New descriptor for object detection using an improved ensemble-based technique

Abstract

Object detection is an essential process for further tasks including, but not limited to, object and event detection, object tracking, object recognition, video indexing, motion estimation, image restoration, image registration, image retrieval, and reconstruction of 3D scene. In the recent past, interest point detectors and their descriptors, as local features, have received a great interest in computer vision areas and technologies. These types of features have shown their robustness against different types of deformation due to geometric transformation, photometric transformation and other disturbances. Therefore, they are more accurate and stable than the global ones. Among all interest point detectors and descriptors, the Scale Invariant Feature Transform (SIFT) and Speeded-Up Robust Features (SURF) are considered as the most common methods that receive interest from researchers in terms of usage and development; but, getting more accurate, invariant and fast descriptor is still needed. Matching technique is often used to recognize the object based on such features; however, it is not proper for some applications such as searching for an isolated object and it is difficult to be used in object category recognition or to recognize the part-based object. Therefore, learning-based technique, that has been proven to be an effective method in object detection, can be used to overcome the previously mentioned challenges. However, the object required to be detected usually represents a small ratio compared to non-object that causes an imbalanced data problem. The aim of this study is to design and develop an effective model for object detection that is faster, more accurate and it can manage aforementioned challenges. To achieve this goal first, new fast and an accurate descriptor is introduced based on interest points; second, an effective classification method, that mitigates the effect of imbalanced data, is designed based on developed ensemble classifiers; third, an updating scheme of interest point detector is presented to speed up the object detection system. Results show that the proposed features are faster and more invariant than the most common interest-point-based features. The developed technique based on ensemble classifiers produces notable results in terms of accuracy and False Positive rate compared to the traditional one. The speed of object detection system has increased by 30% in average based on the proposed scheme.