This paper proposes a framework for 3D obstacle avoidance in the presence of partial observability of environment obstacles. The method focuses on the utility of the Artificial Potential Field (APF) controller in a practical setting where noisy and incomplete information about the proximity is inevitable. We propose a Particle Filter (PF) approach to estimate potential obstacle locations in an input depth image stream. The probable candidates are then used to generate an action that maneuvers the robot towards the negative gradient of potential at each time instant. Rigorous experimental validation on a quadrotor UAV demonstrates the robustness and reliability of the method when robot's sensitivity to incorrect perception information can be concerning. The proposed method is highly compute efficient for real-time applications and agile robots.