Stelmakh O. Methods and models of analysis of transport systems in the conditions of non-stationary parameters of transport flow

The dissertation work is devoted to the development of information technology to increase the accuracy of determining the intensity of traffic based on the analysis of video data in real time in conditions of non-stationary parameters of traffic flow. The basic tendencies of application of information transport systems are covered and modern information transport systems of detection of vehicles are considered. This section discusses the existing transport systems for determining the intensity of traffic. A software component for determining the traffic intensity in two implementations has been developed. The significant advantage of the implementation based on the bioinspired module over the implementation based on the comparison of two frames has been experimentally proved. The software component for determining the intensity of traffic, which is developed, is an integral part of the information system of traffic management. The technology of determining the intensity of traffic using the neural network U-net, which has high efficiency and effectiveness in image segmentation and is known for reliability when working with large data sets. U-net methods were used - encoding and decoding methods for merging basic information and high-level information. To train the neural network, the SGD optimization algorithm was used to analyze the gradient of objects. The input images were cut into 128 by 128 pixel segments. A data set of 10,000 images was used to train the U-net neural network. The advantage of using the U-net neural network for the problem of determining the traffic intensity and TLCR load index has been experimentally proved. The technology of determining the intensity of traffic based on the video data coming from the video surveillance camera has been developed. The algorithm for determining the load index of the transport section of the TLCR has been improved, making it possible to take into account only cars moving in the studied lane. The developed method of determining traffic intensity based on successive values ​​of congestion has the following advantages over other similar systems: data processing speed, accuracy, no need for additional equipment (eg sensors) and low cost. An algorithm and information system for long-term forecasting of the traffic load index of the TLCR transport section for further assessment of the traffic condition have been developed. The forecasting information system is based on a learning model with a recurrent LSTM neural network. The developed system is trained on a training video obtained from traffic cameras recorded for one week to predict the load of the TLCR for each day of the week. An algorithm for detecting traffic jams based on the load index of the TLCR transport section obtained from images obtained from video cameras installed in different parts of the city has been developed. The algorithm allows to classify traffic jams according to three levels of congestion: low, medium and high congestion. The developed algorithm was investigated on the basis of experimental data of recordings from road video cameras and satisfactory results of detection and classification of load levels were obtained. The sequence of data processing and transformations constitute a new technology for determining the intensity of traffic, which provides high accuracy in estimating the intensity of traffic on the road. The developed technology can work in conditions of non-stationary parameters of traffic. Due to the use of segmentation instead of classification, as well as a specific set of data for training, the technology is devoid of such shortcomings as incorrectly selected angle and the lack of a vehicle in existing databases for training. The proposed system successfully calculated vehicles with high accuracy - the average values ​​of F-measure and accuracy (Accuracy) reached 0.9967 and 0.9935, respectively. The accuracy of the developed information system of long-term forecasting of the load index of the transport section TLCR, which is based on the model of training with a recurrent neural network LSTM, is investigated. The obtained experimental average accuracy of 0.914 for five days (two days off and three working days) demonstrates high efficiency of forecasting results.