Scalable Real-time Abnormal Event Detection

State-of-the-art video anomaly detectors typically rely on a costly object detection method to increase precision, limiting the processing bandwidth to one video stream per GPU, at around 20-30 FPS. However, for real-world video surveillance, e.g. monitoring an entire city with hundreds or thousands of cameras, the processing costs of object-centric video anomaly detectors are simply too high, given their power consumption and the cost of GPUs. To this end, we will present two lightweight models, capable of processing over 60 video streams at 25 FPS, significantly reducing the processing costs. Different from competing models performing anomaly detection at the object or spatio-temporal cube levels, we present models that take whole video frames as input, which is significantly more efficient. The presented models employ several techniques to achieve efficiency, e.g. adversarial knowledge distillation, self-distillation and masked auto-encoders. Comprehensive experiments on four benchmarks show that the presented methods are significantly faster than state-of-the-art methods, while achieving comparable accuracy levels.

You may also join via Zoom:  https://umich.zoom.us/j/95715143468

Password: AIseminar

Bio
Radu Ionescu is a Professor at the University of Bucharest, Mathematics and Computer Science Department. He is also Co-founder & CTO at SecurifAI and Lead Scientist at VeridiumID. His research interests lie at the intersection of AI, machine learning, and deep learning.

More details: https://raduionescu.herokuapp.com/