Vi Dimensions is a leading developer of an innovative anomaly detection system that applies AI and ML to surveillance video for the security industry. The patented unsupervised machine learning system ARVAS (Abnormality Recognition Video Analytic System) will autonomously learn normal patterns of behavior to a camera’s FOV without applying rules to detect and alert operators of abnormal behaviors providing real time situational awareness.
Almost all conventional video analytics systems are rule based. Applying these methods of detection and analysis to identify abnormal events is counter-intuitive because they are by nature highly unpredictable.
ARVAS uses advance statistical modelling and deep machine learning to detect anomalies instead. This automated approach enables more accurate detection of complex risk pattern that would otherwise escape human analysts. And since it is data driven, ARVAS becomes smarter and stronger over time, providing more reliable security for cities and enterprises.
ARVAS relies on a multi-server architecture that streams multiple cameras in parallel and provides real-time anomaly detection on each stream. The anomaly detection is handled by a computer vision core engine that learns through time the normal patterns of behaviour and applies an ensemble of deep learning models to categorize the proposed alarms or to discard them in an unsupervised manner. We are working with Intel on every front of this complex and computationally expensive pipeline, from high-performance detection at the edge, to optimized deep learning models at the back end. What makes OpenVINO attractive is its ability to optimize the inference on targeted hardware, thereby abstracting our complex architecture. This is critical for us, because we are progressively pushing the ARVAS anomaly detection engine at the Edge, so as to offer a truly scalable solution to our customers. In our effort to tackle the anomaly detections where humans are involved, we work closely with Intel to design and deploy a more powerful way to detect and track people, analyse their behaviour through time and trigger alarms in case of anomalous behaviours. This is more complex than plugging an OpenVINO pretrained models out-of-the-box and invokes several deep learning models.