In real-world operations, tasks such as urban inspection and forest search place extremely high demands on UAV autonomous obstacle avoidance and navigation. However, the classic DQN falls short in practical deployment due to its limitations in single-sensor input, insufficient exploration, and slow convergence. This paper introduces three key improvements to DQN: First, a dual-stream network is built to fuse visual and LiDAR information. Second, prioritized experience replay (PER) and NoisyNet are introduced to enhance sample efficiency and exploration quality, respectively. Third, a three-stage curriculum learning strategy is designed to gradually reduce the training difficulty. Experiments are conducted using TensorFlow 2.x and Gazebo, and the results show that our improved algorithm significantly outperforms DQN, DDQN, Dueling DQN, and PPO in terms of the success rate, path length, and adaptability to dynamic obstacle avoidance.