Exploration of multisensory integration for adaptive activity recognition

• 主要作者: Ke, Na
• 其他作者: Mao Kezhi
• 格式: Final Year Project
• 語言: English
• 出版: 2017
• 主題: DRNTU::Engineering::Electrical and electronic engineering
• 在線閱讀: http://hdl.handle.net/10356/71706
• 機構: Nanyang Technological University

In order for a machine to start performing a task, we need to first train it the way to solve the problem. When we encounter image classification problem, machine is not like humans, it is not difficult for us to recognize things but for computers, it has a series of steps to conduct in order to get the classification result. The normal way to do image classification by a machine is by extracting the image features from a predefined data set first. Then use these features to train a classifier, and lastly use the classifier to make prediction for unseen images. However, when a human classifies the activities, except for the image features, we can also refer to the additional information to make decision. Especially for the activities that we never met before. This multisensory integration system of humans inspires me to search for other sources to make help improve the accuracy of activity classification. In this report, two classification method will be introduced, image-based classification and text-based classification. A process of how these two ways implement the classification function will explained in detail in the main content. In both classification process, two classifiers, SVM and Naïve Bayes classifiers will be used and the performance will be evaluated respectively. Lastly, a fusion of these two classification is developed, the decision score and the fusion-based classification accuracy is calculated. Followed by that is a conclusion of this project, the future work can be done for the further improvement of the project, and the program codes are included in the appendix.