Content extraction from video signals with applications to the automotive and healthcare sectors

Content extraction from video signals represents a topic of great research interest thanks to the promising characteristics of video processing techniques, which are mainly related to the use of digital cameras as data acquisition sensors. The use of digital cameras to generate and record video signals, to be furtherly processed, provides several advantages regarding the cost and deployment of video-based systems developed to extract information about real world scenarios. These systems are non-invasive, as camera sensors do not require a direct contact with the subject to be framed. Also, their cost is moderate, thanks to the ubiquitous diffusion of digital cameras that makes them accessible and user-friendly devices. Thanks to these aspects, video-processing techniques are versatile and may be adopted in a wide range of application scenarios. Other advantages are related to the characteristics of video signals. Being defined as temporal sequences of still digital images, video signals contain time-related information of framed objects and scenes, that may be associated with relevant evolutionary changes. In this thesis, video-based algorithms to extract properties of dynamic systems are proposed. In particular, specific applications to the automotive and healthcare sectors are presented and innovative video-based solutions are employed in the tasks of motion analysis and human monitoring. Robust motion estimation algorithms are needed to extract various characteristics of dynamic objects related to their motion, e.g., speed and periodicity. Speed estimation, for instance, plays an important role in the context of automotive safety. In this thesis, the topic of estimating the speed of framed objects in video sequences is addressed and a method to deal with geometrical transformations superimposed to the shift of the object under analysis in the camera plane is proposed. Algorithms to extract the periodicity, typical of some movements, are also presented. The respiration act is an example of a periodic movement, that is worth to be investigated in medical applications as it provides important information related to the health status of a subject. As the Maximum Likelihood (ML) principle represents a reliable tool to derive estimators of unknown parameters of interest, it is exploited in this work to implement speed and Respiratory Rate (RR) estimation algorithms based on video processing techniques, adopted to enhance the considered motion signals. The topic of human monitoring in automotive scenarios represents, instead, a cross-sectoral application concerning both healthcare and automotive safety. In this work, a system to assess the stress status of a driver is proposed by combining information extracted from video signals and from physiological sensors. In particular, thermography is exploited to retrieve skin temperature variations, possibly caused by stressful events, from thermal images acquired inside a vehicle during driving performance.