Data Fusion for Deep Learning on Transport Mode Detection: A Case Study
This work addresses methodological gaps in transport mode detection for researchers and practitioners, but it is incremental as it focuses on comparative evaluation rather than introducing new techniques.
The study tackled the problem of incomplete comparisons in transport mode detection by evaluating various methodological choices, particularly data fusion methods, on a public dataset, and found that simple late fusion outperformed other literature methods and that 2D convolutions on spectrograms with logarithmic frequency axes were superior to 1D temporal representations.
In Transport Mode Detection, a great diversity of methodologies exist according to the choice made on sensors, preprocessing, model used, etc. In this domain, the comparisons between each option are not always complete. Experiments on a public, real-life dataset are led here to evaluate carefully each of the choices that were made, with a specific emphasis on data fusion methods. Our most surprising finding is that none of the methods we implemented from the literature is better than a simple late fusion. Two important decisions are the choice of a sensor and the choice of a representation for the data: we found that using 2D convolutions on spectrograms with a logarithmic axis for the frequencies was better than 1-dimensional temporal representations.