Object Detection in Self-Driving Vehicle Using CARLA Simulator and YOLOv8

Abstract

This study develops an object detection system for autonomous vehicles using the YOLOv8 model integrated with the CARLA simulator. The research addresses gaps in multi-camera setups and real-time detection by training YOLOv8 on a custom dataset generated from CARLA simulations. Results show high performance with a mean Average Precision (mAP@50) of 0.990 in single-camera configurations, outperforming multi-camera setups due to computational constraints. While Proportional-Integral-Derivative (PID) and Model Predictive Control (MPC) integration was explored conceptually, empirical evaluation revealed enhanced navigation accuracy in simulated scenarios, though real-world validation is needed. The work highlights trade-offs between accuracy and speed, suggesting optimizations for practical deployment.