Camera traps have become a norm for estimating wildlife populations utilizing random encounter models to express metrics such as abundance and density. One essential parameter required to evaluate such models is the speed of the trapped animal. This speed can be estimated by labeling the contact point of instances with the ground across image sequences and projecting the resulting pixels onto a model describing said ground. Our approach proposes using digital elevation models recorded via airborne laser scanning as an alternative to manually calibrating simpler ground models. A study to investigate the impact of DEMs' level of complexity on the accuracy of projected pixel distance estimates was conducted using a realistic dataset of 2629 2D labels and DEMs of three diferent resolutions. Reducing the resolution of the used DEM from leads to an average skew in distances of 2.84 m and 3.94 m, respectively, with widely dispersed individual errors. Further work is needed to assess the impact of these errors on speed and wildlife density calculations. Despite these challenges, the method shows promise as an alternative to currently used methods.
(D. C. Schedl)
CEUR
ceur-ws.org derived, enabling the computation of the speed of animals across the sequence. Naturally, the quality of the resulting estimate depends on the accuracy of the employed ground model.
Fitting a ground model requires a manual calibration routine per deployment location plus one camera calibration for the operating camera type and model. One such calibration routine involves placing a standardized pole (typically 1 m tall and marked at known intervals of 10 cm to 20 cm) perpendicular to the ground within the FOV and recording images. Afterward, specified pole features, such as both ends, are annotated in each image. These annotations, along with their actual distances and the camera intrinsics (i.e., sensor resolution, FOV, and optical center), allow fitting a ground model. For instance, the R-package CTtracking1 fits a flat or unidirectional (i.e., bending on one axis) planar ground model per deployed camera trap to predict animal positions and speed. The quality of the estimated model depends on the expressiveness of the pole positions, i.e., how well they represent the respective environment.
This laborious manual deployment calibration restricts the use of camera traps to locations that are accessible by humans and need to be repeated for each deployment session. Also, more complex terrains, such as those with ridges and slopes, require extensive calibration to map adequately. Simple ground models might miss such terrain features, leading to inaccuracies in the 3D coordinates. Therefore, we propose utilizing precise digital elevation models (DEMs)—typically recorded by airborne laser scanning—instead of simpler manually fitted models.
Although this approach appears straightforward, there is a deficit of studies investigating the utilization of DEMs for this specific use case. Notably unexplored is the impact of their level of complexity on the quality of the distances they assign to pixel coordinates. While the resource intensity increases with model complexity, a coarser model can lead to substantial inaccuracies in the distance estimates (see Figure 1). Findings in this area could provide valuable insights for improving camera-trap-based analysis. Hence, this work compares DEMs varying in precision to address the question: ”How critical is a DEM’s resolution and detail to the accuracy of projected pixel distance estimates?”
Most of the methods required and applied in the presented approach represent problems in computer graphics and computer vision. Fortunately, as such, these problems are likely to have already been solved in the form of algorithms or even software libraries.
Fundamental to all camera-trap-related data processing is determining its intrinsic metrics. For this
purpose, an OpenCV camera model is computed using substantial captures featuring a calibration
checkerboard pattern and the OpenCV library [
The more variable input required consists of a set of camera shots, each labeled to mark a feature of the triggering target. Further, the orientation data encompassing the camera’s location in GPS coordinates and its rotation in the given deployment are necessary to fit a DEM. Although modern camera traps may include features that automatically determine orientation parameters, these can also be assessed via external tools or by following guidelines, such as consistently facing the camera north.
Before projection, the 2D coordinates undergo undistortion employing the camera calibration and OpenCV. Subsequently, the resulting undistorted coordinates are used to define rays from the camera origin through the corresponding pixel. These rays can then be cast onto the geometry expressed by the DEM, with their intersection point representing the 3D location depicted by the respective pixel.
The projected 3D coordinates allow for further analysis, e.g. speed computations as required for REM computation. The following section goes into experimentally verifying the described method.
The camera-trap dataset, which forms the basis for our experiments, was collected in the field using model BTC-PATRIOT-FHD (Browning International S.A., 84050 Morgan, Utah) camera traps, with a resolution of 2688×1520 (4 MP). The study area, located in See, Austria, is positioned geospatially between the latitudinal values of 47.053 920 41° N to 47.081 984 77° N and longitudinally between 10.450 032 78° E to 10.488 935 68° E as illustrated in Figure 2a. The dataset comprises 2629 labeled instances of the three resident ungulate species red deer (Cervus elaphus), roe deer (Capreolus capreolus) and chamois (Rupicapra rupicapra) and was gathered throughout 23 deployments, each lasting approximately 150 to 160 days. Only two periods, lasting 31 days in September–October 2023 and February 2024 respectively, were annotated, which form the baseline for subsequent experiments. Although 2D annotations exist, they could not be directly employed in our experiments due to missing information about the internal data handling of the applied labeling tool (i.e., Agouti2). Therefore, labels are simulated by randomly selecting 2D coordinates concentrated in the lower half of the image sensor (cf. Figure 2b). Due to camera trap alignment, the upper sensor half would lead to ray casts that do not intersect with the ground. Although the 2D image coordinates are randomly sampled, the number of labels per camera deployment corresponds to the in-field data, as shown in Table 1.
(a) Orthophoto of the study area. (b) Labeled instances.
The evaluation compares DEMs with resolutions of 3 m, 20 m, and 50 m (c.f., Figure 3), regarding the most precise DEM as ground truth. The federal government of Tirol ofers the required elevation data3, originally recorded with a step size of 1 m via airborne laser scanning in 2018. This provided data consists of height texture tiles in the MGI Austria GK West coordinate system (EPSG:31254) and is processed (i.e., merged and transformed) with the cartography software QGIS 4. The coordinate reference system (CRS) WGS84/UTM33N (EPSG:32633) was chosen as the foundation for measuring distances and errors in meters. Given ThreeJS requires a 3D mesh for raycasting, the elevation data is lastly converted into a polygonal mesh using the Python libraries pyproj5 and Rasterio6.
Each camera deployment is given geospatial coordinates, which can be associated with 2D positions on the DEM. However, the precise height of the mounted cameras in situ is not measured. Therefore, these parameters are estimated using the DEMs and a relative height ofset of 0.5 m. Similarly, the camera rotation had to be inferred based on the assumption that the cameras were generally installed parallel to the ground and facing north to avoid direct sunlight. Therefore, first, the forward orientation of the camera was defined by moving 5 m north on the DEM while retaining the relative height ofset. Secondly, the camera roll (defining the upward direction) is estimated by sampling two elevations 3 m east and west of the camera’s deployment position. Placement and orientation computations have been carried out individually for each DEM resolution, resulting in slightly changed parameters per DEM and deployment, as illustrated in Figure 3. Compared to the 3 m DEM, the variation in simulated mounting altitude was between 0.0 m to 3.27 m and on average 0.54 m and 1.14 m (cf. Table 2) for the lower resolution DEMs. 3https://www.tirol.gv.at/als 4https://www.qgis.org/ 5https://github.com/pyproj4/pyproj 6https://github.com/rasterio/rasterio (a) 3 m DEM (ground truth).
(b) 20 m DEM.
(c) 50 m DEM.
Based on the simulated camera mounting, the 2629 labeled instances, consisting of 2D coordinates in images, are assigned to the corresponding deployments and projected onto the DEM to compute 3D coordinates in the CRS of the DEM. In our experiments, ray-casting is performed using the ThreeJS JavaScript framework7 utilizing a bounding-volume-hierarchy implementation8 ensuring accuracy and computational eficiency. After projection, the labels are, on average, ∼7 m away from the mounted cameras. The high-precision DEM has the highest label-to-deployment distance (cf. Table 2). Furthermore, the change in a label’s 3D coordinates with a change in DEM accuracy is measured. Table 2 illustrates the impact on distances (between deployment and labels) and the 3D coordinate alone. All values indicate absolute mean values and are computed considering the altitude (3D) and disregarding altitude (2D). Decreasing the DEMs resolution to 20 m and 50 m skews the average distances by an average of 2.84 m and 3.94 m respectively, with the individual errors dispersing widely, as indicated by the standard deviation.
Note that due to the recalculation of the camera’s mounting position for each DEM and the diferences in the terrain data, not all 2629 labels could be projected to 3D locations in all situations (cf. Table 2). Furthermore, projected 3D locations are clipped if they exceed a distance of 100 m from the deployment camera. Missing 3D coordinates are omitted from the calculations.
The error values for our models indicate a logarithmic behavior as model resolution changes, i.e., changing from a 3 m to 20 m DEM introduces considerably more errors than the resolution decrease from 20 m to 50 m. Thus, as the detail level shrinks and resolution drops, the influence of the level of detail on the divergence decreases, resulting in less pronounced errors between models. However, this behavior is likely situational and might break once a model has reached a certain coarseness, at which critical landscape features are lost due to aliasing.
This paper takes the first steps toward estimating animal speeds required for REMs based on labeled instances and digital elevation models captured via airborne laser scanning. Converting these DEMs to 3D geometry allows instance distance computation by projecting their 2D coordinates onto this geometry. The described method presents a promising alternative to current methods by being less restrictive and reducing deployment calibrations to measure the location and rotation of the camera
DEM
3D (Std) 2D (Std) Error in Distances 3D (Std) 2D (Std) 3 m 2422 — — — — — 20 m 2461 trap. However, said method heavily relies on this calibration and the accuracy of the employed DEMs. This work mainly assesses the impact of a DEM’s level of precision utilizing realistic data from in-situ deployed camera traps. The carried-out evaluation demonstrates that lowering model resolution not only distorts results but also results in strongly dispersed deviations from a more densely sampled DEM.
Further, the evaluation processes revealed significant inaccuracies caused by estimating altitude via a height ofset and a DEM. For this dataset, using the same altitude values across all DEMs causes many projection rays to miss the DEM entirely or cameras to fall below the surface for both lower-resolution models. This discrepancy mainly occurs as the deployment locations do not align with the pixel corners within the height texture data. Therefore, camera deployment parameters have been recalculated for every DEM. This outcome emphasizes the need for models that reduce the regional domain as much as possible while staying faithful and precise. Despite some obstacles, this approach shows potential and, with further development, could ofer a more efective solution than alternative methods.
Since an animal’s speed is calculated based on its projected distances over an image sequence, the actual impact of this error on subsequent calculations such as REM is unclear. For example, this error might attenuate if all distances along a sequence share a similar error. Thus, further work should extend the evaluation to include speed estimations and compute the consequent wildlife density, ideally comparing them to independent estimates.
Although the 3 m DEM has clear advantages over its low-res counterparts, an even higher-resolution model with even further focus, possibly on specific deployment scenarios, might be beneficial. Such a context restriction allows a finer resolution and, thus, a more detailed depiction of the local environment. The grid-based airborne laser-sampling method may inherently not be ideal as it is susceptible to aliasingrelated issues, potentially resulting in the loss of crucial features. It could be beneficial to explore using dynamic resolution approaches.
Additionally, focus should be shifted towards applying the approach in-field to ensure it represents an improvement over current methods. This focus involves researching techniques to accurately measure camera trap locations and rotations to minimize the efort connected to manual deployment calibration. Furthermore, the concept of 3D reconstruction may be incorporated to generate detailed 3D geometry per deployment, replacing laser-sampled height data. Such reconstructions may ofer an adequate model accuracy-simplicity balance and can be created based on a short video capture of the respective area. This project is funded by the Austrian Research Promotion Agency FFG (project BAMBI and Equalize2; program numbers: 892231 and 53988223) for which the budget is provided by the Federal Republic of Austria. Image data collection was part of a project funded by ’Wildbach- und Lawinenverbauung GBL Oberes Inntal.’