Navigating the Future: Advancing Autonomous Vehicles through Robust Target Recognition and Real-Time Avoidance

Abstract

This research addresses the critical challenge of enhancing object recognition and real-time response capabilities in autonomous vehicles (AVs) under varying simulated conditions, which is crucial for ensuring both navigational safety and operational efficiency. Utilizing the Carla 0.9.14 simulator and Unreal Engine 4.26 on Ubuntu 20.04, we focus on improving the detection and classification of key on-road obstacles—vehicles, pedestrians, and cyclists—using the YOLOv7 object detection algorithm. By integrating advanced sensory technologies, specifically stereo vision cameras and LIDAR, we create a dynamic testing environment that simulates diverse urban and rural scenarios. Our methodology enhances the YOLOv7 algorithm's accuracy and speed through extensive training on a meticulously curated dataset of 4,113 images, reflecting a broad spectrum of environmental conditions, including varying lighting and weather conditions. This rigorous approach yielded a significant increase in mean average precision (mAP) from 64.3% to 76.3%, and enhanced the algorithm's reliability, with notable improvements over previous models. The research delineates a clear advancement in AV technology by demonstrating substantial improvements in object detection metrics, contributing foundational insights for future implementations in real-world settings and supporting the further development of real-time avoidance systems. This study not only progresses the field of AV but also sets a new benchmark for object detection performance, aligning with industry and academic goals to optimize AV systems for complex and unpredictable driving scenarios.