Real-time object detection for autonomous driving based on YOLOv8 and LiDAR fusion

Abstract

Intelligent multimodal perception is essential for the reliability and safety of autonomous driving in complex, real-world environments. This study designs and implements a real-time object detection system that tightly couples visual features from advanced YOLOv8 convolutional neural networks with spatial information extracted from LiDAR point clouds. An adaptive sensor fusion pipeline aligns and merges these heterogeneous data sources through feature- and decision-level integration, enhanced by dynamic weighting based on real-time sensor quality estimation. The system is deployed and evaluated on a synchronized hardware platform that replicates automotive deployment, with experimental protocols covering both standard benchmarks and adverse environmental scenarios. Quantitative results show that the weighted fusion approach notably improves mean average precision and maintains high throughput and low latency even under rain, low illumination, and sensor ambiguity. Analytical modeling of mAP, robustness metrics, and error propagation provides insight into the mechanisms behind the observed performance gains and system stability. Ablation studies further confirm that both the vision and LiDAR modules, as well as the adaptive weighting mechanism, are pivotal for consistent detection and operational resilience. The engineering and experimental findings demonstrate that the fusion strategy developed in this work provides a practical and transferable solution for next-generation autonomous vehicle perception in real-world deployments.