Abnormal Target Recognition and Application in Smart Function Area Based on Video Sensing Technology
Author:Jing Wu, Chunxue Wu, Yan Wu, Naixue Xiong Author Unit: School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China Public and Environmental Affairs, Indiana University Bloomington, IN 47405, USA Department of Mathematics and Computer Science, Northeastern State University, Tahlequah, OK 74464, USA
Abstract：Video recognition, which can effectively detect the abnormal objects in the monitoring area and judge their
identities so as to achieve the purpose of real-time monitoring, is an important part of the smart function area
based on video sensing technology. Different from the object detection for static images, video recognition
should focus on the foreground objects and ignore the background in the video. In this paper, an abnormal
target recognition model (ATRM) is proposed. ATRM focuses on a small amount of foreground objects and
ignore irrelevant background. Since the accuracy of foreground object positioning is high, the bounding box
regression operation and some other time-consuming technologies are not required which further reduces
the detection time. Taking full account of the continuity character of video sequence image, ATRM obtains a
small number of Regions of Interest (RoIs) by tracking and positioning the dynamic foreground targets just in
part of video frames. This model can effectively identify the foreground object and has a great performance
in practical application of the smart function area based on video sensing technology. In this paper, the
model is applied to the smart campus monitoring system. ATRM achieves a mean average precision (mAP)
of 78.6％ on our campus area surveillance video set. Compared to Fast RCNN, ATRM greatly improved the
recognition speed and detection precision for some specific objects.
Keywords：Smart function area; Dynamic foreground object; Regions of interest; Object detection; Feature fusion; Convolutional neural network
Download information [SIZE：1.85 MB Amount of downloads： second]
