The Structural Model Learning Based On Local Features And The Application In Object Detection And Localization

Object detection is an important research field in image processing and computer vision. It is the foundation of various applications, such as image understanding, automatic image segmentation, motion tracking, scene understanding and etc. The complicated image, formed in the real-world scene, may exist illumination change, scale and perspective change for object, occlusion, strong background noise and so on. It is difficult for object detection and recognition method to deal with various changes in the complicated image. The structure information, constructed by the significant local information of the object is always the stable characteristic of object. For example, the facial features of human face (eyes, nose, mouth) is a typical example. The local feature is defined by a significance criterion to distinguish its neighborhood image representation, and it has the flexibility to handle a variety of changes in the image (geometric changes, illumination changes, affine distortion, occlusion, background confusion). Therefore, it is feasible to use the local features and the spatial structural model to solve the object detection problem for the complicated images.In this paper, we present a structural model of learning algorithms using SIFT feature and random forest classifier, and the probabilistic voting model is used to detect and locate the object in the complicated images. The main procedures as following:1) In structural model learning, The random forest classifier are constructed by the SIFT features extracted from the training images. For each leaf node of each decision tree, the shift information of the local features to the object center (offest) and the class information of these features, together with the class information of this leaf node are stored. Therefore, all leaf nodes construct a discriminative tree-structured codebook mode.2) In object detection, first, the SIFT features, extracted from the test image, are roughly classified by the random forest model, and the discriminative codebook is used to estimate the object's location via a probabilistic computation, called probabilistic Hough vote, by the generalized hough transform. These probabilistic votes have a better accuracy and reliability. Then, the probability density peak region, also called the candidate object center, are searched in the probabilistic voting space. In order to reduce the mistake detection and improve detection accuracy of our algorithm, the candidate object center should be verified by the low-level segmentation information.We evaluate and analyse the performance of our proposed method on the multi-views motorbike image dataset and the UIUC-Cars image dataset. The experimental results show that the proposed algorithm can provide a better detection results even in a complicated environment such as scaling, occlusion and strong background noise.