Efficient Traffic sign and Object detection implementation for a Self Driving Car

Autonomous vechiles are revolutionizing transportation, yet their efficacy depends on sophisticated object detection systems capable of real-time performance under resource constraints. This thesis delves into the creation and optimization of a object detection system for a self-driving car, leveraging the YOLO-v5 framework to identify and classify traffic signs, pedestrians, and other vechiles. Through the use of a combination of quantization and downsampling, the system achieves high processing speed and accuracy, while running on a modest CPU computational power of an Intel NUC mini-computer. This thesis shows the application of the RDW Self-Driving Challenge 2024 in which this system was used for the University of Twente team. The results highlight third place in the competition and a system that is capable of stopping before red lights, adhering to the speed limit, letting a pedestrian cross and spotting cars. The discussion highlights the trade-offs between accuracy and processing speeds, the possible impact of different optimization techniques, and the practical considerations and adaptations it would need for the real world environment. The findings provide a valuable groundwork for future teams of the RDW challenge and overall object detection for autonomous vechiles.