Currently, autonomous flight control for unmanned aerial vehicles (UAVs) has become increasingly critical in remote-sensing applications, such as high-resolution data acquisition and road disease detection. However, this task also faces significant challenges, particularly the weak GNSS signals in flight areas and the complex flight environment. Furthermore, many existing autonomous-flight-control algorithms for UAVs are computationally demanding, which limits their deployment on embedded devices with constrained memory and processing power, thereby affecting both operational efficiency and the safety of UAV missions. To address these issues, we propose PISCFF-LNet, a lightweight road-extraction network that integrates prior knowledge and spatial contextual features. The network employs a dual-branch encoder architecture to separately extract spatial and contextual features, thus obtaining multi-dimensional feature representations. In addition, to enhance the integration of different features and improve the overall feature representation, we also introduce a feature-fusion module. To further enhance UAV performance, we introduce an improved ray-based eight neighborhood algorithm (RENA), which efficiently extracts road-edge information with a remarkably low latency of just 7 ms, providing accurate flight guidance and reducing misidentification. To provide a comprehensive evaluation of the model’s performance, we have developed a new drone remote-sensing road-semantic-segmentation dataset, DRS Road, which includes approximately 2600 ultra-high-resolution remote-sensing images across six scene categories. The experimental results demonstrate that PISCFF-LNet achieves improvements of 1.06% in Intersection over Union (IoU) and 0.83% in F1-Score on the DeepGlobe Road dataset, and 1.03% in IoU and 0.57% in F1-Score on the DRS Road dataset, compared to existing methods. Finally, we applied the algorithm to a UAV, using a PID-based flight-control algorithm. The results show that drones employing our algorithm exhibit superior flight performance in both simulated and real-world environments.

A registered vehicle is certified safe at a fixed moment, yet its hazards change continuously thereafter: real-world emissions rise with mileage and tampering, mechanical systems wear toward crash-relevant defects, traction batteries age toward thermal runaway, and software-defined functions are rewritten after registration. Periodic inspection certifies a snapshot while the safety-relevant state drifts away from it. This review reframes in-use vehicle inspection as continuous safety assurance and asks what evidence is required to detect, adjudicate and control a hazard once a vehicle is in service. We model assurance as a four-stage safety-evidence chain — capture, characterise, adjudicate, enforce — assessed against five evidence-quality criteria and instantiated using established safety-barrier and control-theoretic models (bow-tie, STAMP/STPA, ALARP), with a cross-cutting integrity layer for tampering, fraud and cyber manipulation. Using a PRISMA-ScR best-fit framework synthesis of 445 studies (2021–2026; Scopus, Web of Science, IEEE Xplore), we map the maturity of the assurance evidence base on its 339 core- and bridge-tier studies. Detection and risk-characterisation are strongly evidenced in every domain. Two findings are comparatively robust — supported by both the regulatory record and the safety-model reading, not publication counts alone: in-use battery safety has no binding adjudication barrier, and software-defined vehicles have no inspection-side feedback path to verify conformity after updates. Enforcement closure is the least visible component of the peer-reviewed record across domains — a property of the published evidence base, not an audit of real-world capacity; separating a genuine capacity gap from this venue effect heads the resulting research agenda.