A Review of the Safety Survey of Intelligent Driving

Abstract

With the rapid advancement of artificial intelligence and sensing technologies, intelligent driving systems have transitioned from high‑end model exclusivity to widespread adoption across diverse vehicle segments. This survey reviews three major object detection methods—traditional object detection, object detection based on Yolo, and object detection based on Transformer—and evaluates their respective strengths and limitations in autonomous driving contexts. We summarize representative improvements to R‑CNN and Faster R‑CNN that enhance small‑object detection, occlusion handling, and anchor optimization, as well as YOLO variants that deliver real‑time inference (up to 200 FPS) without sacrificing accuracy. Object detections which are based on Transformer, such as DETR, Deformable DETR, and Swin Transformer are shown to simplify pipelines, capture global context, and improve robustness against sparse or overlapping targets. Beyond algorithms, this paper identify critical challenges in perception latency, model generalization, data privacy, legal compliance, and hardware constraints. Finally, this paper outline future directions—lightweight end‑to‑end architectures via neural architecture search, self‑supervised and federated learning for privacy‑preserving adaptation, and heterogeneous edge‑cloud hardware co‑design—to accelerate the safe, efficient, and scalable deployment of intelligent driving technologies.