Anomaly detection in human behaviour using computer vision: A review

Abstract

To improve the safety of the public, hundreds and hundreds of CCTV cameras are being installed in public spaces such as roads, shopping centres, parks, etc., but security organisations’ capacity for monitoring cannot keep up. Monitoring relies on human judgement, which is affected by factors such as distraction and stress. Moreover, manual monitoring also comes with the added risk of missing an anomalous event or a delay in reporting. All of this contributes to the need for automation in surveillance. A robust machine learning model that uses computer vision techniques, such as image and video data analysis, object detection, and motion tracking, can help mitigate these problems. Starting with introducing supervised learning methods like Convolutional Neural Networks (CNNs) to learn normal and anomalous data patterns and using Long-Short Term Memory (LSTMs) to capture and analyse time-related dependencies. Moving on to autoencoders, it is an unsupervised learning approach that significantly improves anomaly detection by overcoming the need for a detailed labelled dataset. They learn to reconstruct the normal patterns during training, and during inference, any deviations are marked as an anomaly. Such an autonomous system will allow organisations to focus on tasks that require human attention and help minimise the response time of authorities in the event of anomalous events. This paper presents a comprehensive review of the research done on this topic.

DOI: https://doi.org/10.24321/3117.4787.202602