A vision based intelligent transportation system for smart cities

• Main Author: Liu, Yizhe
• Other Authors: Mohammed Yakoob Siyal
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/168265
• Institution: Nanyang Technological University

With the increase in the number of vehicles worldwide, traffic congestion has become a widespread issue. To manage traffic problems, several systems have been designed. However, each system has its advantages and disadvantages. This project presents the development and implementation of an intelligent transportation system that utilizes image processing algorithms to obtain crucial traffic data information, for the purpose of improving road management, reducing congestion, and enhancing road safety for all users. The system includes a graphical user interface implemented in MATLAB that enables the extraction of essential traffic data such as vehicle types, vehicle count, vehicle speed, overall road usage and traffic incident occurrence. A total of fourteen image segmentation methods, three image enhancement methods and two classification methods have been implemented for the system. Data analysis was conducted to compare the performance of different methods in various environmental conditions such as weather conditions and shooting angles. The results showed that the accuracy of the vehicle count algorithm was affected by environmental conditions, with higher accuracy achieved in sunny condition compared to rainy and snowy conditions due to the presence of noise resulted by raindrops or snowflakes. More advanced and complex segmentation methods such as ViBe, LoG, Canny, and zero crossing performed better than relatively simple ones, such as quadtree decomposition and binary image conversion. The neural network-based classification method maintained its accuracy of vehicle types algorithm even in unusual viewing conditions, while the pixel count comparison-based method performs unsatisfactory results because it had difficulty adjusting threshold values for each lane based on the actual condition. The author has provided possible explanations of the results achieved. Recommendations for future research is also proposed to address the direction of improvement.