for the present study. We are aware that at a spatial level, we only gathered information from a single location, so our results and conclusions cannot be generalized to all window collisions contexts. However, we focused this study on the role of variables related to the behaviour of the animals (velocity, angle of approach, etc.) that are not well-understood. Therefore, we consider that our study can provide some novel insights for future testing despite the limited spatial representation.

The configuration of the backyard study area is shown in Fig. 1. The backyard contained several bird feeders that were maintained with a combination of seed mixes by the homeowner throughout the study period. For the first year of monitoring, feeders were placed at two positions. In the second year, a third feeder was added by the homeowner positioned closer to one of the windows. The backyard also contained two bird baths as well as a large patio table with an umbrella and chairs. The position of the house relative to the mature trees on the property meant that the south-facing windows under study were shaded through most of the day and exposed to direct sunlight only in early morning (approximately 7:00–10:00 am) and late afternoon (approximately 4:00–5:00 pm).

Camera system

We installed four Arlo Pro 2 wireless home security cameras in the study area (VMS4430P; Netgear, San Jose, CA, USA). The cameras recorded in 1,080 p resolution at 24 frames per second. Fluctuations in wireless signal strength occasionally reduced the resolution to 720 p. We positioned three of the cameras along the back wall of the house immediately adjacent to three sets of large glass sliding doors (Fig. 1). We used brackets to mount these cameras to the wall approximately 2 m above ground adjacent to the windows, with the camera lenses aimed approximately parallel to the nearest window and the ground, and offset from the window frame by approximately 20 cm. We positioned a fourth camera on a tripod (approx. 1.2 m above the ground) near the bird feeders to provide a view of the entire study area. Thus, the configuration of the four cameras with overlapping fields of view allowed for most events to be captured from multiple angles.

The cameras were triggered synchronously to record video of a fixed duration (10–15 s post-trigger, plus 3 s pre-trigger) using a combination of infrared motion detection (e.g., a bird passing in front of the lens) and audio detection (i.e., a spike in amplitude associated with the thud of a bird striking the glass). The sensitivity of each type of trigger was adjusted within each camera to optimize the signal-to-noise ratio and reduce the frequency of recordings being triggered by unrelated environmental events such as falling leaves, background noise or sudden changes in ambient light. Audio detection sensitivity was tested by simulating the sound of a collision by clapping hands in front of the windows. The audio trigger sensitivity of three cameras was set to three or medium; the camera placed on the tripod had audio triggers disabled to avoid detecting sounds produced by birds at the feeders. Motion sensitivity on all four cameras was set to “7” or medium-high. The cameras uploaded footage directly to a cloud server. In addition to the motion and audio-triggered recordings, we programmed two of the cameras to record video on a continuous rolling basis to provide a failsafe in case events failed to trigger the cameras.

Video analysis

All the video clips that were captured using audio or motion triggers were stored on the cloud server for a period of 7 days. The continuous recorded footage recorded by two of the cameras was stored for 14 days. Footage was inspected independently throughout the study period by the first author and an assistant. Video clips of birds approaching or interacting with the windows were downloaded for permanent storage and included in later analyses.

A single rater analyzed the video clips using a standardized protocol in VLC Media Player (VideoLAN, Paris, France). All video clips that were saved throughout the study period were processed together in one batch. For each saved video clip, events were classified as either a collision or near-miss. An event was coded as a collision if the bird made physical contact with the glass. We excluded any events in which birds contacted the glass repeatedly at close range, which might indicate they were attacking their reflections on the glass rather than colliding. An event was coded as a near miss if the bird initially appeared to fly along a path that would lead to it colliding with a window, but suddenly changed its trajectory (reducing velocity or sharply shifting direction) within 2 m of approaching the building, and subsequently did not make impact. A change in trajectory preceding a near miss was determined by drawing a straight line along the initial flight path of the bird and confirming that the line projected toward a window, then comparing the bird’s observed displacement across subsequent frames. Examples of collision and near miss events are provided in the Supplemental Material. We included video frames in the flight analyses that began either when the bird first entered the frame from off-camera and passed a reference point within the measurements of the backyard, or from the last spot where the bird took off from perching within the backyard (i.e., the “starting” position). The frames included in the flight analyses ended either when the bird impacted a window or at the frame when the bird was closest to the window. We estimated the duration of each flight by multiplying the total number of video frames included in the analysis by 0.042 s (the duration between frames at the 24 fps recording setting used by the cameras). We estimated the distance covered by the flight by marking the change in the position of the bird across consecutive video frames and comparing the difference relative to known dimensions of the study area. For each event that was included in the analyses, video clips recorded by multiple cameras (typically 2 or 3) was examined by the rater in combination to reduce bias in visual approximation of changes in the birds’ position over time. The distance traversed by a bird between consecutive video frames was noted in a spreadsheet (Microsoft Excel; Microsoft, Redmond, WA, USA). We then calculated average velocity within each flight as the total distance covered in flight divided by the total flight duration. We compared changes in the position of the bird between video frames with dimensions of the backyard to produce an overhead map showing the approximate flight paths followed by birds as they approached the windows (see Supplemental Material). We visually estimated the horizontal angle of approach (i.e., the azimuth) for each flight path relative to the normal angle from the window (e.g., a head-on approach at exactly 90° from the window would be represented as 0°) to the nearest 5° angle. To include horizontal angle of approach in our statistical analyses, we used the absolute value of the estimated angle from perpendicular thus controlling for direction (i.e., using values that ignore whether the bird approached from the left or right side of the window).

Statistical analyses

We ran generalized linear mixed models with the afex R package (Singmann et al., 2022) with two dependent variables (probability of collision, probability of visible injury) with the same independent factors: velocity, horizontal angle of approach, and time of the day. We included time of day as an index of variability in ambient lighting conditions that we expected to produce a consistent effect on the appearance of the window glass; we did not directly measure ambient light within the study area. Because we had more than one collision event per species in some data entries and we did not have individuals tagged, we included species as a random factor to account for the possibility of non-independence of samples in the model (i.e., the same individual birds approaching the windows multiple times over the duration of the study). The probability of collision when birds were on a collision approach included two values: 0, no collision (near miss), and 1, collision. The probability of visible injury considered only birds that collided with the window, including two values: 0, animals flew away immediately, and 1, animals were stunned (i.e., landed on the ground below the window for at least 5 s or past the end of the video recording) or died right after the collision. We used the absolute difference value of horizontal angle of approach relative to the perpendicular, ignoring left or right direction. We excluded vertical angle of approach as most birds flew approximately parallel to the ground and our camera set up did not allow us to quantify vertical angle accurately. Time of the day was transformed into decimals. None of our independent continuous factors showed any significant correlation between them (Pearson product moment correlation r < 0.15, P > 0.40). We focused the results on assessing whether the independent factors were significantly (P < 0.05) accounting for the variation in the dependent variables as well as presenting their effect size (and confidence intervals). Effect sizes were expressed in the odds ratio scale (Sperandei, 2014), which required that we centered our independent variables for the statistical analyses to standardize the effect sizes to the same scale. However, figures were presented using the probability scale and the uncentered independent factors to facilitate interpretation. Statistical analyses were conducted using R 4.2.0 (R Core Team, 2022). We present descriptive results from our video analyses with means and SDs reported for each coded variable.

Ethics statement

This study adhered to the guidelines of the University of Western Ontario and the Canadian Council on Animal Care (CCAC) and all animal handling procedures were approved by the institution’s Animal Care Committee (protocol 2019-118).

Flight characteristics and lethality of collisions

Velocity of approach yielded the highest effect size of the independent variables studied. The probabilities of collision increased with velocity faster in the range of 2 to 9 m/s and then leveled off. Additionally, birds flying faster than 7 m/s faced a higher probability of suffering a visible injury. However, our study involved a small sample size, especially for video recordings of birds that were stunned or killed by a collision and for birds that nearly avoided making impact, so the importance of the observed difference in flight velocity should be interpreted with caution. Our results support the idea that the flight velocity and angle at which a bird makes impact with a window may influence the odds of a collision as well as the outcome of a collision (injury, death, etc.).

Birds admitted to rehabilitation centres following a window collision often have patterns of soft tissue injury sustained by the head and body as force is absorbed upon impact (e.g., Hudecki & Finegan, 2018). Yet, there has been no research to date aimed at characterizing the mechanics of collisions and understanding how these injuries occur. It seems likely that the severity of injury is related to the amount of force that the bird absorbs when it collides. Modelling collision dynamics in three-dimensional space requires more precise measurements than we were able to obtain with our video cameras. However, based on our empirical data, we can conceptualize how energy is distributed when birds impact a window under different configurations. When a bird strikes a window, some of the kinetic energy generated by the impact will be absorbed by the window, while the rest is directed outward at the bird. Generally, if a soft object (e.g., a bird) strikes a harder object (e.g., a window), most of the energy available for producing post-collision (final) velocity is stored in the soft object. Part of that energy pushes the bird outward from the window (i.e., normal force), while the rest of the force is absorbed by the bird’s body, compressing its tissues, and potentially causing injuries such as ruptures of blood vessels and broken bones. In theory, the ratio between these two components of force acting on the bird post-collision may vary with the duration of the impact, resulting from the extent that the bird’s body compresses. The angle at which a bird approaches the window may also affect the amount and direction of force its body experiences upon impact. A collision at a perpendicular angle to the window (i.e., a head-on collision) will result in momentum being conserved along only one axis. In other words, the momentum of the bird as it flies forward into the window will be transferred and redirected outward at the bird. However, a collision at an oblique angle, either in the horizontal or vertical plane, would result in components of momentum being conserved along multiple axes independently (Fig. 4). A bird that collides at a glancing angle may experience less force, reducing the likelihood of it sustaining injury. For example, all five of the collisions we recorded that resulted in birds being stunned or immediately killed involved angles of approach that were approximately perpendicular to the glass, while collisions that resulted in birds immediately flying away included more variable angles of approach.

More energy absorption by the bird’s body may result in lower final velocity and displacement, but this could produce more damage to the bird’s tissues. Perhaps variation in morphology and body composition, and resulting differences in how much energy is absorbed by soft tissues during collisions, could explain why collisions are apparently more lethal for some birds than others (e.g., Nichols et al., 2018). In the real world, collisions occurring in three-dimensional space may be complicated by other factors, such as the bird’s behaviour and wind acting on the bird. Future work using cameras with higher frame rate and resolution could be used to observe differences in the durations of birds impacting windows. One way that transfer of energy during bird-window collisions could be studied further is by simulating crash tests using objects with similar physical and mechanical properties to the body of a bird.

Environmental variables and collision risk

Buildings with larger window surface area generally pose a higher risk of collisions for birds (Hager et al., 2013; Elmore et al., 2021). However, in our study we found that nearly half of the collisions we observed were with the smaller transom windows above larger glass doors. This suggests that small windows can also pose a significant risk of bird-window collisions. It seems likely that other factors may be important for explaining birds’ behaviour in this setting. For instance, when birds take off from the feeders or surrounding trees, they may aim their flight towards the smaller transom windows because they perceive it as safer cover than the larger sliding door windows below. Alternatively, birds may simply prefer to fly higher and are therefore more likely to hit upper windows.

Although the configuration of our cameras did not allow us to continuously monitor birds’ behaviour in the space surrounding the study area, the recorded footage revealed conspecific and interspecific interactions among birds directly preceding a collision in at least 12 instances. These included birds chasing each other, competing for access to feeders, and escaping predation. Our observations support the idea that distractions from social interactions and interactions with predators, that occur more frequently near bird attractants, may influence birds’ behaviour and ability to detect and avoid collisions with windows during fleeing or chasing.

Considerations for future research

Recording bird-window interactions passively in the field allows for behavioural observations that might otherwise require prohibitively invasive methodology with captive animals. In the present study, we captured nine near-miss events in which birds approached but did not collide with the windows. The total number of near-miss events was probably higher, but some events did not trigger video recording because the camera lenses and sensors were not oriented to observe birds several metres from the glass where detection may first occur. Future studies aiming to record near-misses more effectively could deploy more cameras oriented outwards, or combine cameras with other positional sensors. Given that most of the collision and near-miss events that produced triggers and were successfully recorded involved medium to large passerines (i.e., with body mass exceeding 40 g), it is possible that events involving smaller birds may be less likely to be picked up by the cameras. To limit detection bias, future studies using cameras for monitoring bird collisions should conduct preliminary trials to calibrate the camera trigger sensitivity.

The perceptual mechanisms involved in successful detection and avoidance of windows by birds in flight are not clearly defined (see Martin, 2011 for a discussion of bird vision and collisions with obstacles in general). Characterizing perceptual processes that birds use to avoid collisions could have implications for designing effective window collision prevention methods. These perceptual processes may interact with various factors, such as the morphology of the bird, interactions with a predator, or the rate of visual information being influenced by flight speed (Martin, 2017; Jackson et al., 2020, Bhagavatula et al., 2011). Recording pre-collision behaviour could provide an empirical basis to develop understanding of how birds see windows and how window collision deterrent technologies, such as window markings, could function. For example, birds may rely on specific parts of their visual field and depend on visual information from their surroundings prior to takeoff or to detect and avoid obstacles mid-flight. Visual sensitivity that is necessary for birds to detect an obstacle may vary across parts of the visual field and depending on the distance from the object. Birds with laterally-positioned eyes, such as many passerine species, may turn their heads to orient the high-acuity fovea at the centre of one eye towards stimuli of interest. No research to date has explored whether the conspicuousness (or probability of detection) of window collision deterrents to a bird could be affected by the materials being viewed by the bird through different parts of the visual field prior to takeoff or during flight. Our camera system did not provide the quality of footage necessary for analyzing birds’ head or gaze orientation upon approaching the glass in detail, but in all of the clips where head orientation was evident, we found no signs of birds turning their head upon approach. Future research simulating bird flights and collisions in controlled settings may yield more data to indicate how birds orient their eyes to detect objects and avoid collisions in flight.

The risk of bird-window collisions is elevated in environments with bird attractants such as feeders, bathing stations and native plants (Klem et al., 2004, Bayne, Scobie & Rawson-Clark, 2012, Kummer & Bayne, 2015, Kummer, Bayne & Machtans, 2016). Most bird-window collisions that we recorded using cameras at a residence with bird attractants left behind no trace and were not detected by the building occupants who were present for most of the study period. Our results are consistent with Kummer & Bayne (2015) who reported similar collision monitoring research carried at multiple residences with and without bird feeders. Their experimental protocol involved homeowners performing daily perimeter checks, and accepted forms of evidence of collisions beyond injured birds and carcasses, including homeowner observations through hearing or seeing a collision, as well as smudges, feathers or blood left on the glass. They found that the total number of bird collisions that occurred at several residences was comparable to our study, and that only 7.5% of all the collisions that were reported resulted in an observed fatality, whereas 92% of birds that collided were observed to fly away immediately or their status following the collision was unknown. This provides support for the notion that bird-window collisions could be far more common in residential settings than is realized by homeowners. If a large proportion of collision events are not observed, this could have implications for the reliability of datasets built using collision-reporting community science tools (e.g., Kummer, Bayne & Machtans, 2016).