Research On Multi-object Detection Algorithms Based On Shape And Feature Fusion

Along with the development of computer and network technologies,images and videos are widely used due to their rich contents.It has become a hot topic to localize the useful information from massive data by analyzing images and videos.Multi-object detection aims to get the positions and classes of one or more objects in an image.It has been widely used in pedestrian detection,intelligent monitoring,and many other fields.It's a challenging job to detect objects in an image accurately and rapidly,because objects can be affected by uneven illumination,viewpoint variation,different postures,scale variation,rotation,and so on.This paper makes some research on object detection methods based on shape.Currently most of object detection methods are based on the sliding window mechanism and it is very time-costed.This can be optimized by generating object proposals using objectness estimation methods first and then applying class-specific detection among these object proposals.For those objects with good shapes,shape detector can be used to locate them rapidly and accurately.This paper proposes a novel method of objectness estimation for objects with circular shapes.Firstly the edge map of an image is extracted using Sketch Tokens.And then GPU based Hough transform is applied to extract all the circular object proposals.Finally we get CNN features of all the proposals and classify them by SVM.In 8 ball classes of ILSVRC dataset,our method achieves a MAP of 34.33%.This paper also makes some research on object detection methods based feature fusion and presents a multi-object detection method based on feature fusion of CNN and HOG.CNN feature is a kind of high level feature generated by deep learning.When the training data is abundant,CNN can find the nature of the data and have a good ability to represent data.HOG feature is a kind of low level feature based on image gradients and can eliminate the effect of illumination and slight variations.It has made a good performance in object detection task.So this paper combines CNN and HOG features for multi-object detection task.First we generate about 300 object proposals per image using our objectness estimation method.Then we extract both CNN and HOG features of these proposals,and train them into models separately.For a given window,the total score is a weighted sum of the two models' scores.Again in the 8 ball classes of ILSVRC,the combined feature achieves a MAP of 35.08%,and can at most raise the AP of a single class by 3.83%.