Movement trajectory classification using supervised machine learning

Sammanfattning: Anything that moves can be tracked, and hence its trajectory analysed. The trajectory of a moving object can carry a lot of useful information depending on what is sought. In this work, the aim is to exploit machine learning to be able to classify finite trajectories based on their shape. In a clinical environment, a set of trajectory classes have been defined based on relevance to particular pathologies. Furthermore, several trajectories have been collected using a depth sensor from a number of subjects. The problem to address is to evaluate whether it is possible to classify these trajectories into predefined classes. A trajectory consists of a sequentially ordered list of coordinates, which would imply temporal processing. However, following the success of machine learning to classify images, the idea of a visual approach surfaced. On this basis, the plots of the trajectories are transformed into images, making the problem become similar to a written character recognition problem. The implemented methods for this classification tasks are the well-known Support Vector Machine (SVM) and the Convolutional Neural Network (CNN), the most appreciated deep approach to image recognition. We find that the best possible way to of achieving substantial performances on this classification task is to use a mixture of the two aforementioned methods, namely a two-step classification made of a binary SVM, responsible for a first distinction, followed by a CNN for the final decision. We illustrate that this tree-based approach is capable of granting the best classification accuracy score under the imposed restrictions. In conclusion, a look into possible future developments based on the exploration of novel deep learning methods will be given. This project has been developed during an internship at the company ‘Qinematic’.