Traffic Flow Intensity Research Based on Deep Learning

Abstract

In a harmonious transport system, traffic flows are rationally distributed depending on the capacity of roads and streets to ensure transit capacity, considering the traffic light control systems. At the same time, due attention is not paid to changes in weather and natural conditions, which in turn significantly adjusts driving regimes, taking them out of a stable, predictable state. Modern software and hardware systems and information resources of large cities have a wide range of recorded indicators that affect distribution of traffic flows. Their automated processing using algorithmic machine learning tools has formed a comprehensive understanding of the patterns of change in the traffic intensity indicator, which is a new stage of improving road traffic safety, striving for zero mortality. The scientific novelty of the study refers to the techniques and approaches to studying the weather and climate characteristics and factors of the street-and-road network, their preliminary processing using modern statistical and logical methods of normalisation and eliminating random outliers. The deep learning method opens wide opportunities for analysing the intensity of the road traffic flow. By processing large amounts of data, such algorithms are able to identify complex patterns and relationships, which improves traffic forecasting and optimises traffic management. For correct operation of the neural network for training the model and studying the road traffic flow intensity, a set of software tools for preliminary data processing has been developed, which includes a step-by-step analysis of array structures with subsequent replacement of values or elimination of errors. Preliminary data cleaning in accordance with the syntax of the program logic and the rules of statistical analysis is followed by application of a method for searching and eliminating anomalies was used, i.e. the isolation forest method. This research direction was part of a large study on road traffic flow intensity, and the described results are a set of solutions based on the system interaction of software and methods of statistical and analytical transformations developed by the authors.