Identification of animals and recognition of their actions in wildlife videos using deep learning techniques

Biodiversity crisis has continued to accelerate. Studying animal distribution, movement and behaviour is of critical importance to address environmental challenges such as spreading of diseases, invasive species, climate and land-use change. Camera traps are an appropriate technique for continuous animal monitoring in an automated 24/7/52 documentation. This study shows a proof-of-concept for an end-to-end pipeline to detect and classify animals and their behaviour in video clips. Video clips are captured with 8 frames per second by camera traps using infrared cameras and infrared flash-lights. The clips show deer, boars, foxes and hares - mostly at night time. Our approach shows an average precision of 63.8% for animal detection and identification. For action recognition the achieved accuracies range between 88.4% and 94.1%.