Visual motion pattern extraction and fusion for collision detection in complex dynamic scenes

Detecting colliding objects in complex dynamic scenes is a difficult task for conventional computer vision techniques. However, visual processing mechanisms in animals such as insects may provide very simple and effective solutions for detecting colliding objects in complex dynamic scenes. In this paper, we propose a robust collision detecting system, which consists of a lobula giant movement detector (LGMD) based neural network and a translating sensitive neural network (TSNN), to recognise objects on a direct collision course in complex dynamic scenes. The LGMD based neural network is specialized for recognizing looming objects that are on a direct collision course. The TSNN, which fuses the extracted visual motion cues from several whole field direction selective neural networks, is only sensitive to translating movements in the dynamic scenes. The looming cue and translating cue revealed by the two specialized visual motion detectors are fused in the present system via a decision making mechanism. In the system, the LGMD plays a key role in detecting imminent collision; the decision from TSNN becomes useful only when a collision alarm has been issued by the LGMD network. Using driving scenarios as an example, we showed that the bio-inspired system can reliably detect imminent colliding objects in complex driving scenes.