Abstract

Society must always keep pace with the development of science. Due to the recent advances in surveillance technology, especially in the field of deep learning, the concept of smart cities is closer to be realized than ever. An important aspect of smart cities is smart traffic light control system that can keep up with the enormous increase of vehicles. Also, these systems are necessary as more autonomous cars are produced.

The work in this thesis proposes a system that uses deep learning to control traffic light systems to enable smart cities. The system uses deep learning-based object detection in order to detect different types of vehicles on busy intersection. A YOLOv4-tiny model is trained using Darknet framework in Google Colab. The detector achieved a mAP accuracy of 95.6%.

The dataset used in this research is the BIT-vehicles public dataset, that includes six classes of vehicles taken in different times during day and night. The model is then be implemented on PYNQZ2 FPGA. This development board from Xilinx support python coding through jupyter notebook. The board also has the capability to be connected to a real-time camera to feed the system with live input from different busy intersections. Depending on the result of the detector, the traffic light system can be controlled. When objects are detected, the light is turned red while turning the opposite side’s light to green. This way a car does not have to wait at a red light when there is no object coming from the other direction. Implementing such smart system can reduce traffic jams at busy intersections. As a result, saving time and resources in cities.

The work presented an AI-optimised solution that is cheap to implement on FPGAs and it does not involve extra hardware devices such as a computer or a radar system. This can also play a key role in improving mobility and reducing waiting time as well as the emissions caused by polluting gases.

Smart Traffic Light Controller Towards Smart Cities