LSTM-based classification of e-scooter trajectory features for single vs tandem riding detection.

Objectives: Dockless electric scooters (e-scooters) have emerged as a popular mode of short-distance transportation in urban environments, offering convenience and flexibility in rental and usage. However, users often engage in unsafe behaviors while riding, posing risks to themselves and others. In this study, we aim to identify unsafe riding behaviors through analysis of riding trajectories, to reduce safety incidents and promote safer e-scooter usage.

Methods: This study explores the classification of single and tandem e-scooter riding behaviors using a data-driven approach. Leveraging trajectory data from Gothenburg, Sweden, collected over 11 days in November 2023; we utilize Long Short-Term Memory (LSTM) neural networks to analyze dynamic temporal features.

Results: The LSTM model demonstrated significant performance advantages over both RNN and Random Forest models, achieving an accuracy of 92.65%, precision of 91.69%, recall of 93.85%, F1 score of 95.56%, and an AUC of 0.9169.

Conclusions: Additionally, optimizing the input sequence length to 240 s of continuous trajectory features balanced computational efficiency with prediction accuracy and stability. Dynamic trajectory features such as acceleration, turning angle, speed, and start SOC play pivotal roles in differentiating riding patterns. The proposed method can assist city authorities and e-scooter operators in real-time risk detection, operational monitoring, and targeted safety interventions, contributing to safer shared micromobility systems.