Controlling traffic signaling time using optimized kernel extreme learning machines

In recent years, urbanization and the increase in the number of vehicles have led to an increase in traffic congestion. Therefore, developments in transportation systems seeking solutions to this problem have accelerated, and many solution proposals have been made. However, many of these proposals fail to fully respond to the dynamics of traffic flow. Due to the inadequacy of static traffic management, the importance of intelligent transportation systems has increased, and researchers have increased their efforts towards real-time traffic management models to address the variable and multi-parameter nature of traffic flow. The aim of this thesis is to reduce traffic congestion and make traffic flow smoother. In line with this goal, a combination of deep learning and Extreme Learning Machines (ELM) with a single hidden layer feedforward artificial neural network has been used instead of traditional fixed-time management models used in traffic signalization systems. The study was conducted in four steps. In the first step, vehicles in video images of real traffic were detected and classified using the YOLOv8 deep learning algorithm. In the second step, the features of the trained deep model were extracted and provided as input to the Kernel Extreme Learning Machine (KELM). Here, the inner product of the extreme learning machine was replaced with a kernel function to model the complex structure of the dataset and improve the accuracy of traffic flow predictions. Particle Swarm Optimization (PSO), Grey Wolf Optimization Algorithm (GWO), Genetic Algorithm (GA), and Artificial Bee Colony Algorithm (ABC) were used for the optimization of KELM in the third step. Following this step, it was determined that the Genetic Algorithm was more suitable for our problem and dataset compared to other algorithms. The fourth step involves counting the detected vehicles and simulating an adaptive traffic management model that could be created based on the number of vehicles. In the simulation prepared using the Python Pygame library, traffic at a 4-way isolated signalized intersection is animated. In the real-time camera image, vehicles waiting at red lights are detected to obtain vehicle counts, and the duration of the green light is determined using the values assigned to these vehicles. Thus, an adaptive traffic control system tailored to real traffic density, vehicle count, and even vehicle type has been created.