Hybrid Deep and Ensemble Learning for Transportation Mode Detection using Sliding-Window Sensor Data

Abstract

Transportation mode detection is an essential component of Human Activity Recognition (HAR) systems, supporting applications such as smart mobility, personalized navigation, and urban planning. Traditionally, these systems relied on manual observation, surveys, or simple threshold-based methods using smartphone or wearable sensors. Such approaches were often limited in accuracy, required extensive human intervention, and could not capture complex, non-linear patterns in multi-sensor data. The proposed system, MetaFuseNet, is a hybrid deep and ensemble learning framework designed for accurate detection of transportation modes including walking, still, car, bus, and train using sliding-window sensor data. MetaFuseNet integrates a Deep Neural Network (DNN) to model temporal and non-linear relationships, a Random Forest Classifier (RFC) for robust feature-based decision-making, and a Logistic Regression (LR) meta-model to combine predictions from the base models for improved performance. Additionally, traditional classifiers such as K-Nearest Neighbors (KNN) and Logistic Regression (LR) are evaluated for benchmarking. The system is implemented in Python using TensorFlow, Keras for deep learning and scikit-learn for classical machine learning. A Flask-based web interface enables users to upload sensor data files and obtain real-time predictions using the pre-trained models, without requiring retraining. This research delivers a complete end-toend pipeline for transportation mode prediction, combining advanced deep learning, ensemble learning, and meta-model stacking techniques. By providing automated, accurate, and efficient activity recognition, the system demonstrates a practical solution for real-time transportation monitoring and intelligent mobility applications.