A photogrammetric approach to the estimation of distance to animals in camera trap images

The ability to assess accurate distance measurements, meaning the distance from an animal to the observer, is critical for wildlife monitoring methods such as distance sampling. This process has typically been done through manual measurements or using reference points, making fieldwork and analysis labor intensive and error prone. Using Structure from Motion (SfM) photogrammetry and geotagged imagery captured from smartphones offers a simple and cost-effective solution to distance estimation from camera traps. This research evaluates the potential of SfM to create 3D models of the areas within a camera trap field-of-view to produce depth maps that can be used to automate the distance estimation process from bounding box-labeled imagery. These methods have the potential, when paired with deep learning algorithms or crowdsourcing, to automate the process of wildlife identification and distance estimation, reducing the need for manual fieldwork or analysis while increasing the precision and replicability of population estimates. To test the accuracy of these methods and evaluate best practices, three types of GNSS integration into the SfM model were assessed at three camera trap sites placed within the Sevilleta National Wildlife Refuge in New Mexico, USA. The evaluation focused on the geometric accuracy of the resulting model, quantified using Root Mean Square Error (RMSE) against known ground control points, the accuracy of distance estimates relative to field-measured distances, and the comparison of modeled distances of detected animals to ground control points. The first method used only the GNSS data embedded (geotagged) cell phone images (ImageGNSS), the second used ground control points (GCPGNSS), and the third used both GNSS sources (AllGNSS). Models created using ImageGNSS had the highest accuracy, with an overall RMSE of 1.25 pixels and a mean absolute error (MAE) of 1.76 m for distance estimation from GCPs. A comparison of the modeled distance for 10 animals to known distance points produced minimal error (MAE of 3.31 m), demonstrating the potential of photogrammetric approaches to make accurate distance estimations from camera trap imagery. Unlike other digital distance estimation techniques, error rates do not increase significantly ( p = 0.36) at distances greater than 15 m to at least a distance of 35 m. The biggest driver of distance error was scene complexity (complex topography, dense vegetation, etc.) • Developed a novel method of automated wildlife distance estimation using trail cameras and SfM 3D modeling. • Evaluated three different types of georeferencing to determine field applicability and model accuracy. • Found that geotagged imagery improved created depth maps and increased the accuracy of wildlife measurements.