Visual obstacle avoidance path planning is a core technology for enabling efficient and safe operation of autonomous robots and intelligent vehicles. In recent years, methods in this field have gradually evolved from a planning-driven paradigm to a learning-based optimization paradigm. This paper provides a systematic review of typical local and global planning strategies within the planning-driven approach, analyzing their strengths and weaknesses in terms of adaptability to dynamic environments, real-time performance, and robustness. We then focus on learning-based optimization methods, summarizing strategies founded on model predictive control and optimization, as well as end-to-end learning approaches, and discuss their applications and challenges in handling high-dimensional perception, enhancing generalization capability, and achieving adaptive decision-making. A comprehensive comparison reveals that no single paradigm can simultaneously ensure both perception accuracy and path safety; future research should explore hybrid architectures that combine the stability of traditional planning with the flexibility of deep learning to improve system interpretability and safety. Finally, we outline emerging research trends—such as efficient simulation frameworks, sample efficiency, and multi-sensor fusion—with a view to guiding the continued development of visual obstacle avoidance path planning.