Reliable and efficient integration of AI into camera traps for smart wildlife monitoring based on continual learning

In this paper, we comprehensively report on an efficient approach for the integration of artificial intelligence (AI) processing pipelines in camera traps for smart on-site wildlife monitoring. Our work covers hardware, software, and algorithmics. We have built two prototypes of smart camera trap on a maximum bill of materials of 100$. We have also built two datasets, made publicly available, comprising over 17 k images, many of them notably challenging even for humans. Leveraging our broad expertise on embedded systems, specialized software libraries and toolchains, and AI techniques such as transfer learning, explainable AI, and, most importantly, continual learning, we achieve more reliable inference on-site - specifically 10 % higher F1-score - than MegaDetector run off-site on a desktop computer. The paper includes many practical details on system realization and on-site training in addition to a vast set of lab and experimental results. • Implementation of a completely functional 100$ smart camera trap prototype. • Collection of over 17 k real-world images, publicly available. • Code developed for smart camera operation and on-device training publicly available. • 10 % higher F1-score of our prototype, achieved on-site, than desktop MegaDetector. • Realization of on-site continual learning based on real and synthetic images.