A Single-Pathway Biomimetic Model for Potential Collision Prediction

Song Zhang, Guodong Lei, Xuefeng Liang

Abstract: In autonomous driving, predicting the potential collision on road from the first-person view is a very challenging task. Previous study developed a computational model for this task, called LGMD2, which is inspired by a perfect biological visual system, Lobula Giant Movement Detectors, in locusts. LGMD enables locusts to swarm and be free of collision. However, LGMD2 model assumes the moving objects are darker than the background, which rarely happens in real-world scenarios. Meanwhile, its computation structure (ON & OFF pathways) has mutual interference and produces incomplete information. Thus, LGMD2 results in a low prediction accuracy. In this study, we amend the assumption that is more in line with the reality, and propose a novel single-pathway LGMD2 model, SLGMD2. It avoids the interference and generates more complete signal for prediction. To evaluate the effectiveness of SLGMD2, we collect a new first-person view vehicle collision dataset. The experiment on both two public datasets and the newly collected dataset shows state-of-the-art performance of our proposed SLGMD2 model.