17 February 2020
13:30 Master's Defense Auditorium 2 - IC 3
Theme
Anomalous Action Detection in Videos Assisted by High-Level Features Using a Multi-Stream Deep Neural Network
Student
Sarah Almeida Carneiro
Advisor / Teacher
Helio Pedrini
Brief summary
Understanding actions is a topic that has been continuously studied by human beings with, for example, the objective of understanding social behavior or preventing situations. In addition, with the growing number of devices that can capture information from the environment, the ability to easily monitor these actions has been improved. In addition, given that society has sought more imposing security measures, both in the sense of well-being and protection, computational studies have emerged in the area of classification to help them. The study of the classification of anomalous actions has become common for the development of applications useful for detecting and signaling unusual events in a given environment. In the context of anomalous actions, during our work, we decided to focus on just two groups. The first group considered actions involved with well-being in which the data sets were related to falls. The second, related to preventive protection measures, used data sets associated with fighting events. There are many studies on this field of specialization. The best results reported in the literature are from works related to deep learning approaches. Therefore, this study aimed to use a deep learning model based on multi-stream classification systems using high-level characteristics to be able to address the issue of fight detection and video fall detection. In this work, we focus on the use of a multi-stream network, where each stream is a VGG-16 network. In addition, each stream is responsible for receiving a pre-filtered video as input. This pre-filtered video is related to what, in this work, we consider high-level descriptors. Thus, during this study, conceivable high-level descriptors are also investigated, such as spatial, temporal, rhythmic and depth information of a video for the classification of the chosen anomalous actions. We validated our method against five data sets commonly used in the literature, two for detecting fights and three for detecting falls. The experiments demonstrated that the association of information from descriptors, correlated to a multiple flow strategy, increased the classification of our deep learning approach, therefore, the use of complementary characteristics can produce interesting results that correspond to other previous studies.
Examination Board
Headlines:
| Hélio Pedrini | IC / UNICAMP |
| Ronaldo Cristiano Prati | CMCC / UFABC |
| Alexandre Mello Ferreira | IC / UNICAMP |
Substitutes:
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| David Menotti Gomes | UFPR |