Design and Implementation of Online Deep Learning Approaches for Video Anomaly Detection

Anomaly detection in video streams is a hard task of computer vision. Major challenges are poor video quality, data imbalance, context dependence, and models explainability and efficiency. This thesis proposes the design and implementation of novel online approaches for real-world video anomaly detection exploiting a supervised learning methodology. In particular, are presented deep learning architectures based on the analysis of contextual, spatial, motion, and temporal information extracted from the video. A data balancing strategy based on hard-mining and adaptive framerate is used to avoid overfitting and increase detection accuracy. An extended taxonomy is also defined by differentiating anomalies in "soft" and "hard". Moreover, a novel anomaly detection score based on a sigmoidal function has been introduced to reduce false positive rate while maintaining a high level of true positive rate. The proposed methodology has been validated with a set of experiments on a well-known video anomaly dataset: UCF-CRIME. The experiments on the testbed demonstrate the impact of the contextual information and data balancing on the classification performances, considering only "hard" anomalies during training and that the best proposed model can achieve state-of-the-art performances while minimizing resource consumption. The experiments also demonstrate that this model has good generalization capabilities on anomalous subclasses not previously seen during training. Finally, the applicability of the best proposed model on real-world scenarios is qualitatively verified by analyzing its response to various types of anomalies and by exploiting the Grad-Cam explainability tool, in order to assess the model’s decision-making process