Research On Metric Learning Algorithms For Cross-Camera Person Re-Identification

Given one person-of-interest(query),person re-ID aims to tell whether this person has been observed in another non-overlapping camera located in a different place.Person re-ID enables tracking the target pedestrian steadily and long-termly across different cam-eras and has attracted increasing interest in the computer vision and pattern recognition community.This thesis has conducted a thorough research into existing metric learning methods in person re-ID and has proposed several novel algorithms from several aspects,i.e.multiple metric learning,cross-view analysis and ranking optimization.The main contributions of this thesis are as follows:(1)A novel metric learning algorithm based on geometric preserving local Fisher discriminate analysis has been proposed.Local Fisher discriminate analysis(LFDA)is an effective algorithm for person re-ID.However,LFDA is limited by linearity and prone to overfitting,thereby leading to poor generalization performance.To overcome this drawback,this thesis assumes that the re-ID data resides on a manifold,which provides better approximation to real-world data structure than linear assumptions.The data is then projected into a new subspace following the criterion that geometric should be well preserved.A novel Fisher criterion is then proposed by incorporating geometric preserving projection into LFDA through a linear weighted technique.In this way,not only the intra-class compactness and inter-class separation is maximized,the geometrical structure is also effectively preserved.The geometric preserving term can serve as a regularization term thus overfitting is alleviated.(2)A multiple metric learning algorithm based on query adaptive weighting and multi-task learning has been proposed.Most existing metric learning methods only learn one unique Mahalanobis distance metric,which ignores the distinct discriminant power of each individual feature and may easily encounter overfitting.In light of this,this thesis proposes a multiple metric learning method where sub-metrics are separately learned for each feature and the final metric is represented as the weighted sum of multiple sub-metrics.Besides,this thesis investigates how to assign weights to the sub-metrics and proposes a two-step weighting strategy.First,the weight of each feature is measured adaptively according to the query in question.Furthermore,to utilize the relatedness among different features,re-ID is formulated as a ranking task and a multi-task learning framework is exploited to jointly learn the weights for all feature types.Experiments demonstrate that the proposed method can greatly boost the accuracy of person re-ID.(3)A semi-supervised subspace learning method based on self-training has been proposed.A critical issue in person re-ID is how to associate people across different views.One popular solution is Canonical correlation analysis(CCA),but CCA requires man-ually labeled image pairs for training,which is labor-expensive and time-consuming.Therefore,this thesis puts forward a self-training based semi-supervised subspace learn-ing approach,which only needs limited labeled data.First,an initial projection is learmed with the li1ited labelled image pairs.Then the unlabeled images are projected into a low-dimensional subspace with the initial projection,in which pseudo pairwise relationships can be constructed.Next,the pseudo pairwise relationships are encoded into a graph Laplacian regularization term to learn new projections.This process is iterated until the pseudo pairwise relationships remain unchanged.The proposed method yields compet-itive performance while requiring much fewer label information than fully-supervised methods.(4)A novel re-ranking strategy based on contextual similarity and reciprocal content similarity has been proposed.A re-ranking step can improve the re-ID performance further by re-estimating the distances between probe and gallery images.However,most current re-ranking ap-proaches are computationally demanding and lead to instability in performanee.There-fore,we put forward a novel light-weight re-ranking strategy,with the underlying as-sumption that true matches should not only have a multitude of mutual nearest neighbors,but also be ranked highly in each other's ranking list.To be specific,we first build the Expanded Cross Neighborhood for all the images and then re-computes the similarity between probe and gallery images according to the overlap ratio between their neighbors(contextual similarity)and their position in each other's ranking list(reciprocal content similarity).By means of this,more possible correct matches can be found and the influ-ence of highly ranked negatives can be significantly eliminated.The proposed re-ranking algorithm is efficient and can yield significant and robust performance gain.