Student : P Mathee
About the student
As a Computer Science student, I am constantly driven by the need for a deeper understanding of the reasoning and workings of technologies that impact us.
About the Project
Video surveillance has become increasingly commonplace, leading to a growing demand for automated surveillance systems. Anomaly detection on video surveillance is one of these areas and presents a challenge, especially in terms of multi-scene anomaly detection, where lighting and camera angles affect model performance. In this work, I explore various unsupervised and zero-shot learning methods to address anomaly detection in this context. I conduct 4 experiments that employ unsupervised and zero-shot learning methodologies. These include anomaly detection with Isolation Forests and handcrafted features such as Histogram of Oriented Gradients and Lucas-Kanade Optical Flow, anomaly detection with Autoencoders and KMeans clustering on Local Binary Patterns combined with Motion History Images, anomaly detection with Autoencoders using DINOv3 and features from a temporal transformer, and finally, anomaly detection using the FLAIR Vision-Language Model. While handcrafted methods in the first experiments ultimately fall short of reliable anomaly detection in challenging, multi-scene and lighting video footage, Deep Learning methods such as Vision-Language models show more reliable results.
