OBJECT DETECTION USING SHARED LOCAL FEATURES

Abstract

To detect object which is in different views from the cluster of images is still a challenging task. In our proposed method, we developed a new method to detect an object from different views by using shared local features in partially occluded images. Here shared features are the common features which are obtained from different classes and these common features are trained jointly in order to reduce the no of classifiers to detect an object. First we select some random samples (rectangular boxes) with different sizes which cover entire image. Each sample is represented by centre point of rectangular box, length and width of box. After selecting random samples, we apply oriented centre symmetric local binary pattern(OCS-LBP) & HOG for each sample and trained by random forest classifier. Like wise apply the same procedure for different views and generate shred features from all different views. To detect an object view point and its location we use probabilistic method for all views and which has the highest probability that view becomes the view point of detected object. Our proposed method was successfully performed on PASCAL VOC 2012 dataset and obtained better results compared to other methods.