Towards an Event-Based Stereo Vision System for Water Flow Estimation

Accurate estimation of water flow dynamics poses significant challenges for conventional frame-based vision systems due to low temporal resolution, motion blur, and extreme specular reflections. This thesis presents a novel, fully event-based stereo vision system specifically engineered for the 3D kinematic tracking of tracer particles and volumetric flow estimation. Leveraging the asynchronous nature, microsecond latency, and high dynamic range of neuromorphic sensors, the proposed system inherently bypasses the optical saturation issues of standard cameras. The core contribution is a robut offline perception pipeline designed to solve the unique spatial and temporal assignment problems of asynchronous data. Specifically, it introduces a custom dipole bond manager for bipartite polarity matching, seamlessly integrated with a 9-state constant acceleration Kalman filter and strict epipolar constraints. Validated through controlled synthetic simulations and deployed over a physical hydrodynamic setup, the system successfully reconstructs dense 3D lagrangian trajectories and comprehensive eulerian vector fields. The results demonstrate the extreme precision of event-based stereoscopy, providing a highly robust analytical framework for high-frequency volumetric flow measurement.