1 Introduction||2 Fast Motion Detection based on a Velocity and a Shift Histogram||3 Fast Motion Detection based on a Shift Histogram||4 Head Motion Detection based on a Histogram of Transition||5 Final Experimental Results and Evaluation||6 Conclusion, Nowadays, there is an increasing demand for security system by continuously using CCTV systems. Most current surveillance systems need a human operator to constantly monitor them. Their effectiveness and response is largely determined not by the technological capabilities but by the vigilance of the person monitoring the display. To overcome these limitations of traditional surveillance methods, a major effort is under way in the computer vision and artificial intelligence community to develop automated systems. An Automatic Vision Surveillance （AVS） has become an important topic for the academic community, which aims to develop autonomous surveillance schemas to replace the traditional schemes. The human motion detection can be performed in two approaches: The first one is a flow or motion based approach. The second approach is head motion based approach. These approaches are usually used to detect an abnormal motion. The existing methods of the first approach have some inadequacies: Some of them rely on feature points extraction and tracking. Owing to the lack of feature points, it increases the false negative value; Some of them perform the observing of flow directions only; Some of them fail to detect an abnormal motion when the motion flow is not perpendicular with camera direction; Some of them rely on the training process; Some of them need a camera observation from very high position. Some of them should be applied in extremely crowded scenes. The existing methods of the second approach have some inadequacies: Some of them rely on foreground extraction and very sensitive with a pattern which similar to a head. Some of them need a complex computation and big vector dimension. In order to make up for the inadequacies of the existing human motion detection based on the first approach, we propose the improvement of the motion flow method. In this research, our information on motion flow is provided by the analysis of motion history image （MHI）. Our method gives more accurate as well as important information for fast motion detection in a condition where the motion is not perpendicular with a camera view. Another improvement is based on the second approach. In this thesis, we introduce a histogram of transition as a novel feature. Since this feature calculates the transition between the background and the foreground, the computation is very simple and takes a short time compared with the computation of LBP and HOG feature. The originalities of this thesis are as follows :In the first place, we introduce a shift histogram based on MHI representation. To the best of our knowledge, this is a new method to get shift information. Most of the existing methods use optical flow generated by Lucas-Kanade tracker or spatio-temporal gradient .In the second place, we apply an accumulative function of the shift histogram to detect fast motion as an anomaly motion in a crowd. This approach does not necessitate motion learning. Most of the existing methods need a learning stage to learn normal and anomaly motion in a video clip. In the third place, we introduce a function distance and a refined frame differencing as foreground extraction. Foregrounds which extracted from a static image and frame differencing images give an accurate detection of a head. They can distinguish a head and a pattern similar with a head. In the fourth place, we introduce a histogram of transition feature as a feature to be fed into a classifier. The computation of this feature is simple and needs smaller computation time. It is suitable for real time application. The dimension of the employed feature vector is small., 九州工業大学博士学位論文 学位記番号:工博甲第383号 学位授与年月日:平成27年3月25日, 平成26年度