Driver Maneuver Detection and Analysis Using Time Series Segmentation and Classification

The current paper implements a methodology for automatically detecting vehicle maneuvers from telemetry data under naturalistic driving settings. Previous approaches have treated maneuver detection as classification problem, although both time series segmentation and are required since input continuous. Our objective is to develop an end-to-end pipeline the frame-by-frame annotation of studies videos into various events including stop lane-keeping events, lane changes, left-right turning movements, horizontal curve maneuvers. To address study developed energy-maximization algorithm (EMA) capable extracting varying durations frequencies continuous signal data. reduce overfitting false alarm rates, heuristic algorithms were used classify with highly variable patterns such stops lane-keeping. segmented four machine-learning models implemented, their accuracy transferability assessed over multiple sources. duration extracted by EMA was comparable actual accuracies ranging 59.30% (left change) 85.60% (lane-keeping). Additionally, overall 1D-convolutional neural network model 98.99%, followed long-short-term-memory at 97.75%, then random forest 97.71%, support vector machine 97.65%. These consistent across different concludes that implementing segmentation-classification significantly improves driver shallow deep ML diverse datasets.