Abstract: In obstacle environments, persistent tracking of non-cooperative target remains technically challenging for UAV due to factors such as uncertain target motion, visual occlusion, and inherent UAV dynamic constraints. To address this, we propose a fusion control method for target tracking and obstacle avoidance based on deep reinforcement learning. Adopting an end-to-end approach, this method designs a fusion control strategy that directly maps image features to UAV control commands, effectively overcoming the limitations of traditional control methods where tracking and obstacle avoidance modules are designed separately. Firstly, we design a motion-compensated extended Kalman filter incorporating homogeneous transformations to estimate target position and motion direction in real-time within the image plane. Subsequently, we develop a motion controller based on deep reinforcement learning. Sensor observations are fed into the policy network to generate control commands in real-time, enabling the UAV to autonomously track the moving target while avoiding obstacles. Comparative experiments verify the superior performance of the proposed method in terms of target tracking stability, motion smoothness, and computational real-time capability.