Research On Object Recognition Based On Shape Description

As an important topic of computer vision,object recognition has been studied for a long time by researchers all over the world.With the development of computer vision,the research on object recognition is more and more significant in practical application.In this paper,the shape of object contour is used as raw data for the research on shape description,contour simplification,shape matching,object recognition,and the application of hand gesture recognition.Firstly,a novel object recognition method based on invariant multi-scale descriptor is proposed.The shape is represented by the invariant multi-scale descriptor,and the contour is modified by adaptive discrete contour evolution algorithm in this method.Then,the dynamic time warping method is used to match two shapes and calculate the similarity between them.In order to utilize the prior knowledge of shape database and improve the accuracy of object recognition,a metric learning based method is proposed.Firstly,a method called bag of salient feature points is defined.Then,the feature points,which are described by the invariant multi-scale descriptor,and extracted by the adaptive discrete contour evolution algorithm,are clustered to learn the histogram representations of shapes.Moreover,the relevant component analysis algorithm is employed in shape matching to assign large weights to the relevant dimensions and small weights to the irrelevant dimensions.At last,the Euclidean distance between the histograms of different shapes is calculated for object recognition.To analyze complex shapes,a new invariant multi-scale descriptor is proposed in this work.This method introduces two new invariant parameters: the difference of area and the difference of arc length,which is treated as the supplement of the invariant multi-scale descriptor.The experimental results show that this method has good properties of invariance,resolution and robustness.Meanwhile,the relevant object recognition method achieves excellent results on the benchmark shape databases.At last,the proposed methods are employed on the application of hand gesture recognition based on RGB-D data.With the RGB-D data provided by Kinect sensor,the proposed shape recognition method is integrated into a real-time hand gesture recognition system based on the invariant multi-scale descriptor.This system achieves 96.4% recognition accuracy and use only 0.0387 second for each query in the hand gesture recognition experiments,which verifies the accuracy and efficiency of the proposed method.