Automatic detection of abnormal human behavior is one of the primary goals for a “smart” video surveillance system. In this research, we propose a system that is aimed to detect human carrying a weapon using images in video sequences. The methods include: (a) transformation phase; (b) feature detection phase; (c) feature extraction phase; and (d) arm-pose learning phase to classify if the human is carrying a weapon in images. Our results demonstrated preliminary success using a number of “positive” and “negative” images. Ultimately, this system could be incorporated with motion detection or motion tracking techniques, leading to provide a solution for law enforcement in particular area (e.g., banks, convenient stores, etc.).