ビデオからの移動物体の検出に関する研究

摘要

In recent years, the video surveillance system for security and a driving safety system on the car have been growing rapidly. The video surveillance system has grown from a manual system to a fully autonomous system, whereas a driving safety system has evolved from a parking safety system to a collision avoidance system. The system requires a good ability to detect a moving object so as to be a reliable system.The problem that must be addressed in the detection of moving objects on a video is a dynamic background. In this thesis, we proposed a moving object detection method using sequential inference of the background to overcome the problem of the dynamic background. The sequential inference of the background uses a series of previous image frames to create a model of the background image for the current frame. After a background model is obtained, then the background subtraction can be done.The proposed method is applied to the video captured using a static camera and a moving camera. The detection of moving objects in a video captured by a moving camera is not as easy as the case using a static camera. Correspondence of pixels in the current image frame with the pixels of the previous image frame must be known in advance. The background model is formed using a bilinear interpolation of the previous image frame. The judgment of a pixel as the background or the foreground is done by subtracting the model of the background image from the current frame.An important stage in this method is updating normal distribution of the pixels on a background model. A background model is formed based on the value of the normal distribution which is updated with each frame of a video. The originality of this thesis is to propose novel ways of updating the normal distribution to obtain an effective background model.Experiments are performed on several videos. The results show that the proposed method can detect and extract moving objects that appear in a video scene successfully under various situations of the background. The effectiveness of the proposed method is recognized by recall, precision and F measure.