Graph-based Extraction and Tracking of Insect Motion in Event-Camera Data Streams

Monitoring pollinator populations requires accurate reconstruction of individual insect flight trajectories from continuous sensor data. Event-based vision sensors capture motion at microsecond-resolution, producing sparse data streams well suited to detecting fast-moving insects. We present a two-stage pipeline that extracts flight trajectories from event camera recordings. In the first stage, a graph neural network (GNN) trained entirely on synthetic data performs semantic segmentation to separate insect-triggered events from background activity, generalizing to real-world recordings with an F1 score of 0.95. In the second stage, graph-based instance segmentation reconstructs individual trajectories from the segmented events through iterative splitting of over-merged paths and re-merging of fragmented tracks. With post-processing parameters tuned via Bayesian optimization, the full pipeline achieves a mean average precision (mAP) of 0.60 and mean average recall (mAR) of 0.79 on real-world data, up from 0.17 and 0.56 before refinement, while preserving accurate insect counts. These results demonstrate that event-based cameras combined with deep graph learning enable scalable, high-resolution monitoring of insect activity in natural environments.