Research On Typical Object Detection And Change Detection From Unmanned Aerial Vehicle Images

Unmanned Aerial Vehicle(UAV)has been widely used in many fields.Object detection and change detection methods for UAV images play an important role in both military and civilian applications.Although object and change detection is always a hot research field,robust and efficient object and change detection are still a challenging problem because of object appearance variations and complicated background.In this thesis,object detection and change detection methods for UAV images are studied,and the man contents and contributions are as follows:1.An online cascaded boosting framework with histogram of orient gradient(HOG)features for car detection from UAV images are presented.First,the primary gradient direction of the sliding window is used to estimated the car's orientation,and then the sliding window is rotated according to the estimated car's orientation to perform feature extraction and classification.Second,to efficiently compute the HOG features in the rotated window,two fast HOG features extraction methods are proposed,which are based on integral image and circle filter respectively.Third,an efficient online cascaded boosting framework for car detection is poposed by combining online boosting with soft cascade,and two improved weak classifiers based on Fisher discriminant analysis and Bayesian discriminant analysis are employed to enhance the performance of the car detection method.2.It is not very efficient to compute rotation-invariant features based on radial gradient transform and gradients accumulation in annular regions.This problem is particularly serious when this kind of features is densely computed across the whole image to perform object detection.An accelerated rotation-invariant HOG features based on Gaussian filter is presented for this problem.First,a look-up table and the polar coordinators are used to speed up the radial gradient transform.Second,Gaussian filter is employed to accelerate the gradients accumulation in annular regions.The proposed accelerated rotation-invariant HOG features are used to perform airplane detection,and the results show that the proposed features achieve equivalent detection performance but is much more efficient.3.The recently proposed rotation-invariant HOG descriptor using Fourier analysis in polar coordinators is time and memory consuming.An improved efficient rotationinvariant HOG descriptor is presented based on this method.First,spatial convolutions are convert to multiplications in the frequency domain based on fast Fourier transform to speed up the descriptor computation.Then,a backward search feature selection method based on support vector machine are employed to reduce the descriptor dimensionality,and as the descriptor dimensionality reduces,the computational time for generating features reduces too.The proposed efficient rotation-invariant HOG descriptor is used to perform car detection,and the results show that the proposed descriptor yields better detection performance at low time and memory cost.4.For change detection problem from UAV images,a method based on nonrigid image registration and graph cuts is proposed.First,after image coarse registration based SURF features and homography,image fine registration is completed using B-spline based transformation model.The interference of local registration errors on change detection is reduced significantly.Second,several local features are extracted and transformed to a new feature space based on slow features analysis,which makes the change detection method more robust.A min-max bidirectional difference image is used to reduce the effects of local registration errors on change detection.Lastly,a global objective function is constructed based Markov random field,and the graph cuts method is employed to solve this function to obtain the change detection result.In the global objective function,not only the change information of the single pixel is considered,but also the spatial relationship between the pixels.Therefore,the isolated noise and cavitation in the change detection rusults are significantly reduced.