An Automated Pipeline for 3D Flight Path Reconstruction from Multi-Channel Audio of Echolocating Bats

Sammanfattning

This thesis presents an automated pipeline developed for reconstructing the 3D flight trajectories of bats navigating a cave environment purely from multi-channel audio recordings. The pipeline was developed and validated using the Ushichka dataset which contains multiple simultaneous audio and video recordings of bats in their natural habitat, the Orlova Chuka cave in Bulgaria. The proposed pipeline integrates computer vision, signal processing algorithms, and robust multilateration solvers. Echolocation calls are visually detected from audio spectrograms using a trained SqueezeNet Convolutional Neural Network. Time Differences of Arrival across the microphone array are then estimated via Generalized Cross-Correlation with Phase Transform. To robustly estimate 3D call positions and timing in the presence of anomalous data, a minimal multilateration solver is coupled with a Random Sample Consensus framework. Finally, a multi-hypothesis tracking method associates these discrete coordinates into continuous trajectories. Processing of the acoustic data successfully yielded numerous reconstructed trajectories. Validating the acoustic reconstruction against ground-truth trajectories constructed from thermal cameras, the mean absolute orthogonal distance between them is 3.6 cm. This result demonstrates the usefulness of the approach in studying the flight of bats, and, by extension, their group behavior.