A Novel Segment-Based Approach for Improving Classification Performance of Transport Mode Detection

Abstract

Transportation planning and solutions have an enormous impact on city life. To minimize the transport duration, urban planners should understand and elaborate the mobility of a city. Thus, researchers look toward monitoring people’s daily activities including transportation types and duration by taking advantage of individual’s smartphones. This paper introduces a novel segment-based transport mode detection architecture in order to improve the results of traditional classification algorithms in the literature. The proposed post-processing algorithm, namely the Healing algorithm, aims to correct the misclassification results of machine learning-based solutions. Our real-life test results show that the Healing algorithm could achieve up to 40% improvement of the classification results. As a result, the implemented mobile application could predict eight classes including stationary, walking, car, bus, tram, train, metro and ferry with a success rate of 95% thanks to the proposed multi-tier architecture and Healing algorithm.

Vehicle Mode and Driving Activity Detection Based on Analyzing Sensor Data of Smartphones

“…The predicted vehicle modes, ranging from 3 to 8 modes, consist of the non-motorized ones (stationary, walk, run, bike) and the motorized ones (motorcycle, bus, car, train, tram, subway, ferry). Most of these works requires data from GPS or the combination of accelerometer with other sensors (gyroscope, magnetometer) [ 2 , 4 , 5 , 6 , 8 ]. Yet, these input requirements lead to higher power consumption of smartphones.…”

“…Numerous machine learning classification algorithms are applied, for instance, Random Forest (RF), Decision Tree (DT), Naïve Bayes (NB), K Nearest Neighbor (KNN), Support Vector Machine (SVM), Hidden Markov Models (HMM), Gradient boosting decision tree, and XGBoost. Among them, Random Forest usually results in the highest prediction accuracy [ 2 , 8 ]. The similar trend is observed in our experiments.…”

“…Recently, researchers have paid a lot of attention to study various methods of providing assistance and safety awareness to drivers. Indeed, such works primarily fall into the following categories: recognizing vehicle mode (car, bus, train, bike, walking ...) [ 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 ], identifying driving styles (normal, aggressive, drunken, fatigue, drowsy, inattentive ...) [ 10 , 11 , 12 , 13 , 14 , 15 , 16 ], detecting normal/abnormal driving events (moving, stopping, turning left, turning right, weaving, sudden braking, fast u-turn...) [ 17 , 18 , 19 , 20 , 21 , 22 ], accident detection [ 23 , 24 ], estimating energy consumption and pollution [ 25 ], monitoring road and traffic condition [ 26 , 27 , 28 , 29 , 30 ].…”

“…Hence, it is unrealistic to fix such parameter values for all vehicle modes or all driving activities. Up to now, there exist few works that consider inferring the optimal data window size and the optimal overlapping ratio from training datasets [ 8 , 33 , 34 ]. The authors of [ 33 , 34 ] prove that the window size of 1–2 s results in the best accuracy and the best processing speed in predicting various human activities.…”

“…The authors of [ 33 , 34 ] prove that the window size of 1–2 s results in the best accuracy and the best processing speed in predicting various human activities. Then, the authors of [ 8 ] show that the window size of 60 s leads to the highest overall recall rate in detecting vehicle modes from their dataset. However, this long window size causes a slow responding speed as well as a long processing time due to its long feature vectors.…”