In the thesis study, novel incremental learning approaches which take into account all characteristics and limitation of the data stream are proposed. The first contribution to the literature is the presentation of automatic feature extraction and selection approach based on Discrete Cosine Transform and Swarm Intelligence for data stream. The second contribution is to develop a data stream learning approach based on Online Sequential-Extreme Learning Machines and Autoencoders for data stream in light of the obtained results of the first approach. Another contribution of the thesis is the improvement of the developed learning approach to be robust to the concept drift problem of the data stream. The proposed approaches are applied to the surveillance video detection application area, one of the real-world data stream problems. In the video anomaly detection application, anomaly detection process is performed with developed data stream learning approach and a weak-labelling technique, and this is the fourth contribution of the thesis. Moreover, this approach have ability to determine the abnormal event type. As the fifth contribution of the thesis to the literature, the developed video anomaly detection approach is combined with the automatic feature extraction and selection approach to increase the video anomaly detection performance. Finally, a novel data stream learning approach based on active learning, which requires minimum video label values for video anomaly detection, has been implemented. All the developed approaches are tested on real, synthetic data stream and video data sets in comparison with popular approaches in the literature, and the analysis has achieved promising results with high success. |