A Study Of Visual Loop Closure Detection Based On Feature Combinations

Visual loop closure detection(LCD)is a technique to efficiently and accurately determine whether the vehicle revisits places by matching image features.As a key component in the visual simultaneous localization and mapping(SLAM)system,Visual LCD obtains place matches,which provide key information for the successful optimization of the map.With the rise of autonomous cars,outdoor localization systems are required to work in full-time and large-scale environments.However,when illumination variation,season alternation and other dynamic changes present,the appearance of the scenes may change significantly,posing difficulty for high-precision LCD.Two ways can be explored to robustify visual LCD against the appearance variation of scenes.The first way is to represent images by the features which encode illumination-invariant image attributes.The second way is to represent images by multiple image features jointly in order to utilize the complementary information which may be beneficial to recognize places.Based on the two ideas,this thesis makes the following innovative contributions.(1)The local edgel chordiogram,an image feature descriptor based on the shape of edgels,is proposed.It uses the chordiogram to encode the relative geometric primitives among strong edgels.We demonstrate that the illumination invariance of the feature achieves maximum when an identical proportion of strong edgels is selected in all images.When the illumination condition of the scene changes,it is difficult to recognize some parts of the scenes as their appearances change dramatically accordingly.A calculation method of image similarity is further proposed to reject those unrecognisable regions by only adding up the distances of the features from the similar image regions.Experiments show that the novel similarity measure can reject those confusing image regions.The proposed feature with the proposed similarity measure achieves an outstanding performance of LCD.(2)A method to quantitatively evaluate the LCD precision of individual features and feature combinations is proposed.The LCD precision of a feature is quantified by the separation of the distributions of true and false image matches of places in the distance space of feature vectors.Therefore once a statistical distance is available,a feature quality index(FQI)can be acquired to evaluate the precision of feature combinations for LCD.The optimal feature combination(OFC)can be obtained by maximizing the FQI with respect to the weights.A series of FQIs are proposed,based on Kullback–Leibler divergence,Bhattacharyya divergence,R?enyi divergence,1,2-Wasserstein metrics,Fisher criterion and its improved form.The properties of these FQIs and the relations among them are investigated in detail.Their effectivenesses are experimentally verified.The relations between the performance of the OFCs and the statistical properties of the feature components are also revealed.(3)The properties of the optimal feature combinations are analyzed and the algorithms to solve the OFCs are proposed.When an FQI and a collection of features are available,we first show that by taking the correlations among the feature components into consideration,the OFCs can enhance strong and unique feature components while suppress weak or redundant components.Especially,when the number of feature components is large,their correlations are strong and their performances differ significantly,the OFCs perform significantly better than the uniformly weighted feature combination.Then Algorithm 1 is proposed to solve the OFC maximizing the FQI.Since negative weights may cause false-positive loop closures,Algorithm 2 is proposed to obtain the OFC with nonnegative weights.For broad types of FQIs,it is proved that Algorithm 1is globally optimal and Algorithm 2 is locally optimal.As these supervised algorithms require the loop closure information in the training set,Algorithm 3 is further proposed to estimate it,enabling the algorithms to work in unknown environments.Finally,experiments are conducted to compare the performance of the proposed feature combinations and verify their superiorities.