LSTM-Based Trajectory Prediction for Human-Drone Interaction in Structured Environments

Anticipating human movement is essential for robotic motion planning in structured environments such as corridors, hospitals, and factory floors. This paper presents a lightweight, real-time trajectory prediction pipeline for human-aware drone-navigation, using head-orientation motion as the primary cue for intent inference. The approach combines a branched Long Short-Term Memory (LSTM) network with a discrete intent classifier that identifies the most likely motion mode (e.g., left, right, or straight at a hallway junction), subsequently predicting the corresponding continuous trajectory. While prior work has addressed human motion prediction using heavy multimodal predictors, the proposed model is designed for on-board deployability and low-latency inference. In addition, few studies have explored structured environments where movement is constrained to discrete paths. We collect a synchronized dataset of head-motion trajectories using an inertial measurement unit (IMU) and motion-capture system, and train the LSTM to predict future trajectories of humans, with a normalized motion frame calibrated to the environment. To evaluate the model, we integrate the predictions in a structured testbed with finite path options and static obstacles to demonstrate how prediction accuracy affects human-drone coexistence. The results demonstrate that our lightweight model enables accurate trajectory forecasting and classification, enabling early avoidance planning. The work highlights the viability of simple, interpretable sequence models for real-time human-aware navigation in constrained spaces.