Error Analysis and Design Considerations for Stereo Vision Systems Used to Analyze Animal Behavior

This paper presents an analysis of patterns of error in the triangulation of 3D points from stereo camera systems that are used in field work to study the behavior of bats in flight. A measure of the error present in a 3D reconstruction is proposed. A method for empirically testing the performance of a particular stereo camera configuration through a software simulation is presented. Randomly generated 3D calibration points are projected onto the image planes of simulated cameras, and autocalibration is performed using the direct linear transform (DLT) method. The accuracy of the computed 3D reconstruction is determined by computing the proposed measure of geometric error over a grid of reference scene points. A series of experiments are performed with this simulator to evaluate the accuracy of 3D reconstruction with various camera placements and under differing levels of noise. Results of these experiments are used to motivate suggestions for the design and calibration procedure of stereo vision systems used in ecology field work.