Application of Neural Networks in Autonomous Driving Vehicles to Enhance Controllability of Lateral Tip-Over Stability Hazards

Abstract

Background. In case any hazardous situation occurs during driving related to risks of lateral tip-over, the driver is often capable of controlling the situation before an accident occurs using human sensory information. Autonomous vehicles however, must rely on technical sensors, logic and actuators to achieve similar controllability and would require exact knowledge information related to car dimensions and most importantly, the exact location of the center of gravity at all times to be able to detect any hazardous situations. Aim. This paper intends to contribute to the safety of autonomous vehicles by investigation, modelling and implementation of classification mechanisms for tip-over hazards using human sensory information, in particular sound patterns. Method. A generic car model was defined suitable of simulating various lateral tip-over hazardous scenarios and corresponding variations in car tire noise patterns that were fed to neural network controllers trained in advance with sinusoidal patterns to detect lateral tip-over hazards. Results. The system was able to detect most lateral tip-over hazardous situations resulting from cornering, driving on a road bank successfully using only car tire noise patterns. Conclusions. The initial findings show a potential for further research in the field of controllability of autonomous vehicles based upon human-like sensory information, in particular observable noise patterns, rather than exact technical models and parameters or expensive force sensors.