Advances in UAV Visual Navigation for Low-Altitude Air Traffic Safety: An Integrated Perspective on Perception, Decision-Making, and Traffic Management

Abstract

As applications such as logistics delivery, urban air mobility, emergency response, and infrastructure inspection expand, low-altitude UAV operations are shifting from dispersed, low-density flights toward large-scale, high-density operations. Visual navigation, owing to its low weight, low cost, and limited dependence on external infrastructure, has become a key enabling technology for autonomous localization, environmental perception, and obstacle avoidance when global navigation satellite system signals are degraded or unavailable. From the perspective of low-altitude air traffic safety, however, improvements in the performance of individual vision algorithms do not automatically translate into safe system-level operations. Visual observations must be converted into conflict assessments, safety-critical decisions, exchangeable information, and operationally manageable states before a closed safety loop can be established. This review examines UAV visual navigation through three interrelated dimensions: perception capability, decision safety, and compatibility with traffic management. It synthesizes research on visual simultaneous localization and mapping and visual-inertial odometry, object detection and tracking, depth estimation, conflict detection, autonomous avoidance, low-altitude traffic management, operational risk assessment, and safety assurance. Particular attention is given to the propagation of visual-perception uncertainty into safety-critical decisions, the transformation of onboard visual outputs into information that can be used by low-altitude traffic management systems, and the assurance of learning-based vision systems. The review indicates that the main bottlenecks are shifting from the accuracy of individual algorithms to cross-layer metric alignment, standardization of visual safety information, interoperability among heterogeneous systems, and certification of learning-based perception. Future research should therefore extend beyond algorithmic optimization toward operational risk management and regulatory oversight by developing safety-consequence-oriented metrics, standardized information interfaces, runtime assurance mechanisms, and digital-twin-based validation integrated with flight testing.