Detecting animals in challenging outdoor environments

Abstract

Monitoring wildlife is vital for sustainable coexistence between humans, flora, and fauna. Do we have a healthy population? Is an animal near extinction, or did new species arrive in the habitat? Where do the animals stay, hunt, and live? Forest rangers, livestock owners, and other interested people place camera traps to answer those questions. Those camera traps automatically take photos if motion is detected, which easily leads to many images to be analyzed. The quality of those photos often varies and depends on the environment. Animals usually do not stand directly close in front of the lens but are partly covered by the fauna or only barely visible in the distance. Furthermore, precipitation, fog, dirt, and many other conditions additionally decrease the image quality and hinder the detection of animals. Our interest comes from our mAInZaun project. The idea is to place cameras on the poles of pasture fences. Deterrents are automatically activated using artificial intelligence if predators like wolves, golden jackals, stray dogs, etc., are detected. Ideally, this approach should reduce the predator attacks on grazing animals in a non-lethal way. The automatic activation of the deterrents requires a good detection of predators. Both, false alarms and undetected predators, must be prevented and require a good detection model. Training a model detecting animals requires a certain amount of training data in the form of labeled images. The interest is not only in the clearly visible animals but especially in the hard-to-see ones. Furthermore, the model has to be adapted occasionally to ensure good performance in different environments or if new species have to be detected. This work addresses those challenges with our ShadowWolf framework, which offers an assisted approach for automatically labeling camera trap images. Using state-of-the-art machine-learning algorithms in combination with internet-based crowdsourcing significantly increases the detection of animals and reduces the workload for individual domain experts simultaneously. The outcome is a training dataset that can be used to train arbitrary object detection and classification algorithms. The user can select the appropriate model to run on different devices -- from constrained edge devices to a high-end server. We also collected and analyzed our dataset containing more than 100,000 camera trap images to evaluate our ShadowWolf framework. We also discuss deploying a sensor network tailored for our agriculture application. In the end, this work benefits everybody dealing with real-world wildlife camera trap images: Counting animals, detecting predators, and optimizing automatic machine learning models are simplified and also usable by non-technical people.