Project extension overview

The main objectives during the project extension are:

• Development of applications and demonstrators within WP4
• Development of applications and extension of the simulator of WP1
• Dissemination Workshop (June - September 2018)
• Web update
• Publication of Newsletters (June and December 2018)
• Final version of D4.3 "Results Report" version 7 (Demceber 2018)
• International Publications

Project proposal overview

The objective of this project is to tackle the problem of high availability video analysis, that is, the long-term analysis of video sequences. This project will focus on video-based understanding of people behaviour and it will investigate strategies to characterize the monitored scene, detect the existing people and identify their behaviour (e.g., movements and interactions) over long periods of time, either in single or multiple camera settings. The outcome of the project will provide a people behaviour description that can benefit a wide range of potential applications, nowadays mainly related to security (i.e., video surveillance) but also to people care (e.g., independent living) or monitoring (e.g., commercial areas monitoring) where long-term video analysis is a key issue.

This project assumes, based on our experience on the topic, that most of the recent approaches for people behaviour understanding perform well for short periods of time, for which they are designed, but they fail if applied for long periods of time (i.e., for hours or for a continuous operation). This limitation is due to the fact that the scene-changes over time are not correctly captured by the models employed for analysis and, therefore, they become outdated: models of scene background, which should be robust to multimodal situations; models of the argets or objects of interest, which should be robust to scale and appearance changes; behaviour models, which should cope with point of views and occlusions; or context models, which for instance adapt alarms to the status of a changing traffic light. These scene-changes may affect all stages of the video analysis pipeline (e.g., object segmentation, feature extraction, target tracking, behaviour recognition), thus dramatically underperforming decreasing overall performance in long-time operation video sequences systems, thusand limiting their use in real and practical applications.

These situations require the design of new approaches for adapting the visual analysis tools to the scene dynamics. This project investigates approaches, either in single or multiple camera settings, to perform such adaptation by exploiting the scene context, a continuous self-performance analysis and the use of multiple and possibly also complementary,sensors. The project will contribute with innovative approaches that incorporate using an online analysis of the results quality and contextual data to drive online adaptation of models and algorithms (e.g., configuration). Moreover, collaborative approaches will also be considered to provide feedback-based interaction between processing stages in single cameras or to coordinate multiple cameras operation to efficiently exploit for taking advantage of the multiple viewpoints and camera capabilities. Both adaptation and coordination will improve the system performance will allow the maximization of results performance and the optimal usage of resources for long-term analysis, thus enabling the high availability of video analysis tools.

The proposed research hypotheses are based on previous experience and on initial achievements of the proponent research Group, resulting from work in previous and ongoing research and development projects. A sustained effort in the proposal of methodological evaluation frameworks, including the generation of high quality video sequences with ground-truth data, will be the basis for a rigorous validation of the developed algorithms.