Research On Object Detection Method Based On Invariant Features

Object detection is a hot research field of computer vision in the field of imageretrieval, intelligent transportation, intelligent video surveillance, advancedhuman-computer interaction has broad application prospects. Since the appearance ofdifferent object shapes uncertainty, complexity, mutual occlusion between the objectand between the object and background application scenarios cause object detectionbecomes a challenging problem. Although object detection technology has been studiedmany scholars, but so far there is not a universal, robust, accurate and real-time objectdetection algorithm. This paper mainly focus on the object image invariant featureextraction, and conduct research for different goals, object detection in case of differentscenarios. The contents and results are as follows:1. A detailed discussion of the key technologies of object detection, extraction andclassification description and the inclusion of a feature detection process selection. Inthe feature extraction and description, we introduced the theory of color, invariantmoments, SIFT(Scale invariant feature transform) features and characteristics of theimage contours and contour feature highlights for characterization. In the detectionphase, we were using the BP (Back Propagation) neural network and SVM (SupportVector Machine) as a classifier.2. By comparing various characterize the nature and amount of advantages anddisadvantages of proposed target detection method based on global and localcharacteristics. Explored the use of principal component analysis and BP neural networkapproach to solve the problem of object detection. The use of global and local featureextraction and detection methods combining training. Detected by visual similarity tothe object. Experimental results show that object a variety of feature fusion, principalcomponent analysis and reuse its dimensionality reduction, can effectively improve thecharacteristic dimension too time-consuming due to the impact, and improved objectdetection efficiency.3. It is difficult to detect objects in complex scenes, more noise around the objector object only a small portion of the image. In order to solve the problems, a new objectdetection algorithm based local contour features is proposed in this paper. Firstly,improved gPb (globalized probability of boundary, gPb) algorithm is used to extract theoutline of the image. Then the Otsu for automatic threshold processing is applied to obtain the significant contour. Next we extract k connected roughly straight contoursegments (k adjacent segments, kAS) formed by significant contour, and use kAS as alocal feature for object detection in complex scenes. The algorithm combines gPbalgorithm and Otsu to extract significant contour, thus it can remove much noise aroundthe object boundary, and effectively improve the detection efficiency as well.Meanwhile, in the detection phase, the number of irrelevant features in the test set andthe training set, has been largely reduced, therefore detection accuracy has beenimproved. Multiple sets of experimental results demonstrate the effectiveness of thismethod.