Abstract:
This paper addresses 3D trajectory estimation of simulated fruit flies assuming a time-synchronized and calibrated 3-camera system. Because the objects have almost the same appearance, both stereo matching and temporal tracking are challenging. In this paper, a third camera is introduced to verify matching and tracking results based on epipolar geometry and projection consistency. This reduces the ambiguity, fetches missed matches and corrects incorrect matches during tracking. Since matching and tracking affect each other, we process both in interaction instead of separately. Unscented Kalman filters are adopted to track objects by modelling motion information, as no distinguishing appearance features are available.