Abstract
Stop-move detection has been an essential step to construct semantic trajectories and extract meaningful activity sequences of moving objects. Detecting stop and move segments accurately is critical because errors occurred in stop-move detection can be propagated and amplified in later steps in trajectory data analysis. In particular, post-processing that merges or discards the detected stop-move segments can make an impact on the accuracy and characteristics of detected stops and moves. Although many stopmove detection algorithms exist and new methods are continuously proposed in the field, studies on comparing the performance of the stop-move detection methods are still scarce.
In this study, we evaluated the effect of post-processing in stop-move detection with four selected existing stop-move detection algorithms—CandidateStops, SOC, POSMIT, and MBGP—in two input-data scenarios: (1) original data and (2) sampled data. The detected stops were assessed by two quantitative measures that quantify the accuracy at different levels of aggregation in space and time: (1) accuracy based on individual data points (i.e., F-measure) and (2) the shape of detected stops (i.e., shape compactness). With the case study, we found that the impact of post-processing on the detection results can vary by a selected algorithm and input data sparsity. The results can potentially provide insights into how to adopt and maneuver the stop-move detection methods for GPS data analysis.