Researchers have developed a computer vision model to automatically detect wild animals on highways, potentially alerting drivers in real-time through smartphones or car screens, similar to traffic congestion warnings. This technology aims to prevent accidents by providing timely notifications without human intervention.

Wildlife roadkill is a persistent and hazardous problem that impacts both humans and animals and has garnered growing concern from conservationists around the globe. Tackling this issue is challenging because it necessitates substantial investments in road infrastructure to properly mitigate animal-vehicle incidents.

The success and widespread adoption of low-cost and economically viable detection systems, such as those that alert drivers about the presence of animals and collect statistics on endangered animal species, heavily rely on the availability of data for system training despite recent applications of machine learning techniques. Inadequate training data has a detrimental effect on extracting features in machine learning models, which is essential for accurately detecting and classifying animals.

This work assesses the efficacy of various cutting-edge object detection models when trained on restricted data. The chosen models are derived from the YOLO architecture, which is very suitable and widely employed for real-time object identification. The models encompassed in this list are YoloV4, Scaled-YoloV4, YoloV5, YoloR, YoloX, and YoloV7. The researchers concentrate on studying endangered animal species in Brazil and utilize the BRA-Dataset to train their models. In addition, they evaluate the efficacy of data augmentation and transfer learning strategies in their assessment.

Related contributions

Ecology researchers face challenges in identifying and measuring species density. Traditional methods involve slow and costly processes, such as using camera traps over several days. In a study by de Arruda et al., they proposed using Convolutional Neural Networks (CNN) to automatically detect and identify animal species from the Pantanal using thermal and RGB images. 

Likewise, Schneider et al. utilized deep neural networks for object detection tasks in camera trap images, aiming to identify, count, and locate animals despite various challenges like occlusions and lighting variations. 

Similarly, Biswas et al. compared CNNs for detecting bird species in Bangladesh and India, where manual classification of over 800 species could be more practical. They trained seven species using transfer learning on various models and evaluated accuracy, precision, recall, and F1-score. 

Likewise, Adami et al. compared CNNs to create a solution combining edge and cloud computing with computer vision to deter animals like wild boars and deer from agricultural areas. 

To create a species recognition system, the researchers initially constructed a database of mammals susceptible to vehicle strikes. They accomplished this by identifying and retrieving 1,823 photos from the internet that are freely available for public use and do not have copyright protection. If deemed required, the photographs underwent editing to eliminate "noise" (random fluctuations in colour, brightness, etc.) that might impede species identification or to aid identification by including a range of perspectives.

Subsequently, the researchers conducted experiments on various iterations of YOLO (You Only Look Once), a computer vision technique extensively employed for instantaneous identification of objects, including wildlife. An advantage of single-stage detection is its suitability for real-time identification of large animals, prioritizing speed above accuracy. Another determining aspect of the decision was the feasibility of operating the system on edge devices with minimal resources, such as tablets and portable laptops. 

Furthermore, the researchers utilized videos of animals captured within São Carlos Ecological Park to evaluate the technique's effectiveness. Subsequent database additions will incorporate animal photographs from forest camera traps and roadway cameras.

Source: https://www.nature.com/articles/s41598-024-52054-y

Image source: Unsplash