A Three-stage Infrared Person Perception Scheme for Security Patrol Robot

A three-stage infrared person perception scheme is proposed for security patrol robot. Firstly, an active contour model, which provides enclosed curves and defines relative complete target interiors, detects the foreground regions that may contain persons. Secondly, feature extraction of histogram of oriented monogenic energy (HOME) and pattern inference by deep brief network (DBN) runs for person recognition. Thirdly, in a frame-by-frame way, a multiple object tracking strategy runs to track the persons that exist in the scenes. As whole, the proposed scheme copes well with the poor quality of infrared images, the casual transition of robot motion modes, and complex modality of person targets as well. Experimental results shows that the proposed scheme has the advantages in perception accuracy and robustness.