Clustering And Filtration Of Pedestrian Trajectories In Vaccination Centre Using DBSCAN And Random Forest

Valid input pedestrian dynamics data is essential assumption for valid egress (or ingress) model. Standard approach to obtain these data is conducting an experiment, which is expensive and has limited scope and analysis possibilities. Pedestrian detection and tracking algorithms can be used for extraction of trajectories from surveillance cameras. This approach allows gathering huge amount of unique data put in context of modelled object. Due to detection errors caused by occlusions, data are biased and need to be validated in post-processing. This paper proposes two steps machine learning approach for categorization of trajectories based on their origin and destination using DBSCAN clustering method (step 1) and filtration of valid data points in individual trajectory using Random Forest (step 2). Processed trajectories are then used for aggregation of two distributions which can be used as direct input in model: walking speed and waiting times. Proposed approach was tested on long term video footage of high capacity vaccination center operation.