The aim of the proposal is to study innovative surveillance components for autonomous monitoring of complex multi-sensory and networked infrastructure such as underground stations. We will investigate two main research threads relying on two key paradigms:
1. Unsupervised modelling techniques for audio/video streams content characterization. The scientific approaches will explore saliency-guided models and specific topic-modelling Bayesian frameworks for extraction of typical activity patterns from large compressed/uncompressed video collections. Identification of relevant unusual activities in the data streams will be combined with top-down user practices to automatically and adaptively select audio-video streams of interest in the control room.
2. Investigation of behavioral cues (like head or body pose) extraction techniques for the categorization of human activities. Moving one step beyond localization-based human activity, we will rely on strong scientific and social (based on proximity,personal space monitoring, group behaviour) formalisms to infer people activities and perform the live detection of behavioural scenario and contextual analysis of the infrastructure usage.
The above research will be enhanced by applying automatic clustering algorithms to cues collected from multiple scenes over long time periods, for the discovery of collective comprehensive daily routines from individual behaviour patterns. In particular, we will target the development of an online reasoning and learning monitoring system with adaptation capabilities like structural changes propagation.
The proposed approaches will be technically, scientifically and sociologically assessed on two complementary real-scale metro infrastructures operated by the partners.
VANAHEIM gathers a multidisciplinary team of 8 partners (4 research institutes, 2 companies, 2 public operators) from 5 different countries, with the complementary competencies needed to conduct the proposed research.
