Novel algorithms for optimised video surveillance
More and more video surveillance systems are required to meet increased safety and security needs, but the manpower to analyse the video footage is not always available. Better software is needed to automatically analyse the footage without too many false alarms and miss rates, requiring more advanced algorithms, particularly since programmable graphics processor units (GPUs) have evolved significantly. The EU-funded PARALLELYTICS (Real-time video analytics engine optimized for GPUs) project sought to design and implement parallel video analysis algorithms optimised for today's powerful highly parallel processors. PARALLELYTICS focused on three overriding research goals: parallelisation friendly machine learning algorithm selection and design; unsupervised and supervised learning algorithms and feature design for normalcy learning; and development of adaptive message-passing algorithms. Project partners established a machine learning course as part of Istanbul Sehir University's undergraduate computer science curriculum. Two graduate courses on machine learning and probabilistic graphical models were designed. The Data Science Lab was set up comprising 11 graduate and undergraduate students. Researchers developed feature selection measures for unsupervised and supervised decision tree learning and dynamic time warping applications to modelling time series data. These techniques were then applied to model social media discussion on various social networking platforms using real data. They also designed features for shape recognition, crowd activity recognition and individual agent motion modelling. In addition, the PARALLELYTICS team designed intelligent and GPU-friendly message-passing algorithms used to improve estimation of unknown variables and remove invalid information. PARALLELYTICS should change the efficiency and improve the accuracy of video surveillance, giving a welcome boost to safety and security worldwide.
Keywords
Video surveillance, graphics processor units, PARALLELYTICS, video analytics engine
