AI for Anomalous Events Detection

Back to Verticals page

In this vertical we use artificial intelligence to identify anomalous events for enhanced situational awareness. It includes following projects:

a. Weakly Supervised Training for Anomalous Events Detection

In this project, the aim is to develop weakly supervised anomaly detection systems, overcoming challenges related to noisy video-level labels and the rare occurrence of anomalous events. This study incorporates innovations such as a random batch selection to reduce inter-batch correlation, normalcy suppression block minimizing anomaly scores over normal regions, and cluster-based loss to mitigate label noise and to enhance representation learning.

Related Publications

● Clustering Aided Weakly Supervised Training to Detect Anomalous Events in Surveillance Videos, IEEE TNNLS, 2023
● A Self-Reasoning Framework for Anomaly Detection Using Video-Level Labels, IEEE SPL, 2020
● CLAWS: Clustering Assisted Weakly Supervised Learning with Normalcy Suppression for Anomalous Event Detection, ECCV, 2020

● Cleaning Label Noise with Clusters for Minimally Supervised Anomaly Detection, CVPRW: Learning from Unlabelled Videos (LUV), 2020

b Stabilizing One-Class Novelty Detection Using Pseudo Anomalies

In this project we focus on one-class novelty detection methods, observing instability during training. We aim to develop robust novelty detection algorithms. The study will utilize both the current and old states of generators to create diverse pseudo-anomalies for training.

Related Publications

● Stabilizing Adversarially Learned One-Class Novelty Detection Using
Pseudo Anomalies, IEEE TIP, 2022

c. Unsupervised Video Anomaly Detection

In this project we aim to develop fully unsupervised video anomaly detection using new learning strategies. We address the challenges of sparse occurrences and the undefined nature of anomalies. Human annotations are completely avoided during the learning process.

Related Publications

● Generative Cooperative Learning for Unsupervised Video Anomaly Detection, CVPR, 2022

d. Anomaly Detection Using Moving Surveillance Robots

In this project we aim to develop an autonomous anomaly detection system utilizing mobile robots with surveillance cameras. Departing from the limitations of static cameras, the system will employ Siamese network-based algorithms for anomaly detection, comparing test images with a geo-tagged normal image database. The learning process will improve with human verification of anomalous verses normal detections.

Related Publications

● An Anomaly Detection System via Moving Surveillance Robots with Human Collaboration, ICCV Workshop, 2021