NEURAL NETWORK APPLICATION FOR DETECTION OF ROAD ACCIDENTS

Subject of Research.The paper considers the issues of neural network application for detection and prediction of road accidents. The overtaking process of cars with crossing into oncoming traffic is analyzed. The potential possibility of road accident reduction while overtaking is shown owing to intellectual assessment of road situation dynamics development. Method.We proposed to use a two-class classifier based on a neural network. Road situations while overtaking with crossing into oncoming traffic were the objects of classification. The data on them was transmitted to the neural network input in the form of the frame set, that is, a graphical representation of discrete states of the “group of vehicles — section of the road” system. Frame formation was expected to be carried out as a result of information exchange between detectors and vehicle-mounted sensors, and road infrastructure, which is developed within the framework of the “smart city” paradigm. Main Results.The road situation is classified as “Dangerous” in case of high vehicle collision probability while overtaking and “Safe”, otherwise. If the situation is considered as “Dangerous”, the vehicle central processor generates an appropriate effect on the vehicle control elements to prevent an accident. Results of situation simulation implemented on Tensor Flow open software library for machine learning are obtained. They showed high prediction accuracy (0.96) on artificial data set. Practical Relevance. The results of the work can be used in promising unmanned and manned vehicles having radio communication with road infrastructure elements within the “smart city” concept to prevent road accidents caused by dangerous overtaking.