PROJECT TITLE :

The Devil Is in the Details An Efficient Convolutional Neural Network for Transport Mode Detection

ABSTRACT:

The objective of the classification problem known as transport mode detection is to devise an algorithm that, given multimodal signals (GPS and/or inertial sensors), can infer the mode of transport that a user is currently utilizing. It has a wide range of applications, some of which include tracking carbon footprints, analyzing mobility behaviors, and performing real-time smart planning door-to-door. The majority of the currently available methods rely on a classification step that implements Machine Learning techniques. However, similar to the outcomes of many other classification problems, the results of Deep Learning methods typically outperform those of traditional machine learning methods that make use of handcrafted features. However, there is a significant drawback associated with deep models, and that is the fact that they typically take up a lot of memory space and are expensive to process. We demonstrate that a simplified and improved model can achieve the same level of performance as an existing deep model. During our work with the GeoLife and SHL 2018 datasets, we were able to develop models with tens of thousands of parameters. This represents a reduction of between 10 and 1,000 times fewer parameters and operations compared to the state-of-the-art networks, yet these models still achieved comparable levels of performance. We also demonstrate, by making use of the aforementioned datasets, that the preprocessing that is currently employed in order to deal with signals of varying lengths is not the most effective method, and we offer alternatives that are more effective. In conclusion, we present a method for utilizing signals of varying lengths with the more lightweight convolutional Neural Networks, as an alternative to the more resource-intensive Recurrent Neural Networks.