Research On Key Techniques Of Moving Object Detection And Recognition In Complex Background

The main task of moving object detection and recognition research is to detect movingobjects in video sequence and determine an object belongs to what category. Due to thebackground interference, illumination changes, object shielding and object deformation, theappearance of same object will change under different conditions. So we must solve suchproblem in practical application and expand the use of moving object detection and recognition.Moving object detection and recognition has large potential applications such as weaponsautomatic tracking, video retrieval, computer vision-based human-computer interaction,intelligent video surveillance and motion behavior analysis. It has tremendous economicvalues.Combined with the theory of image processing, pattern recognition, computer vision andother fields, this paper did comprehensive and thorough research on moving target detectionand recognition in a series of problems. In this paper, we systematic focus on serials ofpractical problems from video background extraction to moving object recognition modeling.The research content of this thesis includes:(1) moving object detection in dynamicbackground;(2) object detection in static complex background;(3) moving object recognitionin dynamic complex background. According to these problems, the corresponding solutions areproposed in this paper. The main content of this paper are summarized as follows:1. First we researched on the adaptive weight updating Adaboost moving object detectionmethod in dynamic background. Moving object detection methods based on feature selectionand machine learning need to select a large number of features for training, and this will causehigher computation cost. Therefore, the Adaboost method, which uses multiple weak-classifersto generate a strong-classifer with higher detecting rate, has been widely used. The traditionalAdaboost method selects the best features to classify samples by adjusting the feature weights,but it is difficult to converge and classify complexity background samples because of featureweighting expansion. To solve these problems, a moving object detection algorithm based onadaptive weights updating Adaboost method was proposed. This method adaptively updatesfeature weights by calculating the training samples’ FNR and FPR. To avoid repeatedlymisclassified samples’ feature weighting expansion, this method also improves moving objectdetection rate and suppress error detection rate according to the type of sample distribution.Experimental results show that the proposed method can detect moving object in dynamicbackground without serious shield.2. Moving object detection method based on feature selection requires large amounts of training samples with labeled object location. However, manually label the object location isquite time-consuming and labor-intensive. Furthermore, in order to improve the accuracy ofobject detection rate, many methods use high order features to add spatial information, but alsoincrease the computational complexity. To solve this problem, an object detection methodbased on kernel function high order feature is proposed. The main idea is as follows. We firstconstruct the high order local features by using the kernel function. This reduces high orderlocal feature dimension and the computational complexity of feature selection. Then, theextracted features are used for training the feature code book, and this code book is used as theinput training samples of support vector machine to train the classifier. When detecting object,we extract the features of test image using kernel functions high order local feature extractionmethod, and use these features as the input of the SVM classifier to get the final results.Because of without manual intervention to extract kernel function high order features, themethod has the advantages of unsupervised learning.3. In order to reduce the influence of background noise, object recognition methods needto cluster the local features in complex background. But the traditional K means clusteringmethod has high computational complexity. In addition, most recognition methods requiredetection of object location first, and then recognition of object, therefore affect the real timeperformance. To solve this problem, moving object recognition based on feature tree model isintroduced in chapter5. This method consists of three steps: local feature extraction, featuretree construction and maximize the joint probability of object recognition. The method extractsHOG features and flow features, and then construct the feature tree with hierarchical K meanclustering method. Taking advantages of the fast lookup feature localization and classificationcharacteristic of tree-structure, we finally built the probability framework of object detectionand recognition, which reduces computation time.