Research On Object Detection, Tracking And Recognition In Intelligent Video Surveillance

With the rapid growth of video surveillance requirements and the extension ofsurveillance area day by day, current technical method and manpower are difficult toguarantee the promptness and effectiveness of surveillance, and can’t meet the needs ofsurveillance. Therefore the demand for intelligent video surveillance system becomesmore and more urgent.Intelligent video surveillance (IVS) is to use computer vision and image processingtechniques to obtain the image sequence of motion detection, moving targetclassification, moving object tracking and monitoring the scene for the understandingand description of target behavior, and thus provides usefull and key information formonitors. At present, the intelligent video surveillance has been widespread concern ofscholars, and been launched an in-depth study. However, it is still in the phase ofexploration for the wide range of scientific subjects IVS involves, and the comlexproblems and application backgrounds IVS meets.It is a key point of the IVS systemthat researching for robust，precise, real-time and intelligent video content analysisalgorithms, which has important theory significance and application value. In thiscontext, we focus our attention on the key technologies of IVS, such as selection ofimage features, object detection, object tracking and recognition. The main work andinnovation in the thesis are as follows:(1) Considering the noise in information systems, we propose an approximateattribute reduction algorithms which can be used to deal with noise by loosening theconditions of reduction. The proposed approximate attribute reduction algorithm candelete the Information redundancy in SIFT features, and thus improve the effciency andaccuracy of matching algorithm.(2) A robust pedestrian detection algorithm in infrared imagery is presented. First,a keypoint sliding window searching strategy is introduced for candidate regionsgeneration, and thus the size of the sliding window searching space is thus reduceddramatically. Second, the multi-block LBP (MB-LBP) feature is used to represent thepedestrians. Considering the drawback of MB-LBP, that is lots of features, theapproximate attribute reduction algorithm is used to ruduct the information redundancyof MB-LBP. Finally, pedestrian detection is performed by SVM classifiers.Experimental results in various scenarios demonstrate the effectiveness and robustnessof the proposed method. Notice that the proposed algothm is also applied to pedestriandetection for visual images.(3) An object tracking scheme based on reducted SIFT features is presented. Itcan solve the feature selection, description, matching, object part-occlusion and shorttime complete-occlusion problem in the feature-based tracking method. In the scheme, the reduct SIFT features are used to descript the object efficiently. Then, the non-rigidmatching is applied to tracking the moving object. At the same time, object movementestimation method is embed into the feature-based tracking frame to solve the objectpart-occlusion and short time complete-occlusion problem. The paper focus onnon-rigid feature matching problem, and a non-rigid reduct SIFT feature matchingalgorithm based on geometric constraint is proposed. A SIFT features’ similarity basedmatching evaluation function (MEF) is defined. The MEF jointly considers theconstraint about geometric distance between features, thin plate spline transformationbetween the feature sets, and the entropy of the matching matrix. Based on this, thealgorithm turns the non-rigid feature matching problem to a MEF optimization problem,and solves the problem by deterministic annealing iterative frame. In each iterative step,the algorithm directly computes the matching and transformation between the featuresets alternatively, and reaches the optimal matching result iteratively.(4) An object feature-sets net (OFSN) modeling is proposed for object recognitionin intelligent video surveillance. In the method, local invariant feature based objectfeature-set (OFS) is constructed to describe the detected object in the image. Based onthis, the OFSN is constructed, OFS is the node of OFSN, and the similarity of the OFSsis the connecting line between the nodes. The OFSN model can effectively describe theobject at large imaging condition variety, can cluster training images of the same classautomatically into an OFSN tree, and can incrementally train and study images of thenew classes. These provide strong supports for object recognition. At the same time, wealso studied how to use Recursive Self-Organizing Map (RSOM) clustering tree toincrementally train and study the image object features, which can apply for efficientlarge scale image objects retrieval for OFSN model’s training and recognition. Whilerecognizing an image object, we can quickly find the candidate OFS by feature retrievalin the RSOM clustering tree, which realize recognition in the large scale image dataset.