Autonomous collision avoidance is a critical requirement for Maritime Autonomous Surface Ships (MASS), especially in restricted waterways with limited maneuvering space, dense traffic, and strict regulatory constraints. Existing model-based and reinforcement learning approaches often simplify the action or state space, resulting in poor maneuver feasibility and limited rule compliance in narrow channels and harbor environments. This paper proposes a POMDP-based collision avoidance framework tailored for restricted-water navigation. The proposed method integrates vessel maneuverability constraints, navigational boundary limitations, and rule-aware semantic representations to capture encounter situations and local regulatory priorities. A two-dimensional continuous action space, combining heading rate and speed adjustments, is introduced to improve maneuver realism and feasibility. The enhanced POMDP is trained using Proximal Policy Optimization (PPO), enabling stable policy learning under partial observability and dynamic multi-vessel interactions. Simulation results demonstrate that the proposed approach effectively ensures safe, efficient, and rule-compliant navigation in scenarios involving both static obstacles and dynamic vessels, validating the advantages of the improved state and action representations.