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

Area of Science:

• Urban mobility and transportation safety
• Machine learning applications in behavioral analysis
• Micromobility risk assessment

Background:

• Dockless electric scooters (e-scooters) are increasingly popular for urban short-distance travel.
• Unsafe riding behaviors pose significant risks to riders and the public.
• There is a need to identify and mitigate these unsafe behaviors for improved safety.

Purpose of the Study:

• To classify single and tandem e-scooter riding behaviors using a data-driven approach.
• To identify key dynamic trajectory features indicative of unsafe riding.
• To develop a system for real-time risk detection and safety interventions in shared micromobility.

Main Methods:

• Utilized trajectory data from e-scooters in Gothenburg, Sweden.
• Employed Long Short-Term Memory (LSTM) neural networks for dynamic temporal feature analysis.
• Optimized input sequence length to 240 seconds for computational efficiency and prediction accuracy.

Main Results:

• The LSTM model achieved high performance: 92.65% accuracy, 91.69% precision, 93.85% recall, 95.56% F1 score, and 0.9169 AUC.
• Demonstrated significant performance advantages over traditional models like RNN and Random Forest.
• Identified acceleration, turning angle, speed, and start state of charge (SOC) as pivotal features.

Conclusions:

• The proposed LSTM-based method effectively classifies e-scooter riding behaviors.
• This approach aids city authorities and operators in real-time risk detection and safety interventions.
• Contributes to the development of safer shared electric scooter systems and urban micromobility.