Rank Optimization For Person Re-identification Through Intelligent Machine Learning Techniques

In computer vision,person re-identification has recently received the significant attention of the researchers and is becoming an emerging research domain with various challenges.Especially,re-ranking or post-rank optimization is a significant challenge.The objective of person re-identification is to re-identify the person when he/she has seen again in multi-cameras.These vision systems are mostly used by security agencies for surveillance in areas like public places,airports,banks and shopping malls etc.Various complexities such as occlusion,illumination variations,camera viewpoint,camera/object motion,and background variations pose difficulties on re-identification process.These re-identification methods mainly focus on two aspects,I)generating robust feature representation or feature descriptors 2)learning effective distance metric.In most approaches,the similarity between the probe and each of the gallery images is computed based on extracted discriminative features and then ranking is performed.Such type of pairwise similarity is unable to explore the complex and high-order relationships between images.Therefore,it leads to suboptimal matching results,especially at rank-1.Existing re-identification methods perform well in some particular scenarios but their performance at rank-1 remains a major concern.Moreover,human-in-the-loop effort makes it more laborious for end users.In order to address such issues,in this dissertation,we present rank optimization and prioritization techniques for person re-identification.Particularly we present two types of the methods.First,that is dealing with the images before the final results are produced is called pre-rank categorization method.Second,those are dealing with the images produced after the final results are called post-rank prioritization and optimization methods.Major work of this study is focused on the second type of methods.In this dissertation,we propose novel post-ranking algorithms and also give insight to the designing and evaluation of post-rank optimization methods.The main research contributions of this dissertation are listed as follows1)Particularly,for handling the large gallery set problem in person re-identification,this dissertation proposes a pre-rank categorization method that is based on the idea of color baskets.In which,we exploit six colors to create these baskets.Further,for signature generation,we use salient dense color features.The scale invariant feature transform(SIFT)and convex hull techniques are used to explore objects in the image.In this method,the person images are divided into three horizontal stripes for obtaining the features.For training incremental linear discriminant analysis(i-LDA)is used,while to reduce the computational cost we exploit maximum relevance and minimum redundancy(mRMR)technique.2)Two novel methods named as hypergraph-based post-rank optimization and multi-feature fusion based re-ranking are proposed regarding post-rank optimization.The graph-based methods proved to be effective in computer vision applications,especially in image retrieval and recognition.But they are unable to explore the higher order relations among samples.Therefore,we exploit the hypergraph that is more effective than simple graph-based methods and this dissertation proposes a hypergraph-based learning scheme that not only improves the rank-1 accuracy but also models the complex and higher order relationships among the images.In this method,after obtaining the rank list using baseline method,we introduced a new refinement algorithm to classify ranks accordingly.This algorithm calculates the position of each image in the list and then finds the correlation information automatically.Furthermore,to discover the relationship among samples,we utilize the hypergraphs for re-rank learning.Soft assignment technique is used to perform weight learning of hyperedges.The advantages of this method are 1)it reduces human in the loop effort;2)it also reduces the initial rank list that ultimately reduces computational cost;3)final results are noise free and relevant.3)In order to address the issues of utilizing single feature types in person re-identification system and especially in post-ranking optimization,this dissertation proposes a multi-feature fusion based re-ranking framework.In most of the traditional techniques,a long feature vector is obtained utilizing an individual modality.However,in the presented approach,multiple features from the samples are extracted and combined into a unified/hybrid vector.Subsequently,a joint feature vector is obtained via fusion.To checking the similarity between the image pairs,the Mahalanobis distance is computed.A novel tree-based re-ranking algorithm is also proposed that exploits this combined feature vector and the distance metric.Therefore,by effective use of each feature type better re-ranking results are achieved using multiple features.Additionally,this method has the advantage that it efficiently manages the memory usage.4)The proposed methods achieve better re-ranking performance by addressing the computational cost,exploring higher order relations among samples and integration of multiple features into the re-ranking process;consequently,the re-identification and re-ranking results are improved.An extensive experimental analysis on challenging and publically available datasets e.g.,VIPeR,CUHK,GRID,and ETHZ,in which we utilize cumulative match characteristic(CMC)curves for performance evaluation,reveals that the proposed re-ranking schemes perform better than the existing methods.Further,the proposed rank optimization and prioritization methods can be integrated with the baseline methods to gain improved and robust re-ranking results.5)Person re-identification post-ranking methods are under-investigated research area nowadays.Therefore,this dissertation presents design and performance evaluation considerations for post-rank optimization(POP)methods.We provide a detailed analysis of the post-ranking algorithms,baseline method selection mechanism and in-depth analysis of benchmark datasets.Further,we discussed future research directions for designing and evaluating POP methods.