Video surveillance and ERP-based BCIs as anomaly detection: New methods and dataset

Official URL: https://risc01.sabanciuniv.edu/record=b3157891_(Table of contents)

Security cameras are widely used to detect and prevent crimes, but the number of surveillance videos has increased due to this prevalence. By processing these videos with the help of a suitable machine learning algorithm, unfavorable events can be brought to the attention of experts to manually monitor. Since these unfavorable events are of various types and few in number, this problem can be addressed in the anomaly detection framework. In this thesis, several new anomaly detection algorithms have been developed using the UCF-Crime dataset. First of all, features were extracted from the videos in the dataset with the help of a pre-trained artificial neural network (ANN). Then, the size of these features was reduced with a semisupervised ANN, an autoencoder (AE) or principal component analysis (PCA). Lastly, anomaly detection was performed using an recurrent neural network (RNN) based on future feature estimation by regression. In addition to these algorithms, a large scale anomaly detection dataset has been introduced. Due to their non-invasive nature, one of the most commonly used event related potentials in brain–computer interface (BCI) system designs is the P300 electroencephalography (EEG) signal. In order to train and test P300-based BCI speller systems in more realistic high speed settings, there is an arising requirement for a large and challenging benchmark dataset. Various datasets already exist in the literature but most of them are not publicly available, and they either have a restrictive number of subjects or utilize relatively long stimulus duration (SD) and inter-stimulus intervals (ISI). The use of long ISI, in particular, not only reduces the speed and the information transfer rates (ITRs) but also simplifies the P300 detection. This leaves a limited challenge to the state-of-the-art machine learning and signal processing algorithms. Therefore, one certainly needs a large-scale dataset in challenging settings to fully exploit the recent advancements in algorithm design (machine learning and signal processing) and achieve high-performance speller results. To this end, by using 32-channels EEG, here we introduce a new freely and publicly accessible P300 dataset, hoping to enhance research findings towards building efficient BCIs. The introduced dataset is composed of 18 subjects performing a 40-target (5 × 8) cued-spelling task, with reduced SD (66.6 ms) and ISI (33.3 ms) for fast spelling. We have also processed, analyzed, and character-classified the introduced dataset and presented the accuracy and ITR results as a benchmark.