Fig 1.
Trajectories processing procedure with their use in agent-based pedestrian model.
Fig 2.
Raw trajectories of vaccination center waiting room with red polygon for transformation matrix estimation (left) and perspective transformation of trajectories (right).
Fig 3.
K-distance graph for estimation of optimal value of Eps.
Fig 4.
Clustering of start (dot) and end points (cross) of all 408 trajectories.
Fig 5.
Data visualization for individual classes.
Fig 6.
Visualization of cluster analysis results with t-SNE.
Fig 7.
Trajectories from training dataset.
Fig 8.
Pairplot of downsampled training dataset.
Fig 9.
Pairplot of SMOTE training dataset.
Fig 10.
Covariance matrix of training data set.
Table 1.
Comparison of classification algorithms.
Fig 11.
Graph of the dependence of accuracy and training time on the number of trees.
Table 2.
Parameters of random forest classifier.
Fig 12.
Confusion matrix of undersampled train set (left) with accuracy: 0.836 and recall: 0.7481, undersampled trainset mixed with pseudo labelled data 90% (middle) accuracy: 0.7626 and recall: 0.8217 and undersampled trainset mixed with pseudo labelled data 99% (right) with accuracy: 0.8463 and recall: 0.7558.
Fig 13.
Graph of accuracy score (or recall score) dependence on the size of the training set.
The gray area shows when only data from the original training set was used. The accuracy score for the training set was determined by 10-fold cross validation.
Fig 14.
Confusion matrix of the model using SMOTE data set with accuracy: 0.8471 and recall: 0.7016.
Fig 15.
Graph of the dependence of accuracy and recall scores on the size of the training set.
The accuracy score for the training set was determined by 10-fold cross-validation.
Table 3.
Results of model comparisons.
Fig 16.
View corrected trajectories from the test dataset (left column) and their comparison with manually defined erroneous points (right column). The green curve indicates the corrected trajectory.
Table 4.
Model parameters used for calibration.
Fig 17.
Waiting times for 1/1 waiting (left), 1/2 waiting (middle), 2/2 waiting (right).
Fig 18.
Waiting points location and qualitative validation of trajectory shapes of cleaned trajectories from pedestrian traffic detector, model A–basic, and model B–advanced.