Moving Object Detection And Recognition Based On Multi-feature

How to detect and recognize moving target quickly and accurately is important in the field of computer vision and image processing. This paper mainly studies the moving target detection and recognition method based on multi features, the main research contributions are listed as follows:For the problem of detecting moving objects in complex scenes, combining rgb color feature and scale invariant local ternary pattern is presented for moving target detection.Firstly, the temporal median method is adopted to estimate background image and initialize background model quickly. Fusing the similarity measure of color and texture features, to obtain the probability of the pixel as background. Furthermore, apply the lateral inhibition filter model to the probability map to enhance the contrast to sort out the foreground and background pixels. Using the background pixel to update the background model, on the other hand, shadow detection is made for the foreground pixels, and the shadow pixels are classified to background pixels but not used for the model update. The experimental shows that the proposed method can accurately handle scenes containing moving backgrounds,shadows and Camera Jitter in acceptable time.In view of the traditional template matching method in target recognition, only use the single shape descriptor and compare the similarity just between the two shapes, we propose to combine the two shape descriptors in the binary images and match the template under the manifold learning. Firstly, extract the contour points equidistant, and use the inner distance shape context descriptor and height function descriptor to describe the shapes in different characteristics. Secondly, use the two descriptors to shape matching respectively in dynamic programming under different distance metric. And we add the two distance matrixes after respective normalized. Finally, we compare the shape similarity by considering the underlying structure of the shape through manifold learning. The experimental shows the proposed algorithm can improve the accuracy of the target identification and in MPEG-7 dataset, the bull’s eye score is 93.49%.For edge extraction of moving object in real images is very complex, and the methods based classifier is more general than the template matching, presenting a object recognition based on bag of words model and feature fusion. Firstly, using a fixed step size and fixed scale-intensive to extract key points, and then extract the Sift and LBP around the key points in the grids to describe the shape features and texture features. Secondly, k-means clustering algorithm is introduced to generate a visual dictionary, and the local descriptors are encoded by approximated locality constrained linear coding, and max pooling and spatial pyramid matching are used to generate the feature representation. Finally, the feature representations are fused in image level and sent to the svm for classification. The experimental shows the proposed algorithm can improve the recognition accuracy rate, and can achieve good results in the case of recognition with less training images.Finally, for the one moving object detection method and two object recognition methods,considering the template matching method will failure when the perspective change is toolarge. And the statistical pattern recognition method based on the trained classifier is more suitable than template matching. So we combine the moving object detection method and the latter object recognition method based on classifier to constitute the moving object detection and recognition system based on multi features. The experiments show that it can achieve a good recognition rate.