| With the rapid development of video capture technology and large-scale data storage technology, large camera networks are increasingly deployed in public places. The traditional manual monitoring technology has been difficult to cope with the resulting massive videos, and intelligent video surveillance technology is becoming the main way to solve the problem. As an emerging topic in the field of intelligent video surveillance, person re-identification has attracted wide attention of researchers. Person re-identification is to identify the same person's images in an existing database coming from non-overlapping camera views. Images for re-identifying are mainly captured by different cameras. Therefore, this topic faces challenges of perspective change, posture change, scale change and illumination change. How to re-identify person accurately and efficiently is the current research direction of this subject.This dissertation focuses on person re-identification in the intelligent video surveillance system. It firstly extracts moving pedestrians in surveillance videos by using the background subtraction technique, and then re-identifies pedestrians by the methods of main color matching and keypoints matching. When the target pedestrian is detected, it adopts a target tracking method to identify. Finally, the dissertation designs and implements a person re-identification system.In the respect of pedestrian detection, a Vibe person detection algorithm based on Lab color space is proposed. Firstly, the algorithm establishes a background model of video scene by an improved Vibe algorithm, makes difference between current frame and background model and obtains moving foreground. Then, it extracts human body from moving foreground according to ergonomic theory. For making Vibe algorithm insensitive to illumination change and object shadow and extract movement area incompletely, this dissertation adopts CIE 1976 Lab color-difference formula with weight coefficient to measure distance between pixel points and sample points and makes use of space consistency of pixels to correct pixel classification results.In the respect of person re-identification, a person re-identification algorithm based on HSV color space and keypoints matching is proposed. It firstly utilizes an improved color quantization strategy of HSV space to pre-test pedestrian images and quickly rule out images of which main colors are different from the target. For the remaining images, it firstly takes advantage of circularly symmetrical Gabor filters to generate multi-scale images, then extracts and describes keypoints by FAST algorithm combined with Shi-Tomasi algorithm and BRIEF algorithm, matches and purifies keypoints by Brute Force algorithm and Random Sample Consensus algorithm. Experimental results show that this algorithm, which achieves high recognition accuracy and fast execution speed, is invariant to image scaling and rotation and robust to perspective change and image noise.In the respect of pedestrian tracking recognition, an enhanced Mean Shift tracking algorithm is proposed. It utilizes changing bandwidths of kernel function for tracking, updates the target model when not being occluded, and introduces Kalman filter to predict target position. Experimental results show that this algorithm, which can meet the demand of real-time image processing, is insensitive to scale changes of targets and occlusion.Finally, this dissertation makes use of Microsoft Visual Studio 2010 and OpenCV to design a person re-identification system, which is based on MFC dialog and has the functionality of semantic recognition and example image recognition. The system consists of human-computer interaction module, pedestrian detection module, pedestrian feature extraction module, pedestrian recognition module and recognition results display module. It applies the technologies of object-oriented programming and MFC widget redrawing and makes users obtain pedestrian targets from surveillance videos easily and quickly. |
PDF Full Text Request