Near range pedestrian collision detection using bio-inspired visual neural networks

New vehicular safety standards require the development of pedestrian collision detection systems that can trigger the deployment of active impact alleviation measures from the vehicle prior to a collision. In this paper, we propose a new vision-based system for near-range pedestrian collision detection. The low-level system uses a bio-inspired visual neural network, which emulates the visual system of the locust, to detect visual cues relevant to objects in front of a moving car. At a higher level, the system employs a neural-network classifier to identify dangerous pedestrian positions, triggering an alarm signal. The system was tuned via simulation and tested using recorded video sequences of real vehicle impacts. The experiment results demonstrate that the system is able to discriminate between pedestrians in dangerous and safe positions, triggering alarms accordingly.