Pixel-Level Precision Saccade Control of Arbitrary 3D Spatial Points in Active Binocular Vision Systems

Abstract

We propose a vision-based method to achieve Pixel-Level precision saccadic movements in an active binocular vision system (ABVS). Traditional methods for precise saccadic motion rely heavily on extensive training or accurate kinematic models. However, extensive pre-training reduces the flexibility of ABVS applications, and inaccuracies in the kinematic model can lead to significant errors or failures in saccadic movements. Our proposed saccade control method does not rely on accurate kinematic models. Instead, it utilizes virtual stereo rectification (VSR) and online extrinsic parameters correction (OEPC) to accurately estimate the extrinsic parameters of the binocular cameras. This enables rapid target search based on epipolar constraints using monocular target information, significantly reducing interference from numerous similar targets in the scene and achieving accurate saccades. Experimental results with ABVS demonstrate that our method achieves a 0.994 correct saccade rate and an average effective saccadic motion error of only 3.252 pixels without accurate kinematic models. With accurate kinematic models, the average effective saccadic motion error further decreases to 1.999 pixels when using our method.