Autonomous indoor AAV navigation in GPS-denied environments is challenging. Issues include low lighting and obstacles. This work proposes a dual guidance framework to address these challenges. A CNN-based model with attention mechanisms and scene aware features serves as input to deep reinforcement learning. The attention module focuses on key visual elements. It highlights important regions and helps the AAV avoid collisions. The scene aware component uses computer vision techniques like pedestrian detection and brightness assessment. It gathers environmental information to assist navigation. These two models enhance the ability of DRL to adjust the AAV trajectory. This ensures safe and efficient navigation in complex indoor spaces. The approach was tested in a simulated indoor environment with pedestrians, created using Unreal Engine. The results show its effectiveness in enabling safe AAV navigation. Both the scene and attention models contribute significantly. Compared to ResNet, EfficientNet, DenseNet, and transformers, the CNN-based attention model performed better in accuracy and F1-score. The proposed navigation method also outperformed PPO and similar approaches in the literature.