Detection of Suspicious Person with Kinect by Action Coordinate.

Surveillance cameras have been popular as measures against criminals. They are useful for deterrence of climes such as theft and murder, but they cannot prevent climes since real-time watching must be required for real-time detection of the climes. On the other hand, technologies of automatic recognition of people have been developed and the accuracy of image recognition by artificial intelligence has especially increased. That is, it is possible to detect crimes by installing artificial intelligence to surveillance cameras. However, a large amount of behavior data must be learned to determine a behavior pattern of a person by image recognition, and the data size is large since they are image data. Also, it is not confirmed whether a behavior pattern can be judged precisely when a part of the body is hidden by a shielding object such as another person. In this paper, we propose a method to judge a behavior patter of a person using Kinect in order to detect crimes at real-time. When a suspicious person is standing in front of the entrance door, there are cases where more than one person stands at the same time. In those cases, a part of the skeleton coordinates of the person who unlocks the door may be hidden, and it may be difficult to distinguish between behavior patterns. Therefore, in this research, we propose a method for detecting suspicious person even when skeleton coordinates are hidden by shielding objects. As a result of experiments which distinguish the picking operation from the unlocking operation by a key, the average of F-measure was 47%. We know that the score was quite poor to distinguish them. On the other hand, the average F-measure of some examinees showed more than 80%. We also mention about the value and the reason as consideration in this paper.