Identification of Urban Last-Mile Courier Delivery Stops Using GPS Trajectory Data
Main article
Abstract
Urban last-mile logistics has grown substantially in recent years, propelled by the rapid expansion of e-commerce and on-demand delivery services. Despite its operational significance, fine-grained monitoring of last-mile courier activities remains challenging: a single courier shift generates hundreds of stops, many lasting only a few minutes, and the GPS records collected by on-board fleet-tracking devices carry no explicit label indicating the purpose of each stop. This paper presents an end-to-end framework for identifying delivery stops from passively collected GPS trajectories by combining vehicle traces with electronic waybill records. Drawing on a real-world courier dataset spanning the full calendar year 2022 and covering more than 80 delivery vehicles in a major metropolitan area, the framework cleans raw trajectories, extracts stop candidates through speed-threshold segmentation and spatial merging, and constructs ground-truth labels by matching stop candidates to waybill records within calibrated temporal and spatial tolerance windows. Five interpretable features include dwell time, pre-stop speed, heading change at the stop, local stop density, and distance from the departure hub are then derived to represent each candidate. To counter the severe class imbalance between delivery and non-delivery stops, SMOTE resampling is applied exclusively to the training partition before three supervised classifiers—Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Decision Tree (DT)are trained and evaluated under a unified protocol. All three models achieve test-set identification accuracies exceeding 98.9%, confirming that the proposed interpretable feature set enables robust delivery-stop identification and supports scalable operational monitoring of urban last-mile courier services.
