Object Detection And Loaction System Based On PTZ Active Multi-camera

The research on intelligent multi-camera surveillance system is of great attraction. Systems of this sort, which include image processing, communication technology, industry control and other aspects of theoretical knowledge, are comprehensive.This paper aims to construct an object detection and location system based on active multi-camera. In the process of designing each module, many problems, such as object detection, object handoff and multi-camera coordination are discussed. The main content of this paper is as follows:1. This paper elaborates the theoretical analysis of detecting and locating objects using active multi-camera. We also analyze overall framework of the system and technologies used in each module.2. This paper studies algorithms of object detection and tracking in single camera streams. In order to meet the requirements of the system, we choose appropriate algorithms of detection and tracking. For object detection, this paper studies optical flow method, frame difference method and background modeling method. For special pedestrian detection, we extract the HOG feature, collect positive and negative samples in INRIA library and train classifier. According to the actual system of this paper, the pedestrian detection through the movement of detection window on the picture is avoided. Instead, after the detection of object, we change the scale of the object area to fit the detection window. Then we determine whether the object is a pedestrian by means of the HOG+SVM classifier. For object tracking, this paper mainly studies particle filter algorithm and CamShift algorithm.3. This paper studies the algorithms of object handoff among multi-camera. For object handoff between multi-camera streams, two kinds of situations are discussed. For multiple cameras with overlapping view, we conduct object handoff using homography matrix. We extract the feature points using SIFT feature, and remove the mismatch points using RANSAC algorithm. We can get the handoff conditions of objects through the homography matrix. When the objects’ distribution is relatively concentrated, the original algorithm is not effective to get the handoff conditions of objects. In this case, the solution to object handoff of this paper is using D-S fusion of chromaticity feature of objects and SIFT feature of object areas. For multiple cameras without overlapping view, we conduct object handoff by means of multi-feature fusion. This paper fuses SIFT feature, shape feature and chromaticity feature of objects by D-S theory.4. This paper studies the algorithms of object location. To avoid the complex camera calibration, the location of object is obtained through the geometric position of cameras. We calculate the two-dimensional coordinates of objects by means of a multi-channel video intersection algorithm.5. This paper studies the algorithms of coordination of camera network. This paper mainly studies two algorithms: the coordination algorithm based on MDP and the coordination based on POMDP. According to the actual situation, this paper improves the algorithms and introduces the object priority into the algorithms.6. We design an object detection and location system base on multi-camera, and verify the validity of the system.