SOCRATES: Introducing Depth in Visual Wildlife Monitoring Using Stereo Vision

Haucke, Timm; Kühl, Hjalmar S.; Steinhage, Volker

doi:10.3390/s22239082

Cited by 3 publications

(2 citation statements)

References 59 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In addition, the high volume of blank images generated by this particular setting can be analyzed using automatic image classification software like MegaDetector (Fennell et al., 2022), Wildlife Insights (Ahumada et al., 2020), or Agouti (Casaer et al., 2019). Finally, measuring the distance between the camera and the detected animal, a critical measurement used to accurately estimate the effective detection area of cameras, is becoming more feasible and less time consuming thanks to the recent development of automatic approaches to measure distances from images and videos (Haucke, Kühl, Hoyer, & Steinhage, 2022; Haucke, Kühl, & Steinhage, 2022; Johanns et al., 2022). We highlight the importance of making these programs easy to use and accessible to conservation programs around the globe to facilitate the proper application of the STE and other viewshed density estimators.…”

Section: Discussionmentioning

confidence: 99%

Estimating animal density using the Space‐to‐Event model and bootstrap resampling with motion‐triggered camera‐trap data

Lyet

Waller

Chambert

et al. 2023

Remote Sens Ecol Conserv

View full text Add to dashboard Cite

Over the past few decades, the use of camera‐traps has revolutionized our ability to monitor populations of wild terrestrial mammals. While methods to estimate abundance from individually‐identifiable animals are well‐established, they are mostly restricted to species with clear natural markings or else necessitate invasive and often costly animal tagging campaigns. Estimating abundance or density from unmarked animals remains challenging. Several models recently developed to deal with this issue are promising, but are not widely used by field ecologists. Here, we developed a framework for applying the Space‐To‐Event (STE) model—originally designed to be used with time‐lapse images—on motion‐triggered camera‐trap data. Our approach involves performing bootstrap resampling on the photographic dataset to generate multiple datasets that are then used as input to the STE model. We tested our approach on 29 datasets, including 17 ungulate species from eight sites, in six different countries and various ecosystems. Then, we conducted a regression analysis to evaluate how variations in ecological and sampling conditions across studies affected the bias and precision of our STE density estimates. Our study shows that with a bootstrap resampling approach and information on animal activity and effective detection distances to animals, the STE model can be used to analyze motion‐trigger datasets and provide population density estimates that are similar to those from other methods. We found that measuring the camera viewshed was critical to prevent major negative biases in density estimates. Moreover, using a 1‐s sampling window was important to avoid the positive bias that results from violating the instantaneous‐sampling assumption. We found that precision increased with greater sampling effort and higher density populations. Based on these results, we highlight several issues from past studies that have applied the original timelapse‐based STE to motion‐trigger datasets, issues that our bootstrap resampling approach addresses. We caution that the STE model, whether applied to timelapse or motion‐triggered datasets, relies on strict assumptions. Any violations of these assumptions, such as non‐instantaneous sampling or the application of angle and distance of detection provided by the camera manufacturer, can cause biases in multiple directions that may be difficult to differentiate.

show abstract

Section: Discussionmentioning

confidence: 99%

Estimating animal density using the Space‐to‐Event model and bootstrap resampling with motion‐triggered camera‐trap data

Lyet

Waller

Chambert

et al. 2023

Remote Sens Ecol Conserv

View full text Add to dashboard Cite

show abstract

“…Next, we describe the object detection algorithm, where the segment anything model (SAM) is adopted. As illustrated in Figure 2, the object detection algorithm is consisted of three parts, namely, an image encoder, a prompt encoder, and a mask decoder [28][29][30]. First, the original image is processed by the image encoder, where the characteristics of the original image are extracted.…”

Section: Object Detectionmentioning

confidence: 99%

Intelligent Algorithm for Rock Core RQD Based on Object Detection and Image Segmentation to Suppress Noise and Vibration

Xiong,

Wang,

Zhang

et al. 2024

Advances in Civil Engineering

View full text Add to dashboard Cite

In the construction of the civil engineering infrastructure, the noise and vibration are affected by the geological conditions, adopting specific construction techniques based on the geological conditions is of great significance in suppressing the noise and vibration caused by the construction. To classify and evaluate the rock mass quality, the rock quality designation (RQD) is adopted widely in the geological and mining engineering. Traditionally, to obtain RQD, lengths of drilling core pieces are measured and RQD is calculated manually, which is labor-expensive and time-consuming. With the development of the computational power, the image treatment driven by the computer vision creates a potential approach to obtain RQD automatically. In the present work, the image treatment process with the aid of the object detection and the image segmentation is adopted to obtain RQD automatically, based on the similarity of features such as color and texture, the segment anything model is adopted to detect the rock cores in the image, then, the YOLOv8 algorithm is adopted to train the model, and the gap features of the rock chip segments are extracted for segmentation of different rock core segments. To test the performance of the proposed approach, 10 boreholes from Shapingba Railway Comprehensive Reconstruction Project are adopted to conduct the case study. Compared to the traditional manual approach, RQD obtained by the proposed approach is relatively accurate and obviously efficient, namely, the average error is less than 5% and the time consumed is less than 70%.

show abstract