Boosting Collision Perception Against Noisy Signals with a Probabilistic Neural Network

Accurate perception of collision acts as the basis of artificial vision systems for autonomous collision detection and avoidance. The lobula giant motion detector (LGMD) in the locust's brain is a natural and simple detector responding most strongly to images of approaching objects, which has been extensively studied and modeled. However, noisy signals greatly obstruct the LGMD models upon reliability and robustness. To address this problem, the existing models incorporate additional pre-processing methods or network layers which, nevertheless, further increase the computational complexity of LGMD models. On the other hand, the signal processing within neurons and synapses has been demonstrated as non-deterministic, and this may help neural computation resist noise. This raises the question that whether LGMD incorporated with indeterminacy realized by probabilistic mechanism will improve its anti-noise performance. This paper investigates this point with a probabilistic LGMD model, an attempt to conform to biological facts. The probabilistic model introduces indeterminacy to signal operations including transmission, interaction, integration between multiple layers of a classical LGMD neural network. The proposed model has been systematically tested with a range of stimuli including simple scenes with artificial noise, and real-world complex scenes. We intriguingly found that the performance of LGMD on the aspect of combating noise has been significantly improved, demonstrating that adding a probabilistic mechanism is an easy but effective way of fighting noisy signals in collision perception.