TemporalSwin-FPN Net: A Novel Pipeline for Metadata-Driven Sequence Classification in Camera Trap Imagery

In wildlife conservation, camera traps generate a significant volume of data that requires extensive manual analysis to extract insights, particularly for species identification. Existing tools and methods for analysing camera trap data typically classify images individually, while the devices are configured to capture multiple pictures in burst mode, often accompanied by metadata. In this paper, we introduce a novel metadata-driven pipeline based on the enhanced Swin Transformer architecture, named TemporalSwin-FPN Net, which utilises metadata to improve classification through a sequential analysis pipeline. This approach shifts the focus from analysing individual snapshots to sequences of images, thereby enriching the architecture with more comprehensive data for species classification. TemporalSwin-FPN Net has achieved a remarkable accuracy of 99.7% on the Australian camera trap dataset, and consistently outperformed the baseline models with 94.89% and 93.32% on the Snapshot Safari dataset for Camdeboo and Mountain Zebra projects, surpassing them by 9.62% and 8.54%. Moreover, integrating the TemporalSwin-FPN Net into our proposed pipeline has significantly reduced the end-to-end processing time for a test dataset of 12,922 images, from 10 minutes 54 seconds to just 6 minutes 41 seconds, thereby substantially enhancing data processing efficiency.