Realtime Hardware System Design for Collision Detection Based on LGMD2 Visual Interneuron

In this paper, we designed a realtime collision detection system using a visual neural network including LGMD2 of insects. The algorithm was designed to be hardware-friendly, and a pipeline structure was applied through critical path analysis. The exponential function was replaced with a linear approximation through error analysis to satisfy complexity and accuracy. By modifying the algorithm in the on/off channel, the space for storing the intermediate values of one frame size was reduced. Through performance analysis, errors were minimized and appropriate bit widths for each signal were determined. The simulation shows that collisions were detected at the correct time when five artificial and real images were applied. The proposed architecture was synthesized using Dongbu HiTek