IEEE PerCom2025
4 research presentations at PerCom2025, organizing PerVehicle Workshop, winning an award, and presenting a panel talk.
Publications & Talks Awards
International Conference · 2025
Behavioral-Aware Early Destination Prediction Using LiDAR Point Clouds and Movement Trajectories (PerVehicle)
Abstract
Large-scale facility management increasingly relies on understanding human movement patterns, yet existing person re-identification (Re-ID) methods face challenges in tracking individuals across non-overlapping sensor regions. This paper presents a novel approach combining 3D LiDAR point cloud data with trajectory information to predict pedestrian destinations in early stages of movement. By predicting likely destinations, our method can significantly narrow down the search space for matching individuals between different sensor regions, thereby improving Re-ID accuracy. Our method leverages deep learning architectures—PointNet for processing point cloud data and LSTM for analyzing trajectory patterns—to capture both instantaneous behavioral characteristics and temporal movement patterns. Through experiments conducted with 33,110 trajectories in a university campus setting, our combined approach achieved 90.0% accuracy in predicting three destination categories (entrance, elevator/stairs, and meeting rooms), significantly out-performing single-modality approaches. The results demonstrate that early behavioral cues indeed carry significant predictive power for destination estimation. These findings have important implications for improving person tracking across distributed sensor networks and enabling proactive facility management systems.
Research Note
Tracking people through a large facility inevitably involves gaps where no sensor sees them. Linking a person who leaves one sensor's coverage to the person who appears in another is the task of person re-identification, and it becomes error-prone as the number of candidates grows.
The observation behind this work is that knowing where someone is heading narrows those candidates considerably. Somebody walking towards the elevators will reappear at a different sensor from somebody walking towards the exit, so predicting the destination early in the movement shrinks the search space for matching.
Two kinds of information are used. Three-dimensional LiDAR point clouds capture instantaneous behavioural characteristics such as posture and body orientation, while trajectories capture the temporal pattern of how the person has moved so far. PointNet processes the point clouds and an LSTM analyses the trajectories. PointNet operates on point clouds directly without depending on the ordering of the points, which suits a data format that has no natural order.
In experiments with 33,110 trajectories collected on a university campus, the combined approach reached 90.0% accuracy in predicting three destination categories, entrance, elevator or stairs, and meeting rooms, substantially outperforming either modality alone. Early behavioural cues therefore do carry real predictive power, with implications for tracking across distributed sensor networks and for proactive facility management.
Related Lab Activities
