Research Of Object Detection Based On Improved Implicit Shape Model

Object detection is an important research area in the computer vision community and has been extensively investigated in recent years. It is the fundamental module of other subsequent applications such as object tracking, image segmentation and scene understanding. According to the different types of feature used, we divide current object detection methods into global features based methods and local features based methods. Because local features based methods can better deal with geometric transformation, illumination variation and partial occlusions of objects in the image, they have been widely used in recent years.In this paper, we propose an improved object detection model based on the Implicit Shape Model (ISM). In particular, we apply different types of local features to learn object spatial structural models by a Random Forests framework. However, ISM uses unsupervised clustering method to generate unreliable codebook and employs only one kind of feature to dercribe an object. Instead, we improve the ISM framework in the following two aspects:(1) We propose to optmize the codebook using a supervised learning method. The codebook in ISM is generated using an unsupervised clustering during the training, which leads to many unreliable or redundant offsets in it and causes unreliable probabilistic votes. Thus, we use a Random Forests framework to learn the spatial structure of local features. All the leaf nodes in each decision tree form a discriminative codebook model. Applying the labeled local features to construct the random forests, we optimize the codebook and make the object structure more reliable. Meanwhile, we address a verification method based on BING (Binarized Normed Gradients) to further improve the detection accuracy.(2) We propose to generate the object codebook using mixed types of local features. Because different types of local features characterize the object from different aspects, a combination of these features provides a comprehensive representation for the object. It is worth onting that only one local feature type (grayscale value or SIFT) is used in the ISM for object representation. We use a combination of different kinds of local features to learn object codebook. By fully employing different features’advantages, we can make the model with better classification ability.Experimental results indicated that training the model in a supervised way can make it more reliable and get more precise probabilistic votes. Object description with mixed kinds of features makes the codebook include more rich information and provide better detection result compared with single feature description.