Enhancing UAV Safety through Semantic Scene Completion for Powerline Detection

3D semantic scene completion (SSC) is a task that infers complete scene geometry with semantic information from partial observations to better understand the world. Applying SSC to detecting power lines is crucial for UAV safety due to the high risk of collisions, which can lead to significant operational hazards. This thesis leverages AirSim to create a simulation dataset specifically designed for power line detection tasks, enabling the evaluation of different network architectures. Two sets of data were synthesized: LiDAR FOV and Extended FOV. Both datasets have the same input, but the former includes ground truth covering the LiDAR field of view, while the latter extends the field of view in both vertical and range dimensions. We compare a 2D convolution-based network, LMSCNet, against a 3D convolution-based network, SSCNet, in terms of their performance in standard LiDAR and extended FOV scenarios. Additionally, we extend LMSCNet to incorporate multimodal inputs by fusing LiDAR volumes and semantic RGB images. The extended model, RedLMSCNet, integrates an image branch utilizing ResNet-50 and employs a full decoder. Our results demonstrate that RedLMSCNet significantly outperforms ResLMSCNet, which uses a simple decoder, by improving completion behavior and noise management. These findings underscore the importance of fusing semantic image features with LiDAR data to enhance UAV power line detection capabilities.