Scene Analysis Of Multi-view Videos Based On Social Camera

Scene understanding can effectively improve the analytical and cognitive ability of computer in coping with complex indoor and outdoor scenes; hence it is one of the most important researching topics in the field of computer vision. Recently, with the rapid development of wearable technologies (such as sports camera, smart glasses and eye tracker), the scene understanding of multi-view videos based on Social Camera has drawn great attention, which requires that object detection algorithm and group detection algorithm should be able to overcome the problems of Social Camera, namely dramatic background changes, distinct object scale and high time-variability of visual angle.The paper further studies the core matters of scene understanding in Egocentric videos, with the main focus on object detection algorithm and group detection algorithm. A data set of multi-view videos is created as well so as to fulfill related studies. Main focuses of this paper are as follows:Firstly, people is the common subject of multi-view videos based on Social Camera, which makes the detection of people one of the research topics in the field of scene understanding of multi-view videos. Face detection algorithm is the regular used method for human-subject detection; however, the violent movement of camera often leads to missing detection. A face-subject detection based on CMT is proposed in this paper, which integrates motion information into face detection algorithm, improving the accuracy of detection for distorted face images in Social Camera videos.Secondly, the group detection discussed in the existing literature is mainly based on the attention on the spatial location and orientation of the subject, while Social Camera videos could not give corresponding information accurately. Therefore, in order to realize group detection, the paper puts forward a subject attention algorithm that takes into account the proportion of subject on the screen, the distance between subject and image center, as well as the features of subject orientation. The algorithm replaces the complicated three-dimensional calculation required by group detection with the analysis and understanding of the subjects in the images.Finally, directing at the problems of Social Camera video-limited dataset, over complicated system and inefficiency for inherent law studies of a single attribute-the paper created a Multi-Social Camera video dataset applicable for single study of subject attention attributes. Calculation and analysis of MSRA face detection algorithm, IRPM group detection algorithm. GCFF group detection algorithm, GTCG group detection algorithm and the algorithm proposed in this paper are all carried on the mentioned dataset.