Research On Bird Detection Algorithm Combined With The Deformable Part And Aggregate Channel Features

Bird detection as an important aspect in the object detection,has a broad application prospect and practical value in bird repelling system,video supervision,ecological of health analysis and other fields,bird detection has attracted extensive attention and in-depth research in the field of computer vision.However,there are a wide variety of bird and bird is non-rigid object,and the appearance changes greatly,and there are different degrees of light changes and occlusion in the natural scene.The object detection algorithm based on the Deformable Part Model(DPM)achieves better results.This paper based on DPM,conduct a research on the bird detection in natural scenes,and the main work is as follows.1.DPM algorithm using PASCAL VOC date set,the date set only contains a few bird samples(training set have 297 samples,test set have 282 samples),and most bird samples is not in natural scene.In view of this problem,this paper select and analyze bird samples from other date set at first,then choose 1500 birds pictures which is representative in appearance and posture,they can basically include the general characteristics and attitude of all bird,and finally establish a specialized bird database for all bird samples manually.2.Aiming at single HOG feature in traditional DPM algorithm description ability is limited,the robustness is not strong,in view of this defect.this paper on the basis of the DPM,considering the Aggregate Channel Features(ACF)compared to the single HOG,increased the other types of features channel,which can better describe the characteristics of bird.Therefore,this paper puts forward a new kind of bird detection in nature scene method which combines with DPM and ACF features,build a new ACF-DPM model used for bird detection,both theoretical analysis and experimental results indicate the validity of the model,in the complex natural scene,the model can effectively detect birds,the overall accuracy is better than the other detection algorithm.3.On the basis of the ACF-DPM model,the optimization of the model effect is mainly included: the optimization of the number of components and parts in the ACF-DPM model;the optimization of the size of each filter in the ACF-DPM model.Experimental results show that the optimized ACF-DPM model has betterdetection effect than the original ACF-DPM model.4.For the problem of ACF-DPM model easily missed object after shade,this paper proposes a weighted ACF-DPM model of birds by a new method for weight calculation,to determine the importance of different parts,different weights of the filter are given to different parts of each component to make the model more consistent with the characteristics of birds.In order to verify the validity of the improved model,chooses the data set this paper established and standard data sets PASCAL VOC 2007,and add shade birds pictures to test,compared with the original ACF-DPM model of birds and other detection algorithm,the experimental results show that the weighted ACF-DPM model was improved in performance,and better than the other detection algorithm.