Carried Object Detection Of Walking Pedestrians

With the development of the society and science technology, intelligent video surveillance has found an increasingly wide utilization in the field of computer vision. The technology of surveillance is booming and advances rapidly. Video surveillance technology has been widely used in various public places and with the development of the technology and the decrease of its cost; it will enter average families and make greater progress in house security and entertainment application. As an important content of intelligent surveillance, carried objects detection can be applied to search for the lost items, prevent from theft and monitor the terrorist activities. It has broad prospects and received widespread attention of researchers.Although there are achievements in carried object detection, some problems still need to be solved and there is room for improvement in robustness and accuracy. Since the environment in real life is more complex, the difference of pedestrians’posture, movement or carried object will affect the detection result; on that account we study the detection techniques of carried objects and the main research work is as follows:First, we propose a method to estimate target’s movement direction. The estimation method divides the image plane into8areas and gets the movement direction according to the changes of pedestrians’position. Compared with traditional calibration estimation method, our method is simpler and the estimation result is more accurate.Second, we propose an algorithm for template matching. An EPFL exemplar is selected on the basis of movement direction estimation and transformations of the exemplar (scaling, rotation and translation) are performed to get the best match. After matching, we obtain the best matching exemplar and the protruding area.Third, the paper presents a classification approach combining fuzzy c-means clustering with Graph Cuts. The protruding area is separated into two categories, carried object and non-carried object and the clustering result is included in Graph Cuts energy function. We classify and segment the protruding area by minimizing the energy function. At last we get the carried object area.Finally, a detection system is built according to our method and some experiments have been performed to test the validity of the algorithm.