Belief-integrated timed elastic band planning for robust indoor person following under intermittent observations

Indoor person-following robots routinely lose sight of their target at corners, doorways, and in crowds, and most local planners then stall at the last seen position. We present IO-TEB (Intermittent-Observation Timed Elastic Band), a local planner that maintains a probabilistic estimate of the hidden person’s position and its growing uncertainty, feeds this estimate directly into trajectory optimization, and actively searches likely paths when the person stays hidden. Across 4,320 simulated episodes covering eight indoor scenes, three occlusion durations, and real pedestrian trajectories, IO-TEB completed 87.36% of following tasks, versus 55.14% for the standard timed elastic band planner. It was statistically equivalent to a conventional design that handles the same estimate in a separate finite-state supervisor, indicating that embedding the estimate in the optimizer costs neither accuracy nor computation. Compared with model predictive control (MPC) using trajectory prediction and three recent person-following planners, IO-TEB was not the most accurate, but in our implementation it missed the 100 ms control-cycle deadline least often (3.32% of cycles versus 5.14–44.21%). Ablation shows that replacing the last observation with the predicted estimate as the following reference accounts for most of the gain, whereas the uncertainty-dependent terms added no consistent benefit.