Shtykalo O. Electronic System for Urban Traffic Monitoring

The dissertation is devoted to developing an electronic urban traffic monitoring system based on passive Bluetooth Low Energy (BLE) scanning and intelligent data-processing methods. The relevance stems from the growing load on urban transport systems and the need for approaches that balance informational value, moderate infrastructure cost, and minimization of interference with personal data. In contrast to video surveillance with object detection or the processing of personalized navigation data, the proposed approach relies on registering technical events of Bluetooth-device appearance within the observation range of scanning nodes. It is established that direct use of BLE scanning entails a methodological problem: a single MAC address does not correspond to a single road user, since one participant may carry several devices, which causes systematic overestimation of traffic. Eliminating this redundancy is defined as the central task of the work. A structural and functional organization of the system is proposed, in which a BLE node is treated as a source of a standardized event stream of the form "MAC address – time – node ID," separating the hardware acquisition layer from the intelligent-processing layer. The main methodological contribution is a MAC-address grouping model based on a Siamese neural network with an LSTM encoder, which, through metric learning, estimates the probability that two MAC addresses belong to the same road user, replacing unsuitable multiclass classification with a pairwise-affiliation formulation. Experimental evaluation is performed in the SUMO simulation environment on a road network derived from OpenStreetMap. The LSTM4 model shows the best performance: the normalized root mean square error of traffic reconstruction is 10.64%, compared with 32.96% for the heuristic method and 63.35% for raw data. For optimizing BLE-system placement, a multiobjective problem (number of nodes and aggregate monitoring error) is formulated and solved with the NSGA-II algorithm. Starting from 3,360 candidate nodes, a solution with 115 systems is obtained (infrastructure reduction of about 96.6%), with εgrp = 0.093, a non-covered fraction of 1 − c = 0.116, and an aggregate error of ε = 0.198. Adaptation of placement solutions to new areas is implemented using a graph neural network (grouping error 0.071; non-covered 0.126; aggregate error 0.188). The numerical results pertain to simulation-based conditions and are not equated with full-scale field measurements.