Research On Small Dim Target,Salient Object Detection And Model Based Pose Estimation

Object detection and pose estimation are key technologies in weapon system test,Unmanned Aerial Vehicle(UAV)reconnaissance and strike,precise guided weapon,equipment allocation,autonomous landing,etc.Although they have been stuied for a long time in computer vision,robust and efficient object detection and pose estimation are still challenging problems because of clutter background,noise,complex light condition,texture variations,etc.In order to fulfill the requiments of applications,this dissertation starts to research on object detection and pose estimation in typical cases,including small infraered target detection from a single image,visual enhancement for dim infrared target in video,cloud detection in down-looking infrared images,salient object detection in natural images,3D model based target pose monitoring form a monocular image.(1)An infrared small target detection method from a single image based on spike suppression in frequency domain is proposed.Traditional methods always model the background as low frequency components and the target as a Gaussian spot in the infrared small target image.They have troubles in dealing with clutter backgrounds and targets of arbitrary shapes.Different from the traditional methods,the proposed method models the background and the noise in the infrared image as spikes of the amplitude spectrum in frequency domain and makes no assumption about the small target.This dissertation suppresses the background and noise through suppressing the spikes.In order to highlight larger size targets uniformly,the segmenation result in spatial domain is introduced.Comparing with the traditional methods,the proposed algorithm detects small infrared targets effectively,even for images with multi-target,targets of arbitrary shapes,clutter background,etc.(2)A visual enhancement method based on energy accumulation for infrared small dim target in video is presented.Traditional enhancement algorithms tend to over-enhace the background,especially for images in which the target is so faint that hardly visible.This dissertation introduces temporal cues to perform the visual enhancement.The target intensity is enhanced via accumulation along the target tarcks in temporal domain.According to the characteristics of the target's shape,the proposed method suppresses the clutter background around the target.The enhanced target region is embedded into the original image adaptively by using the target intensity as the weight.Comparing with traditional methods,the proposed algorithm can enhance the small dim target effectively and improve the visual quality of the infrared small dim target image significantly,even for images which contain hardly visible targets.(3)A method based on texture classification for cloud detection in down-looking infrared images is proposed in this dissertation.Comparing with the underlying surface,the cloud textures change smoothly and exhibit heavily correlation in spatial domain.Based on the texture differences,this dissertation detects cloud using support vector machine.The proposed method is consist of three parts: pre-processing,feature extraction and classification.We build a dataset using large numbers of down-looking infrared images collected in this dissertation.Experiments are conducted to evaluate the proposed method.Results show that the proposed method can detect the cloud in down-looking infrared images effectively and efficiently.(4)Traditional salient object detection algorithms have troubles in highlighting the salient object in natural images completely and uniformly,especially for salient object consists of several parts with different appearances.This dissertation solves the problem by introducing objectness.The role of the objectness in salient object detection in natural images should contains two aspects: salient region detection and consistency enforcement.This dissertation calculates the region-wise saliency by fusing objectness,uniqueness and center bias.By treating the regions as nodes,this dissertation models the natural image as a Gaussain Markov Random Field.The edge weight is calculated according to the objectness.Consistency enforcement is conducted by minimizing the energe function.The salient object is highlighted completely and uniformly in the final result.Experiments performed on several widely used benchmark datasets indicated that the proposed algorithm detect salient objects in natural images accurately,completely and uniformly.(5)A method for 3D model based target pose estimation form a monocular image is presented.3D model makes it ease to bulid the sample database and fulfill use the 3D information of the target.For target detection and initial pose estimation,the proposed method achieves them using classifier traind on sample database.In order to eliminate the adverse effects of complex illumination and texture variatios,only structure information is used in traing the classifiers.The following pose tracking is achieved by finding the pose which maximizes the image segmentation into foreground and background.Comparing with traditional methods,the monitoring method built in this dissertation doesnot need cumbersome database preparation and performes well in target pose monitoring.It robust to variations of texture,target size,target number and illumination,etc.