Browsing by Author "Fonseca, Alexandre Daniel Ramos"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Sistemas Não Cooperativos para Registo de Assiduidade em Ambiente de Sala de AulaPublication . Fonseca, Alexandre Daniel Ramos; Proença, Hugo Pedro Martins Carriço; Inácio, Pedro Ricardo MoraisOver the years, high school dropout, college dropout, in particular, has always been a hot topic. With the advancement of technology and Artificial Intelligence and Machine Learning, we necessarily have to think of ways to help mitigate this problem. If there are factors that we cannot control, such as the economic ones, there are others where our actions can be directed. One factor that allows us to evaluate the risk of dropping out of school is student attendance. Although this data can be manually analyzed to make these detections, it would be more efficient to have a capable system of recording this attendance, since the human capacity to analyze data is finite, and often can only infer this situation too late. Of course, a system that only registers attendance will not give a definitive answer, but it will be an essential first step. Thus, a system that can reconcile the detection of a subject and his face while being able to constantly monitor where the subject is, always to be able to identify him even if he moves from one place to another, together with facial recognition, seem to be determining factors to bring a system of this calibre to a successful conclusion. This type of system is generally very much related to the quality of the data and its annotations, so it is vital to collect or obtain quality data to help solve the various problems presented. Considering what has been described, the main goal of this dissertation is to try to start solving the problem of school dropout, namely through the study, validation and testing of several stateoftheart methods in the area of object detection, namely people and faces, but also tracking. The same work will have to be done on face recognition methods, being able to indicate the best stateoftheart methods for each task. As mentioned in the previous paragraph, a significant limitation to this type of task is the data quality since it is not always possible to find a set that perfectly fits our context. Thus, to solve this gap, we will also present a dataset with about 40,000 images, thoroughly annotated frame by frame and that we believe to be an asset in solving this problem. In addition to the above, and in order not only to give a more meaningful and targeted response to our detailed data but also to provide a preliminary view of how one of the system’s tasks might work, we will present two experiments with our data in the area of detection. The first will involve finetuning our data, while the second will involve training it from scratch and then presenting its results as proof of the correct choice of the stateoftheart method.