Driver Maneuver Detection and Analysis Using Time Series Segmentation and Classification.

Area of Science:

• Automotive Engineering
• Machine Learning
• Data Science

Background:

• Vehicle maneuver detection is crucial for driver safety and autonomous systems.
• Existing methods often treat maneuver detection solely as a classification problem, neglecting the time-series segmentation inherent in continuous telemetry data.

Purpose of the Study:

• To develop an end-to-end pipeline for automatic, frame-by-frame annotation of vehicle maneuvers from naturalistic driving data.
• To address both the segmentation and classification challenges in vehicle maneuver detection.

Main Methods:

• An energy-maximization algorithm (EMA) was developed for time-series segmentation of driving events.
• Heuristic algorithms were employed for classifying highly variable events like stops and lane-keeping.
• Four machine learning models (1D-CNN, LSTM, Random Forest, SVM) were implemented and evaluated for classifying segmented events.

Main Results:

• The EMA algorithm extracted driving events with durations comparable to actual events, achieving accuracies from 59.30% (left lane change) to 85.60% (lane-keeping).
• The 1D-convolutional neural network (1D-CNN) achieved the highest classification accuracy at 98.99%, followed closely by LSTM, Random Forest, and SVM models.
• All machine learning models demonstrated consistent accuracy across different datasets, indicating good transferability.

Conclusions:

• The proposed segmentation-classification pipeline significantly enhances the accuracy of vehicle maneuver detection.
• The methodology improves the transferability of both shallow and deep machine learning models across diverse driving datasets.