Vision-Based Activity Recognition System with a Deep Neural Network for Surveillance

Computer vision has gained tremendous attention recently due to what visual data can provide in terms of meaningful information and predictions. Videos, not like other types of data, can carry lots of information about the captured scene. Information such as objects detection, face recognition and action classification can be beneficial to monitoring systems such as traffic monitoring and security systems. Activities recognition, in particular, is a quite significant part of visual data analysis and can provide pragmatic predictions on people’s behaviour. The absence of a well-labelled video dataset makes it more challenging to develop machine learning algorithms for irregular actions recognition. These prediction models can ease the process of monitoring buildings, roads and other common areas monitored by CCTV systems. This paper proposes a method to utilise deep learning in classifying people’s behavior by identifying normal behaviours and classifying any unusual activities and provide a well-trimmed and labelled dataset for abnormal behaviors in CCTV videos