Statistical models of intelligent video-content analysis for cognition

Abstract

Video content has a pronounced and varied cognitive impact. This thesis develops several statistical models of video and demonstrates that these models can be used as a means of quantifying how video impacts cognition. This work takes two approaches. For children, systems are developed to classify content based on expert recommendation. The second approach can be applied to adults and works by developing methods to determine extreme ranges of emotions that impact cognition. This thesis first develops decision fusion methods for cognitive classification of children’s video content. It then introduces the novel concept of positive developmental classification of videos for children into videos that are deemed to have a negative or positive impact on cognition from a literature review; a novel system was developed to classify and segment the content accordingly. This study also introduces automatic age-based classification. The work focuses specifically on several high-level audio features as they relate to the cognitive capacity of children. As the impact on cognition of adults is dependent on the intensity of emotions, there is a focus on affective ranking. The main contributions include developing a method to rank and cluster sequences based on their affective content without the granularity problem. Furthermore, this thesis compares the accuracy of several regression methods on the LIRIS database and develops a method to incorporate prior knowledge into the cluster assignments. Then several state-based methods to predict valence and arousal are developed. The first method is the dynamic prediction-hidden Markov model for arousal-time curve estimation in sports videos. This method determines the arousal-time curve by selecting a state sequence that maximizes the joint probability density function between the arousal states and the arousal-time curve. The second method is a novel kernel-based mixture of experts model for linear regression. The latter method outperforms other mixtures of experts models in predicting valence and arousal. As the use of animation as a means of obtaining childrenˆas attention, this thesis introduces a method to automatically categorize different animation genres in a video database made for children by statistically modelling the temporal texture attributes of the video.