On Algorithms Of Detecting,matching And Tracking Of Local Features In 2-Dimensional Images

Local features in 2-dimensional images are a kind of low-level features,which have been broadly applied to researches on machine vision technology.Techniques of detecting,describing,matching and tracking over local features have been developing for decades,and researches on them are still vivid so far.Focusing on problems of detecting local features,matching local features and tracking objects by local features,this thesis achieves the following four works:(1)We present an algorithm of extracting scale-invariant local features with high repeatability.Many classical algorithms of extracting scale-invariant local features build on scale space representations calculated from discretized kernels.However,these algorithms do not consider quantization error,which can influence their performance.From the scale space theory,it is known that the computational error is bigger when the scale is smaller,and it tends toward zero as the scale increases.According to this,we suggest a scheme which increases the kernel width with only one pixel per iteration,aiming to mitigate the drawback of sampling scales by doubling intervals,which leads to the case that the data points are denser in finer scales whereas they are sparser in coarser scales.To reduce impacts on causality from quantization error,a global comparison method is taken to obtain local features,and therefore the method gets smaller number of unstable extrema than local comparison methods.Disk-shaped templates for convolution are discussed to enforce the stability of features under rotational situation,and computational error from the finiteness of radii of templates is also discussed to suppress the influence from the computational error.Based on these strategies,the algorithm of feature extraction presented in this thesis achieves prominent advantage in recall rate and repeatability over algorithms of local comparison.(2)We present an algorithm of matching local features assisted by geometrical information.Since different features may have similar descriptors,matched features through naively comparing descriptors of local features are not necessarily correct.By the Bayes' theorem,analyzing the factors that influence the confidence of matched features,we know that the confidence of matches can be enhanced by introducing an independent event of measurements with nonmatching as the condition.Hence a scheme is studied,in which the geometrical information derived from distances between projects of feature points is utilized to assist the procedure of feature matching.Different from many methods of feature matching assisted by geometrical information,where the scale information of local features is not exploited,our method employs the scale information,which are obtained in the procedure of feature extraction,to compute projects of feature points on a hypothetical plane.If two feature points are not matched,the distances of their projects on the hypothetical plane to an referential point are hardly equal.Moreover,these distances are independent of feature descriptors.Therefore these distances are exploited to create events of measurements under the condition of features being not matched and then to improve the confidence of matched features.Experiments show that the method assisted by these distances reaches higher precision than the methods merely depending on similarity of descriptors.(3)We present an algorithm of least squares consensus for matching local features.Because of gross error and random error,features with similar descriptors can not be considered as truly matched features.Although methods of traditional random sample consensus are widely exploited in algorithms of feature matching and reduce the influence from gross error,they do not diminish the influence from random error.Introducing a second-order stationary stochastic process into a model which transforms points through the homography between two images,we obtain a sufficient condition for the variance of estimation to decrease monotonically when the model is utilized to estimate the homography matrix between two images.Based on this condition and order statistics on distances between descriptors of features,we present an algorithm to estimate homographies with iterative computation of least squares,which lowers gross error in the procedure of iteratively adding a new data point into the sample and simultaneously depresses random error as the sample size increases.The algorithm overcomes its counterparts which exploit the random sample consensus,reaching higher computational efficiency and matching precision.(4)We propose a new method of conditional importance sampling and exploit the new method to present an algorithm of conditional bootstrap filtering.There are two important problems in particle filtering: how to depress the phenomenon of sample degeneracy,and how to improve the efficiency of estimation.In traditional methods of importance sampling,a sample point is accepted with probability 1,and therefore these methods result in too many particles with weights close to 0,causing sample degeneracy rapidly and together lowers the efficiency of estimation.In particular when the proposal density is significantly different from the target density,the situation becomes severer.We introduce a random threshold as a condition into the proposal density,splitting the sampling procedure into two branches: directly sampling and acceptance/rejection sampling.It is proved that the novel method has a good property,namely that the estimation is unbiased or asymptotically unbiased while improving efficiency of estimation.This conditional importance sampling can depress sample degeneracy and improve estimation efficiency when it is utilized in particle filtering.Based on the method of conditional importance sampling and the bootstrap filtering,we present the algorithm of conditional bootstrap filtering,which has prominent advantage outperforming the bootstrap filter and other particle filters in preserving the effective sample size and lowering the root mean square error.For verifying the performance of the novel particle filtering,we work out algorithms for visual object tracking by local features,in which the conditional bootstrap filter and the bootstrap filter are employed respectively.Through tests under a simple scene and a complex scene,results illustrate that the algorithm with the conditional bootstrap filtering achieves larger effective sample size and more precise estimation to the positions of objects in the two scenes.