Abstract:

To address the requirements for low-latency local planning and safe obstacle avoidance in UAV forest-fire patrol under dense tree-trunk occlusion and locally invalid depth observations, this paper proposes a learning-based local trajectory planning method that fuses depth images, local occupancy distance maps, candidate-level collision-risk assessment, and pre-execution safety reranking. The method uses depth images, local occupancy distance maps, motion states, and target direction as inputs and generates 15 quintic-polynomial candidate trajectories in parallel using 3 × 5 angular motion primitives. During training, candidate safety costs, ranking supervision, and near-collision risk labels are constructed. During inference, online LiDAR point clouds are used to recheck and rerank the candidate trajectories before execution. The local occupancy distance map compensates for the lack of horizontal spatial priors in the depth-only modality under trunk occlusion, missing depth, and near field-of-view boundaries, while the auxiliary risk head and online rechecking improve the safety redundancy of candidate selection. Simulation results show that at 6 m/s, the proposed method achieves success rates of 81% and 57% in low- and high-density forest scenarios, with collision rates of 2% and 5% and average planning latencies of 20.81 and 23.27 ms, respectively. Compared with the depth-only baseline, the success rates increase by 29 and 12 percentage points, respectively. The method improves the safety of local obstacle avoidance and the stability of closed-loop execution while maintaining real-time performance.

Key words: forest obstacle environment, unmanned aerial vehicle, trajectory planning, multimodal perception, safety reranking, collision risk