Fly visual system inspired artificial neural network for collision detection

This work investigates one bio-inspired collision detection system based on fly visual neural structures, in which collision alarm is triggered if an approaching object in a direct collision course appears in the field of view of a camera or a robot, together with the relevant time region of collision. One such artificial system consists of one artificial fly visual neural network model and one collision detection mechanism. The former one is a computational model to capture membrane potentials produced by neurons. The latter one takes the outputs of the former one as its inputs, and executes three detection schemes: (i) identifying when a spike takes place through the membrane potentials and one threshold scheme; (ii) deciding the motion direction of a moving object by the Reichardt detector model; and (iii) sending collision alarms and collision regions. Experimentally, relying upon a series of video image sequences with different scenes, numerical results illustrated that the artificial system with some striking characteristics is a potentially alternative tool for collision detection.