Bioinspired extraction of relative nearness map from expanding optic flow for collision avoidance

Perceiving the depth-structure of the environment is critical for visually guided navigation. It has long been recognized that flying insects rely on translational optic flow to obtain relative nearness map, which measures how soon they would collide with static obstacles. However, the underlying algorithm remains elusive, because in optic flow the map is entangled with viewing angles that require real-time measurement relative to the ego-motion direction. Further difficulties arise from the high inaccuracy of optic flow measured based on biological motion detectors. Here, we propose a network model that realizes the 2D map estimation of the depth ordering of 3D scene structure solely based on the fluctuating optic flow experienced during forward ego-motion. The model shows effectiveness and robustness when validated with real-world video sequences. Further simulations show that the model enables a virtual robot to avoid collision in cluttered environments, even with a narrow field of view. Besides enhancing the understanding of insect depth perception mechanism, the model advances the development of low-complexity and energy-efficient bioinspired algorithms on optic flow-based navigation.