|
This paper presents a novel approach to sensing both suspicious, and task-specific behaviours through the use of advanced computational intelligence techniques. Locating suspicious activity in surveillance camera networks is an intensive task due to the volume of information and large numbers of camera sources to monitor. This results in countless hours of video data being streamed to disk without being screened by a human operator. To address this need, there are emerging video analytics solutions that have introduced new metrics such as people counting and route monitoring, alongside more traditional alerts such as motion detection. There are however few solutions that are sufficiently robust to reduce the need for human operators in these environments, and new approaches are needed to address the uncertainty in identifying and classifying human behaviours, autonomously, from a video stream. In this work we present an approach to address the autonomous identification of human behaviours derived from human pose analysis. Behavioural recognition is a significant challenge due to the complex subtleties that often make up an action; the large overlap in cues results in high levels of classification uncertainty. False alarms are significant impairments to autonomous detection and alerting systems, and over reporting can lead to systems being muted, disabled, or decommissioned. We present results on a Computational- Intelligence based Behaviour Recognition (CIBR) that utilises artificial intelligence to learn, optimise, and classify human activity. We achieve this through extraction of skeleton recognition of human forms within an image. A type-2 Fuzzy logic classifier then converts the human skeletal forms into a set of base atomic poses (standing, walking, etc.), after which a Markov-chain model is used to order a pose sequence. Through this method we are able to identify, with good accuracy, several classes of human behaviour that correlate with known suspicious, or anomalous, behaviours.
|
