Research On Algorithm In Object Detection Based On Improved Edge Boxes

Object Detection is one of the most challenging topics in computer vision.It is widely used in the driving assistance system or intelligent surveillance.With R-CNN,Fast R-CNN and Faster R-CNN proposed,object detection is increasingly tending to apply the framework of region proposals.It means some object candidates are quickly proposed by some methods,then they are further classified by classifiers in order to filter what we want.This paper aims to study two key problems,one is the object proposing method,the other is feature extraction method.For object proposing methods,some are applying image segmentation to distinguish foreground and background,and some are using low-level feature to quickly differentiate foreground and background.Meanwhile,there are many feature extraction methods in the process of object detection development.Each has its own advantages and disadvantages.This paper aims to study this two key topics,object proposing method and feature extraction method.The main work and contributions are as follow:In the object proposing method,this paper mainly studies Edge Boxes,and analyzes its advantages and disadvantages.Against its disadvantages,we propose object saliency award and object location award to improve the performance of Edge Boxes in order to enhance its recall.Meanwhile,we compare between our improved Edge Boxes and other popular object proposing methods for showing the robustness of improved Edge Boxes.In the feature extraction method,this paper compares between HOG feature,sparse auto encoder feature and convolutional neural network feature according to the process of object detection development.This work aims to give advice for feature selection.Finally,this paper applies improved Edge Boxes into detailed application,vehicles detection.In samples collection stage,improved Edge Boxes is used to collect samples automatically.And in the vehicles detection stage,improved Edge Boxes is applied to propose object candidates for the following classifier.The experimental results validate that improved Edge Boxes algorithm proposed in this paper has a certain practicality and robustness.