Semantic-Driven Informed Planning and 3D Reconstruction for the Quadrotor Unmanned Aerial Vehicle

Autonomous aerial reconstruction tackles the problem of deploying the quadrotor unmanned aerial vehicle in an unknown environment for vision-based surface reconstruction. The state-of-the-art methods mainly extend the classic next-best-view (NBV) framework with the semantic information to concentrate on the reconstruction target. However, the greedy decision of the next viewpoint or branch without global consideration leads to the low efficiency. This paper proposes a novel semantic-driven aerial reconstruction planner, which integrates the task relevance into the planning, making the concentrative reconstruction more efficient. Specifically, a new surface-frontier information structure is extended to maintain the rich semantic information for the fine-grained planning. Driven by an attention-based frontier filter, a novel relevance-aware frontier sequence generation method is proposed to achieve the concentrative reconstruction of the target. Then a semantic-informed local refinement, which incorporates the semantic gain into the graph search, is proposed to further reduce the flight cost and improve the reconstruction quality. Comparative experiments are conducted in simulation environments, demonstrating that the proposed method outperforms the state-of-the-art methods in terms of both the reconstruction efficiency and quality.