Object Relationship Based Dangerousness Analysis In Scene

Over the past decade, research in the field of computer vision has a greatdevelopment. Among them, scene understanding also has attracted more and moreresearchers’ interest in this fie ld and a series of delightful fruits are achieved. Themain research contents in visual scene understanding can be divided into three parts:object detection, scene classification and geometric spatial reasoning. And each parthas been deeply researched. However, although the researchers made greatcontributions to scene understanding, there are still many problems which need tobe further explored.The main contents of this paper are to analyze the dangerousness existed in thevisual scenes of images and videos. It belongs to the realm of scene understanding.To our knowledge, there is not any research related to this. Detection andestimation of the possible dangers hidden in visual scenes is of important realisticmeaning, especially in public places which have a dense crowd (e.g., railwaystation, the airport waiting hall and town square). If any danger is discovered, wecan take corresponding measures in time to safeguard life and play down the lossto the minimum.For static scene images, we propose a framework of dangerousness analysisbased on objective relations. The deformable part model (DPM), which is one of thebest object detectors, is applied to search the target objects related to dangerousnessin scenes. The objects considered by us include human and gun. And we gain afavorable detection results in the experiment. After obtained the spatial position ofobjects, the relationship between objects can be modeled according to definitivedangerousness related criteria. Then we treat the relationship as feature descriptionand utilize method of regression analysis to estimate its dangerousness.Different from static scene images, dangerousness analysis in videos has arigorous demand for real-time performance. So the deformable part model can’t beused to handle it. Here we employ the tracking-learning-detection (TLD) framework,which is the state of art tracking algorithm, to track the relevant objects. Besides,the excellent deformable part model is joined with it for assistant optimization.Finally, in consideration of a variety of human-dominated scenes, variouskinds of dangerous situation always have some connection with human pose onsome level. The problem of human pose estimation can be regarded as a kind ofmaximum a posteriori inference (MAP) built on Markov random field (MRF). Wepropose to solve this problem by a branch and bound algorithm. And the human pose information captured by branch and bound approach can provide more usefulclues for dangerousness analysis in scene.